1 Executive Summary¶
The LSST baseline database architecture for its massive user query access system is an MPP (massively parallel processing) relational database composed of a single-node non-parallel DBMS, a distributed communications layer, and a master controller, all running on a shared-nothing cluster of commodity servers with locally attached spinning disk drives, capable of incremental scaling and recovering from hardware failures without disrupting running queries. All large catalogs are spatially partitioned horizontally into materialized chunks, and the remaining catalogs are replicated on each server; the chunks are distributed across all nodes. The Object catalog is further partitioned into sub-chunks with overlaps, materialized on-the-fly when needed. Chunking is handled automatically without exposure to users. Selected tables are also partitioned vertically to maximize performance of most common analysis. The system uses a few critical indexes to speed up spatial searches, time series analysis, and simple but interactive queries. Shared scans are used to answer all but the interactive queries. Such an architecture is primarily driven by the variety and complexity of anticipated queries, ranging from single object lookups to complex \(O(n^2)\) full-sky correlations over billions of elements.
The LSST baseline database architecture for its real time Alert Production relies on horizontal time-based partitioning. To guarantee reproducibility, no-overwrite-update techniques combined with maintaining validity time for appropriate rows are employed. Two database replicas are maintained to isolate live production catalogs from user queries; the replicas are synchronized in real time using native database replication.
Given the current state of the RDBMS and Map/Reduce market, an RDBMS-based solution is a much better fit, primarily due to features such as indexes, schema and speed. No off-the-shelf, reasonably priced solution meets our requirements (today), even though production systems at a scale comparable to LSST have been demonstrated already by industrial users such as eBay using a prohibitively expensive, commercial RDBMS.
The baseline design involves many choices such as component technology, partition size, index usage, normalization level, compression trade-offs, applicability of technologies such as solid state disks, ingest techniques and others. We ran many tests to determine the design configuration, determine limits and uncover potential bottlenecks. In particular, we chose MySQL as our baseline open source, single-node DBMS and XRootD as an open source, elastic, distributed, fault-tolerant messaging system.
We developed a prototype of the baseline architecture, called Qserv. To mitigate major risks, such as insufficient scalability or potential future problems with the underlying RDBMS, Qserv pays close attention to minimizing exposure to vendor-specific features and add-ons. Many key features including the scalable dispatch system and 2-level partitioner have been implemented at the prototype level and integrated with the two underlying production-quality components: MySQL and XRootD. Scalability and performance have been successfully demonstrated on a variety of clusters ranging from 20-node-100TB cluster to 300-node-30TB cluster, tables as large as 50 billion rows and concurrency exceeding 100,000 in-flight chunk-queries. Required data rates for all types of queries (interactive, full sky scans, joins, correlations) have been achieved. Basic fault tolerance recovery mechanisms and basic version of shared scans were demonstrated. Future work includes adding support for user tables, demonstrating cross-matching, various non-critical optimizations, and most importantly, making the prototype more user-friendly and turning it into a production-ready system.
If an equivalent open-source, community supported, off-the-shelf database system were to become available, it could present significant support cost advantages over a production-ready Qserv. The largest barrier preventing us from using an off-the-shelf system is spherical geometry and spherical partitioning support.
To increase the chances such a system will become reality in the next few years, in 2008 we initiated the SciDB array-based scientific database project. Due to lack of many traditional RDBMS-features in SciDB and still nascent fault tolerance, we believe it is easier to build the LSST database system using MySQL+XRootD than it would be to build it based on SciDB. In addition we closely collaborate with the MonetDB open source columnar database team – a successful demonstration of Qserv based on MonetDB instead of MySQL was done in 2012. Further, to stay current with the state-of-the-art in peta-scale data management and analysis, we continue a dialog with all relevant solution providers, both DBMS and Map/Reduce, as well as with data-intensive users, both industrial and scientific, through the XLDB conference series we lead, and beyond.
2 Introduction¶
This document discusses the LSST database system architecture. Chapter 3 discusses the baseline architecture. Chapter 4 explains the LSST database-related requirements. Chapter 5 covers in-depth analysis of existing off-the-shelf potential solutions (Map/Reduce and RDBMS). Chapter 6 discusses design trade-offs and decision process, including small scale tests we run. Chapter 7 covers risk analysis. Chapter 8 discusses the prototype design (Qserv), including design, current status of the software and future plans. Chapters 9 and 10 describe large scale Qserv tests and demonstrations.
3 Baseline Architecture¶
This chapter describes the most important aspects of the LSST baseline database architecture. The choice of the architecture is driven by the project requirements (see 4 Requirements) as well as cost, availability and maturity of the off-the-shelf solutions currently available on the market (see 5 Potential Solutions - Research), and design trade-offs (see 6 Design Trade-offs). The architecture is periodically revisited: we continuously monitor all relevant technologies, and accordingly fine-tune the baseline architecture.
In summary, the LSST baseline architecture for Alert Production is an off-the-shelf RDBMS system which uses replication for fault tolerance and which takes advantage of horizontal (time-based) partitioning. The baseline architecture for user access to Data Releases is an MPP (multi-processor, parallel) relational database running on a shared-nothing cluster of commodity servers with locally attached spinning disk drives; capable of (a) incremental scaling and (b) recovering from hardware failures without disrupting running queries. All large catalogs are spatially partitioned into materialized chunks, and the remaining catalogs are replicated on each server; the chunks are distributed across all nodes. The Object catalog is further partitioned into sub-chunks with overlaps,[1] materialized on-the-fly when needed. Shared scans are used to answer all but low-volume user queries. Details follow below.
[1] | A chunk’s overlap is implicitly contained within the overlaps of its edge sub-chunks. |
3.1 Alert Production and Up-to-date Catalog¶
Alert Production involves detection and measurement of difference-image-analysis sources (DiaSources). New DiaSources are spatially matched against the most recent versions of existing DiaObjects, which contain summary properties for variable and transient objects (and false positives). Unmatched DiaSources are used to create new DiaObjects. If a DiaObject has an associated DiaSource that is no more than a month old, then a forced measurement (DiaForcedSource) is taken at the position of that object, whether a corresponding DiaSource was detected in the exposure or not.
The output of Alert Production consists mainly of three large catalogs – DiaObject, DiaSource, and DiaForcedSource - as well as several smaller tables that capture information about e.g. exposures, visits and provenance.
Catalog | 1st Data Release | Last Data Release |
---|---|---|
DiaObject (unique count) | 0.8 | 15 |
DiaObject (actual in the table) | 15 | 44 |
DiaSource | 23 | 45 |
DiaForcedSource | 15 | 301 |
These catalogs will be modified live every night. After Data Release Production has been run based on the first six months of data and each year’s data thereafter, the live Level 1 catalogs will be archived to tape and replaced by the catalogs produced by DRP. The archived catalogs will remain available for bulk download, but not for queries.
Note that existing DiaObjects are never overwritten. Instead, new versions of the AP-produced and DRP-produced DiaObjects are inserted, allowing users to retrieve (for example) the properties of DiaObjects as known to the pipeline when alerts were issued against them. To enable historical queries, each DiaObject row is tagged with a validity start and end time. The start time of a new DiaObject version is set to the observation time of the DiaSource or DiaForcedSource that led to its creation, and the end time is set to infinity. If a prior version exists, then its validity end time is updated (in place) to equal the start time of the new version. As a result, the most recent versions of DiaObjects can always be retrieved with:
SELECT * FROM DiaObject WHERE validityEnd = infinity
Versions as of some time t are retrievable via:
SELECT * FROM DiaObject WHERE validityStart <= t AND t < validityEnd
Note that a DiaSource can also be reassociated to a solar-system object during day time processing. This will result in a new DiaObject version unless the DiaObject no longer has any associated DiaSources. In that case, the validity end time of the existing version is set to the time at which the reassociation occurred.
Once a DiaSource is associated with a solar system object, it is never associated back to a DiaObject. Therefore, rather than also versioning DiaSources, columns for the IDs of both the associated DiaObject and solar system object, as well as a reassociation time, are included. Reassociation will set the solar system object ID and reassociation time, so that DiaSources for DiaObject 123 at time t can be obtained using:
SELECT *
FROM DiaSource
WHERE diaObjectId = 123
AND midPointTai <= t
AND (ssObjectId is NULL OR ssObjectReassocTime > t)
DiaForcedSources are never reassociated or updated in any way.
From the database point of view then, the alert production pipeline will perform the following database operations 189 times (once per LSST CCD) per visit (every 39 seconds):
- Issue a point-in-region query against the DiaObject catalog, returning the most recent versions of the objects falling inside the CCD.
- Use the IDs of these diaObjects to retrieve all associated diaSources and diaForcedSources.
- Insert new diaSources and diaForcedSources.
- Update validity end times of diaObjects that will be superseded.
- Insert new versions of diaObjects.
All spatial joins will be performed on in-memory data by pipeline code, rather than in the database. While Alert Production does also involve a spatial join against the Level 2 (DRP-produced) Object catalog, this does not require any database interaction: Level 2 Objects are never modified, so the Object columns required for spatial matching will be dumped to compact binary files once per Data Release. These files will be laid out in a way that allows for very fast region queries, allowing the database to be bypassed entirely.
The DiaSource and DiaForcedSource tables will be split into two tables, one for historical data and one containing records inserted during the current night. The current-night tables will be small and rely on a transactional engine like InnoDB, allowing for speedy recovery from failures. The historical-data tables will use the faster non-transactional MyISAM or Aria storage engine, and will also take advantage of partitioning. The Data Release catalogs used to seed the live catalogs will be stored in a single initial partition, sorted spatially (using the Hierarchical Triangular Mesh trixel IDs for their positions). This means that the diaSources and diaForcedSources for the diaObjects in a CCD will be located close together on disk, minimizing seeks. Every month of new data will be stored in a fresh partition, again sorted spatially. Such partitions will grow to contain just a few billion rows over the course of a month, even for the largest catalog. At the end of each night, the contents of the current-night table are sorted and appended to the partition for the current-month, then emptied. Each month, the entire current-month partition is sorted spatially (during the day), and a partition for the next month is created.
For DiaObject, the same approach is used. However, DiaObject validity end-time updates can occur in any partition, and are not confined to the current-night table. We therefore expect to use a transactional storage engine like InnoDB for all partitions. Because InnoDB clusters tables using the primary key, we will likely declare it to consist of a leading HTM ID column, followed by disambiguating columns (diaObjectId, validityStart). The validity end time column will not be part of any index.
No user queries will be allowed on the live production catalogs. We
expect to maintain a separate replica just for user queries,
synchronized in real time using one-way master-slave native database
replication. The catalogs for user queries will be structured
identically to the live catalogs, and views will be used to hide the
splits (using a “UNION ALL
”).
For additional safety, we might choose to replicate the small current-night tables, all DiaObject partitions, and the remaining (small) changing tables to another hot stand-by replica. In case of disastrous master failure that cannot be fixed rapidly, the slave serving user queries will be used as a temporary replacement, and user queries will be disallowed until the problem is resolved.
Based on the science requirements, only short-running, relatively simple user queries will be needed on the Level 1 catalogs. The most complex queries, such as large-area near neighbor queries, will not be needed. Instead, user queries will consist mainly of small-area cone searches, light curve lookups, and historical versions of the same. Since the catalogs are sorted spatially, we expect to be able to quickly answer spatial queries using indexed HTM ID columns and the scisql UDFs, an approach that has worked well in data-challenges to date. Furthermore, note that the positions of diaSources/diaForcedSources associated with the same diaObject will be very close together, so that sorting to obtain good spatial locality also ends up placing sources belonging to the same light curve close together. In other words, the data organization used to provide fast pipeline query response is also advantageous for user queries.
3.2 Data Release Production¶
Data Release Production will involve the generation of significantly larger catalogs than Alert Production. However, these are produced over the course of several months, pipelines will not write directly to the database, and there are no pipeline queries with very low-latency execution time requirements to be satisfied. While we do expect several pipeline-related full table scans over the course of a Data Release Production, we will need to satisfy many user queries involving such scans on a daily basis. User query access is therefore the primary driver of our scalable database architecture, which is described in detail below. For a description of the data loading process, please see 8.15.2 Data loading.
3.3 User Query Access¶
The user query access is the primary driver of the scalable database architecture. Such architecture is described below.
3.3.1 Distributed and parallel¶
The database architecture for user query access relies on a model of distributing computation among autonomous worker nodes. Autonomous workers have no direct knowledge of each other and can complete their assigned work without data or management from their peers. This implies that data must be partitioned, and the system must be capable of dividing a single user query into sub-queries, and executing these sub-queries in parallel – running a high-volume query without parallelizing it would take unacceptably long time, even if run on very fast CPU. The parallelism and data distribution should be handled automatically by the system and hidden from users.
3.3.3 Indexing¶
Disk I/O bandwidth is expected to be the greatest bottleneck. Data can be accessed either through index, which typically translates to a random access, or a scan, which translates to a sequential read (unless multiple competing scans are involved).
Indexes dramatically speed up locating individual rows, and avoid expensive full table scans. They are essential to answer low volume queries quickly, and to do efficient table joins. Also, spatial indexes are essential. However, unlike in traditional, small-scale systems, the advantages of indexes become questionable when a larger number of rows is to be selected from a table. In case of LSST, selecting even a 0.01% of a table might lead to selecting millions of rows. Since each fetch through an index might turn into a disk seek, it is often cheaper to read sequentially from disk than to seek for particular rows via index, especially when the index itself is out-of-memory. For that reason the architecture forgoes relying on heavy indexing, only a small number of carefully selected indexes essential for answering low-volume queries, enabling table joins, and speeding up spatial searches will be maintained. For an analytical query system, it makes sense to make as few assumptions as possible about what will be important to our users, and to try and provide reasonable performance for as broad a query load as possible, i.e. focus on scan throughput rather than optimizing indexes. A further benefit to this approach is that many different queries are likely to be able to share scan IO, boosting system throughput, whereas caching index lookup results is likely to provide far fewer opportunities for sharing as the query count scales (for the amounts of cache we can afford).
3.3.5 Clustering¶
The data in the Object Catalog will be physically clustered on disk spatially – that means that objects collocated in space will be also collocated on disk. All Source-type catalogs (Source, ForcedSource, DiaSource, DiaForcedSource) will be clustered based on their corresponding objectId – this approach enforces spatial clustering and collocates sources belonging to the same object, allowing sequential read for queries that involve times series analysis.
SSObject catalog will be unpartitioned, because there is no obvious fixed position that we could choose to use for partitioning. The associated diaSources (which will be intermixed with diaSources associated with static diaSources) will be partitioned, according their position. For that reason the SSObject-to-DiaSource join queries will require index searches on all chunks, unlike DiaObject-to-DiaSource queries. Since SSObject is small (low millions), this should not be an issue.
3.3.6 Partitioning¶
Data must be partitioned among nodes in a shared-nothing architecture. While some sharding approaches partition data based on a hash of the primary key, this approach is unusable for LSST data since it eliminates optimizations based on celestial objects’ spatial nature.
3.3.6.1 Sharded data and sharded queries¶
All catalogs that require spatial partitioning (Object, Source, ForcedSource, DiaSource, DiaForcedSource) as well as all the auxiliary tables associated with them, such as ObjectType, or PhotoZ, will be divided into spatial partitions of roughly the same area by partitioning then into declination zones, and chunking each zone into RA stripes. Further, to be able to perform table joins without expensive inter-node data transfers, partitioning boundaries for each partitioned table must be aligned, and chunks of different tables corresponding to the same area of sky must be co-located on the same node. To ensure chunks are appropriately sized, the two largest catalogs, Source and ForcedSource, are expected to be partitioned into finer-grain chunks. Since objects occur at an approximately-constant density throughout the celestial sphere, an equal-area partition should spread a load that is uniformly distributed over the sky.
Smaller catalogs that can be partitioned spatially, such as Alert and exposure metadata will be partitioned spatially. All remaining catalogs, such provenance or SDQA tables will be replicated on each node. The size of these catalogs is expected to be only a few terabytes.
With data in separate physical partitions, user queries are themselves fragmented into separate physical queries to be executed on partitions. Each physical query’s result can be combined into a single final result.
3.3.6.2 Two-level partitions¶
Determining the size and number of data partitions may not be obvious. Queries are fragmented according to partitions so an increasing number of partitions increases the number of physical queries to be dispatched, managed, and aggregated. Thus a greater number of partitions increases the potential for parallelism but also increases the overhead. For a data-intensive and bandwidth-limited query, a parallelization width close to the number of disk spindles should minimize seeks and maximizing bandwidth and performance.
From a management perspective, more partitions facilitate rebalancing data among nodes when nodes are added or removed. If the number of partitions were equal to the number of nodes, then the addition of a new node would require the data to be re-partitioned. On the other hand, if there were many more partitions than nodes, then a set of partitions could be assigned to the new node without re-computing partition boundaries.
Smaller and more numerous partitions benefit spatial joins. In an astronomical context, we are interested in objects near other objects, and thus a full \(O(n^2)\) join is not required–a localized spatial join is more appropriate. With spatial data split into smaller partitions, an SQL engine computing the join need not even consider (and reject) all possible pairs of objects, merely all the pairs within a region. Thus a task that is \(O(n^2)\) naively becomes \(O(kn)\) where \(k\) is the number of objects in a partition.
In consideration of these trade-offs, two-level partitioning seems to be a conceptually simple way to blend the advantages of both extremes. Queries can be fragmented in terms of coarse partitions (“chunks”), and spatial near-neighbor joins can be executed over more fine partitions (“sub-chunks”) within each partition. To avoid the overhead of the sub-chunks for non-join queries, the system can store chunks and generate sub-chunks on-demand for spatial join queries. On-the-fly generation for joins is cost-effective due to the drastic reduction of pairs, which is true as long as there are many sub-chunks for each chunk.
3.3.6.3 Overlap¶
A strict partitioning eliminates nearby pairs where objects from adjacent partitions are paired. To produce correct results under strict partitioning, nodes need access to objects from outside partitions, which means that data exchange is required. To avoid this, each partition can be stored with a pre-computed amount of overlapping data. This overlapping data does not strictly belong to the partition but is within a preset spatial distance from the partition’s borders. Using this data, spatial joins can be computed correctly within the preset distance without needing data from other partitions that may be on other nodes.
Overlap is needed only for the Object Catalog, as all spatial correlations will be run on that catalog only. Guided by the experience from other projects including SDSS, we expect to preset the overlap to ~1 arcmin, which results in duplicating approximately 30% of the Object Catalog.
3.3.6.4 Spherical geometry¶
Support for spherical geometry is not common among databases and spherical geometry-based partitioning was non-existent in other solutions when we decided to develop Qserv. Since spherical geometry is the norm in recording positions of celestial objects (right-ascension and declination), any spatial partitioning scheme for astronomical object must account for its complexities.
3.3.6.5 Data immutability¶
It is important to note that user query access operates on read-only data. Not having to deal with updates simplifies the architecture and allows us to add extra optimizations not possible otherwise. The Level 1 data which is updated is small enough and will not require the scalable architecture – we plan to handle all Level 1 data set with out-of-the box MySQL as described in 3.1 Alert Production and Up-to-date Catalog.
3.3.7 Long-running queries¶
Many of the typical user queries may need significant time to complete, at the scale of hours. To avoid re-submission of those long-running queries in case of various failures (networking or hardware issues) the system will support asynchronous query execution mode. In this mode users will submit queries using special options or syntax and the system will dispatch a query and immediately return to user some identifier of the submitted query without blocking user session. This query identifier will be used by user to retirieve query processing status, query result after query completes, or a partial query result while query is still executing.
The system should be able to estimate the time which user query will need to complete and refuse to run long queries in a regular blocking mode.
3.3.8 Technology choice¶
As explained in Chapter 5, no off-the-shelf solution meets the above requirements today, and RDBMS is a much better fit than Map/Reduce-based system, primarily due to features such as indexes, schema and speed. For that reason, our baseline architecture consists of custom software built on two production components: an open source, “simple”, single-node, non-parallel DBMS (MySQL) and XRootD. To ease potential future DBMS migrations, the communication with the underlying DBMS relies on basic DBMS functionality only, and avoids any vendor-specific features and additions.
4 Requirements¶
The key requirements driving the LSST database architecture include: incremental scaling, near-real-time response time for ad-hoc simple user queries, fast turnaround for full-sky scans/correlations, reliability, and low cost, all at multi-petabyte scale. These requirements are primarily driven by the ad-hoc user query access.
4.1 General Requirements¶
Incremental scaling. The system must to tens of petabytes and trillions of rows. It must grow as the data grows and as the access requirements grow. New technologies that become available during the life of the system must be able to be incorporated easily. Expected sizes for the largest database catalogs (for the last data release, uncompressed, data only) are captured in the table below. For further storage, disk and network bandwidth and I/O analyses, see [LDM-141].
Table | Size [TB] | Rows [billion] | Columns | Description |
---|---|---|---|---|
Object (narrow) | ~107 | ~47 | ~330 | Most heavily used, for all common queries on stars/galaxies, including spatial correlations and time series analysis using summarized information |
Object (all extras) | ~1,200 | Largest ~1,500 | ~7,650 | Specialized analysis of objects, including photo-Z, covariances, B+D samples |
Source | ~5,000 | ~9,000 | ~50 | Time series analysis of bright objects and detections |
ForcedSource | ~1,900 | ~50,000 | 6 | Specialized analysis of faint objects and detections |
SSObject | ~0.003 | 0.006 | ~80 | Analysis of solar system (moving) objects |
DiaObject | ~27 (76 max) | ~15 unique (~44 max) | ~260 | Analysis of variable and transient objects, often alert-related |
DiaSource | ~23 | ~45 | ~70 | Time series analysis of variable and transient objects, often alert-related |
DiaForcedSource | ~13 | ~300 | 8 | Specialized analysis of faint diaObjects and detections |
Reliability. The system must not lose data, and it must provide at least 98% up time in the face of hardware failures, software failures, system maintenance, and upgrades.
Low cost. It is essential to not overrun the allocated budget, thus a cost-effective, preferably open-source solution is strongly preferred.
4.4 Discussion¶
4.4.1 Implications¶
The above requirements have important implications on the LSST data access architecture.
- The system must allow rapid selection of small number of rows out of multi-billion row tables. To achieve this, efficient data indexing in both spatial and temporal dimensions is essential.
- The system must efficiently join multi-trillion with multi-billion row tables. Denormalizing these tables to avoid common joins, such as Object with Source or Object with ForcedSource, would be prohibitively expensive.
- The system must provide high data bandwidth. In order to process terabytes of data in minutes, data bandwidths on the order of tens to hundreds of gigabytes per second are required.
- To achieve high bandwidths, to enable expandability, and to provide fault tolerance, the system will need to run on a distributed cluster composed of multiple machines.
- The most effective way to provide high-bandwidth access to large amounts of data is to partition the data, allowing multiple machines to work against distinct partitions. Data partitioning is also important to speed up some operations on tables, such as index building.
- Multiple machines and partitioned data in turn imply that at least the largest queries will be executed in parallel, requiring the management and synchronization of multiple tasks.
- Limited budget implies the system needs to get most out available hardware, and scale it incrementally as needed. The system will be disk I/O limited, and therefore we anticipate attaching multiple queries to a single table scan (shared scans) will be a must.
4.4.2 Query complexity and access patterns¶
A compilation of representative queries provided by the LSST Science Collaborations, the Science Council, and other surveys have been captured [common-queries]. These queries can be divided into several distinct groups: analysis of a single object, analysis of objects meeting certain criteria in a region or across entire sky, analysis of objects close to other objects, analysis that require special grouping, time series analysis and cross match with external catalogs. They give hints as to the complexity required: these queries include distance calculations, spatially localized self-joins, and time series analysis.
Small queries are expected to exhibit substantial spatial locality (refer to rows that contain similar spatial coordinates: right ascension and declination). Some kinds of large queries are expected to exhibit a slightly different form of spatial locality: joins will be among rows that have nearby spatial coordinates. Spatial correlations will be executed on the Object table; spatial correlations will not be needed on Source or ForcedSource tables.
Queries related to time series analysis are expected to need to look at the history of observations for a given Object, so the appropriate Source or ForcedSource rows must be easily joined and aggregate functions operating over the list of Sources must be provided.
External data sets and user data, including results from past queries may have to be distributed alongside distributed production table to provide adequate join performance.
The query complexity has important implications on the overall architecture of the entire system.
5 Potential Solutions - Research¶
The two most promising technologies able to scale to LSST size available today are Relational Database Management Systems (RDBMS) and Map/Reduce (MR): the largest DBMS system, based on commercial Teradata and reaching 20+ petabytes is hosted at eBay, and the largest MR-based systems, reaching many tens of petabytes are hosted at Google, Facebook, Yahoo! and others.
5.1 The Research¶
To decide which technology best fits the LSST requirements, we did an extensive market research and analyses, reviewed relevant literature, and run appropriate stress-tests for selected “promising” candidates, focusing on weak points and the most uncertain aspects. Market research and analyses involved (a) discussions about lessons learned with many industrial and scientific users dealing with large data sets, (b) discussions about existing solutions and future product road-maps with leading solution providers and promising start-ups, and (c), research road-map with leading database researchers from academia. See 16 Appendix E: People/Communities We Talked To.
5.2 The Results¶
As a result of our research, we determined an RDBMS-based solution involving a shared-nothing parallel database is a much better fit for the LSST needs than MR. The main reasons are availability of indexes which are absolutely essential for low-volume queries and spatial indexes, support for schemas and catalogs, performance and efficient use of resources.
Even though a suitable open source off-the-shelf DMBS capable of meeting LSST needs does not exist today, there is a good chance a system meeting most of the key requirements will be available well before LSST production starts. In particular, there are two very promising by-products of our research:
- we co-initiated, co-founded, and helped bootstrap SciDB – a new open source shared nothing database system, and
- we pushed the development of MonetDB, an open source columnar database into directions very well aligned with LSST needs. We closely collaborate with the MonetDB team – building on our Qserv lessons-learned, the team is trying to add missing features and turn their software into a system capable of supporting LSST needs. In 2012 we demonstrated running Qserv with a MonetDB backend instead of MySQL.
Both SciDB and MonetDB have strong potential to become the LSST database solution once they mature.
Further, our research led to creation a new, now internationally-recognized conference series, Extremely Large Databases (XLDB). As we continue leading the XLDB effort, it gives us a unique opportunity to reach out to a wide range of high-profile organizations dealing with large data sets, and raise awareness of the LSST needs among researchers and developers working on both MR and DBMS solutions.
The remaining of this chapter discusses lessons learned to-date, along with a description of relevant tests we have run.
5.3 Map/Reduce-based and NoSQL Solutions¶
Map/Reduce is a software framework to support distributed computing on large data sets on clusters of computers. Google’s implementation [Dean04], believed to be the most advanced, is proprietary, and in spite of Google being one of the LSST collaborators, we were unable to gain access to any of their MR software or infrastructure. Additional Google-internal MR-related projects include BigTable [Bigtable06], Chubby [Chubby], and Sawzall [Pike]. BigTable addresses the need for rapid searches through a specialized index; Chubby adds transactional support; and Sawzall is a procedural language for expressing queries. Most of these solutions attempt to add partial database-like features such as schema catalog, indexes, and transactions. The most recent MR developments at Google are Dremel [Dremel] - an interactive ad-hoc query system for analysis of read-only data, and Tenzing – a full SQL implementation on the MR Framework [Chattopadhyay11].[3]
[3] | Through our XLDB efforts, Google has provided us with a preprint of a Tenzing manuscript accepted for publication at VLDB 2011. |
In parallel to the closed-source systems at Google, similar open-source solutions are built by a community of developers led by Facebook, Yahoo! and Cloudera, and they have already gained wide-spread acceptance and support. The open source version of MR, Hadoop, has became popular in particular among industrial users. Other solutions developed on top (and “around”) Hadoop include HBase (equivalent of BigTable), Hive (concept similar to Google’s Dremel), Pig Latin (equivalent to Google’s Sawzall), Zookeeper (equivalent to Google’s Chubby), Simon, and others. As in Google’s case, the primary purpose of building these solutions is adding database-features on top of MR. Hadoop is commercially supported by Cloudera, Hortonworks [Yahoo] and Hadapt.
We have experimented with Hadoop (0.20.2) and Hive (0.7.0) in mid 2010 using a 1 billion row USNO-B data set on a 64 node cluster. Common LSST queries were tested, ranging from low-volume type (such as finding a single object, selecting objects near other know object), through high-volume ones (full table scans) to complex queries involving joins (join was implemented in a standard way, in the reduce step). The results were discussed with Hadoop/Hive experts from Cloudera. Periodically we revisit the progress and feature set available in the Hadoop ecosystem, but to date we have not found compelling reasons to consider Hadoop as a serious alternative for managing LSST data.
Independently, Microsoft developed a system called Dryad, geared towards executing distributed computations beyond “flat” Map and Reduce, along with a corresponding language called LINQ. Due to its strong dependence on Windows OS and limited availability, use of Dryad outside of Microsoft is very limited. Based on news reports [Zdnet], Microsoft dropped support for Dryad back in late 2011.
Further, there is a group of new emerging solutions often called as NoSQL. The two most popular ones are MongoDB and Cassandra.
The remaining of this section discusses all of the above-mentioned products.
Further details about individual MR and no-SQL solutions can be found in 12 Appendix A – Map/Reduce Solutions and 13 Appendix B – Database Solutions.
5.4 DBMS Solutions¶
Database systems have been around for much longer than MR, and therefore they are much more mature. They can be divided into many types: parallel/single node, relational/object-oriented, columnar/row-based; some are built as appliances. Details about individual DBMS products and solutions we considered and/or evaluated can be found in 13 Appendix B – Database Solutions.
5.4.1 Parallel DBMSes¶
Parallel databases, also called MPP DBMS (massively parallel processing DBMS), improve performance through parallelization of queries: using multiple CPUs, disks and servers in parallel. Data is processed in parallel, and aggregated into a final result. The aggregation may include computing average, max/min and other aggregate functions. This process is often called scatter-gather, and it is somewhat similar to map and reduce stages in the MR systems.
Shared-nothing parallel databases, which fragment data and in many cases use an internal communications strategy similar to MR, scale significantly better than single-node or shared-disk databases. Teradata uses proprietary hardware, but there are a number of efforts to leverage increasingly-fast commodity networks to achieve the same performance at much lower cost, including Greenplum, DB2 Parallel Edition, Aster Data, GridSQL, ParAccel, InfiniDB, SciDB, and Project Madison at Microsoft (based on DATAllegro, acquired by Microsoft in 2008). Most of these efforts are relatively new, and thus the products are relatively immature. EBay’s installation used to be based on Greenplum in 2009 and reached 6.5 PB, but their current Singularity system is now approaching 30 PB and is based on Teradata’s appliances. Some of these databases have partition-wise join, which can allow entity/observation join queries to execute more efficiently, but none allow overlapping partitions, limiting the potential performance of pairwise analysis.
Microsoft SQL Server offers Distributed Partitioned Views, which provide much of the functionality of a shared-nothing parallel database by federating multiple tables across multiple servers into a single view. This technology is used in the interesting GrayWulf project [Szalay08] [Simmhan09] which is designed to host observational data consisting of Pan-STARRS PS1 [Jedicke06] astronomical detections and summary information about the objects that produced them. GrayWulf partitions observation data across nodes by “zones” [Gray07], but these partitions cannot overlap. Fault tolerance is built in by having three copies of the data, with one undergoing updates – primarily appending new detections – and the other two in a hot/warm relationship for failover. GrayWulf has significant limitations, however. The object information for the Pan-STARRS PS1 data set is small enough (few TB) that it can be materialized on a single node. The lack of partition-wise join penalizes entity/observation join queries and pairwise analysis. The overall system design is closely tied to the commercial SQL Server product, and re-hosting it on another RDBMS, in particular an open source one, would be quite difficult.
The MPP database is ideal for the LSST database architecture. Unfortunately, the only scalable, proven off-the-shelf solutions are commercial and expensive: Teradata, Greenplum. Both systems are (or recently were) behind today world’s largest production database systems at places such as eBay [dbms209] [dbms210] and Walmart [eweek04]. IBM’s DB2 “parallel edition”, even though it implements a shared-nothing architecture since mid-1990 focuses primarily on supporting unstructured data (XML), not large scale analytics.
The emergence of several new startups, such as Aster Data, DataAllegro, ParAccel, GridSQL and SciDB is promising, although some of them have already been purchased by the big and expensive commercial RDBMSes: Teradata purchased Aster Data, Microsoft purchased DataAllegro. To date, the only shared-nothing parallel RDBMS available as open source is SciDB – its first production version (v11.06) was released in June 2011. ParAccel is proprietary, we did not have a chance to test it, however given we have not heard of any large scale installation based on ParAccel we have doubts whether it’ll meet our needs. After testing GridSQL we determined it does not offer enough benefits to justify using it, the main cons include limited choices of partitioning types (hash partitioning only), lack of provisions for efficient near neighbor joins, poor stability and lack of good documentation.
SciDB is the only parallel open source DBMS currently available on the market. It is a columnar, shared-nothing store based on an array data model. The project has been inspired by the LSST needs [scidb], and the LSST Database team is continuously in close communication with the SciDB developers. SciDB’s architectural features of chunking large arrays into overlapping chunks and distributing these chunks across a shared nothing cluster of machines match the LSST database architecture. Initial tests run with the v0.5 SciDB release exposed architectural issues with SciDB essential for LSST, related to clustering and indexing multi-billion, sparse arrays of objects in a 2-dimensional (ra, declination) space. These issues have been addressed since then and re-testing is planned.
There are several reasons why SciDB is not our baseline, and we currently do not have plans to use it for LSST catalog data. First, as an array database, SciDB uses a non-SQL query language (actually, two) appropriate for arrays. Adapting this to SQL, likely through a translation layer, is a substantial burden, even more difficult than parsing SQL queries for reissue as other SQL queries. (Given the widespread use of SQL in the astronomy community and the ecosystem of tools available for SQL, moving away from SQL would be a major endeavor.) Second, while relations can be thought of as one-dimensional arrays, SciDB is not optimized to handle them as well as a traditional RDBMS, in particular for the variety of joins required (including star schema, merge joins, and self joins). Standard RDBMS features like views, stored procedures, and privileges would have to be added from the ground up. Third, SciDB’s fault tolerance is not yet at the level of XRootD. Overall, the level of coding we would have to do to build on the current SciDB implementation appears to be larger than what we are planning on top of XRootD/MySQL. As SciDB’s implementation progresses, though, this trade-off could change.
5.4.2 Object-oriented solutions¶
The object-oriented database market is very small, and the choices are limited to a few small proprietary solutions, including Objectivity/DB and InterSystems Caché. Objectivity/DB was used by the BaBar experiment in 1999 – 2002, and the BaBar database reached a petabyte [Becla05]. The members of LSST database team, being the former members of the BaBar database team are intimately familiar with the BaBar database architecture. The Objectivity/DB was used primarily as a simple data store, all the complexity, including custom indices had to be all built in custom, user code. Given that, combining with the challenges related to porting and debugging commercial system led as to a conclusion Objectivity/DB is not the right choice for LSST.
InterSystems Caché has been chosen as the underlying system for the European Gaia project [25, 26], based on our limited knowledge, so far the Gaia project focused primarily on using Caché for ingest-related aspects of the system, and did not have a chance to research analytical capabilities of Caché at scale.
5.4.3 Row-based vs columnar stores¶
Row-based stores organize data on disk as rows, while columnar store – as columns. Column-store databases emerged relatively recently, and are based on the C-store work [Stonebaker05]. By operating on columns rather than rows, they are able to retrieve only the columns required for a query and greatly compress the data within each column. Both reduce disk I/O and hence required hardware by a significant factor for many analytical queries on observational data that only use a fraction of the available columns. Current column stores also allow data to be partitioned across multiple nodes, operating in a shared-nothing manner. Column stores are less efficient for queries retrieving small sets of full-width records, as they must reassemble values from all of the columns.
Our baseline architecture assumes all large-volume queries will be answered through shared scans, which reduces wasting disk I/O for row-based stores: multiple queries attached to the same scan will typically access most columns (collectively). We are also vertically partitioning our widest table into frequently-accessed and infrequently-accessed columns to get some of the advantage of a column store.
Nevertheless, a column store could still be more efficient. Work done at Google (using Dremel) has claimed that “the crossover point often lies at dozens of fields but it varies across data sets” [Dremel]. In our case, the most frequently accessed table: Object, will have over “20 dozens” columns. The Source, DiaObject, and DiaSource tables will each have about 4 dozen columns. These could be wide enough that all simultaneously executing queries will still only touch a subset of the columns. Most other large tables are relatively narrow and are expected to have all columns used by every query. Low query selectivity (expected to be <1% for full table scans) combined with late materialization (postponing row assembly until the last possible moment) is expected to further boost effectiveness of columnar stores.
The two leading row-based DBMSes are MySQL and PostgreSQL. Of these two, MySQL is better supported, and has much wider community of users, although both are commercially supported (MySQL: Oracle, MontyProgram+SkySQL, Percona. PostgreSQL: EnterpriseDB). PostgreSQL tends to focus more on OLTP, while MySQL is closer to our analytical needs, although both are weak in the area of scalability. One of the strongest points of PostgreSQL used to be spatial GIS support, however MySQL has recently rewritten their GIS modules and it now offers true spatial relationship support (starting from version 5.6.1). Neither provides good support for spherical geometry including wraparound, however.
Many commercial row-bases DBMSes exist, including Oracle, SQL Server, DB2, but they do not fit well into LSST needs, since we would like to provide all scientists with the ability to install the LSST database at their institution at low licensing and maintenance cost.
Columnar stores are starting to gain in popularity. Although the list is already relatively large, the number of choices worth considering is relatively small. Today’s most popular commercial choice is HP Vertica, and the open source solutions include MonetDB and Calpont’s InfiniDB. The latter also implements shared nothing MPP, however the multi-server version is only available as part of the commercial edition.
With help from Calpont, we evaluated InfiniDB and demonstrated it could be used for the LSST system – we run the most complex (near neighbor) query. Details are available in 14 Appendix C: Tests with InfiniDB.
We are working closely with the MonetDB team, including the main architect of the system, Martin Kersten and two of his students who worked on porting MonetDB to meet LOFAR database needs. In 2011 the MonetDB team has run some basic tests using astronomical data (USNOB as well as our DC3b-PT1.1 data set). During the course of testing our common queries they implemented missing features such as support for user defined functions, and are actively working on further extending MonetDB to build remaining missing functionality, in particular ability to run as a shared-nothing system. To achieve that, existing MonetDB server (merovingian) has to be extended. Table partitioning and overlaps (on a single node) can be achieved through table views, although scalability to LSST sizes still needs to be tested. Cross-node partitioning requires new techniques, and the MonetDB team is actively working on it.
In 2012 with help from the MonetDB team we demonstrated a limited set of queries on a Qserv system integrated with MonetDB on the backend rather than MySQL. While the integration was left incomplete, the speed at which we were able to port Qserv to a new database and execute some queries is convincing evidence of Qserv’s modularity. Because basic functionality was ported in one week, we are confident that porting to another DBMS can be done with modest effort in a contingency or for other reasons. The experience has also guided Qserv design directions and uncovered unintended MySQL API dependence in Qserv and broader LSST DM systems.
5.4.4 Appliances¶
Appliances rely on specialized hardware to achieve performance. In general, we are skeptical about appliances, primarily because they are locking us into this specialized hardware. In addition, appliances are usually fast, however their replacement cost is high, so often commodity hardware is able to catch up, or even exceed the performance of an appliance after few years (the upgrade of an appliance to a latest version is usually very costly).
5.5 Comparison and Discussion¶
The MR processing paradigm became extremely popular in the last few years, in particular among peta-scale industrial users. Most industrial users with peta-scale data sets heavily rely on it, including places such as Google, Yahoo!, Amazon or Facebook, and even eBay has recently started using Hadoop for some of their (offline, batch) analysis. The largest (peta-scale) RDBMS-based systems all rely on shared-nothing, MPP technology, and almost all on expensive Teradata solutions (eBay, Walmart, Nokia, for a few years eBay used Greenplum but they switched back to Teradata’s Singularity).
In contrast, science widely adopted neither RDBMS nor MR. The community with the largest data set, HEP, is relying on a home-grown system, augmented by a DBMS (typically Oracle or MySQL) for managing the metadata. This is true for most HEP experiments of the last decade (with the exception of BaBar which initially used Objectivity), as well as the LHC experiments. In astronomy, most existing systems as well as the systems starting in the near future are RDBMS-based (SDSS – SQL Server, Pan-STARRS – SQL Server, 2MASS – Informix, DES – Oracle, LOFAR – MonetDB, Gaia – Caché). It is worth noting that none of these systems was large enough so far to break the single-node barrier, with the exception of Pan-STARRS. Geoscience relies primarily on netCDF/HDF5 files with metadata in a DBMS. Similar approach is taken by bio communities we have talked to. In general, MR approach has not been popular among scientific users so far.
The next few sections outline key differences, strengths and weaknesses of MR and RDBMS, and the convergence.
5.5.1 APIs¶
In the MR world, data is accessed by a pair of functions, one that is “mapped” to all inputs, and one that “reduces” the results from the parallel invocations of the first. Problems can be broken down into a sequence of MR stages whose parallel components are explicit. In contrast, a DBMS forces programmers into less natural, declarative thinking, giving them very little control over the flow of the query execution; this issue might partly go away by interacting with database through a user defined function (UDFs), which are becoming increasingly popular. They must trust the query optimizer’s prowess in “magically” transforming the query into a query plan. Compounding the difficulty is the optimizer’s unpredictability: even one small change to a query can make its execution plan efficient or painfully slow.
The simplicity of the MR approach has both advantages and disadvantages. Often a DBMS is able to perform required processing on the data in a small number of passes (full table scans). The limited MR operators on the other hand may lead to many more passes through the data, which requires more disk I/O thus reduces performance and increases hardware needed. Also, MR forced users to code a lot of operations typically provided by an RDBMS by-hand – these include joins, custom indexes or even schemas.
5.5.2 Scalability, fault tolerance and performance¶
The simple processing framework of MR allows to easily, incrementally scale the system out by adding more nodes as needed. Frequent check-pointing done by MR (after every “map” and every “reduce” step) simplifies recoverability, at the expense of performance. In contrast, databases are built with the optimistic assumptions that failures are rare: they generally checkpoint only when necessary. This has been shown through various studies [Pavlo09].
The frequent checkpointing employed by MR, in combination with limited set of operators discussed earlier often leads to inefficient usages of resources in MR based systems. Again, this has been shown through various studies. EBay’s case seems to support this as well: back in 2009 when they managed 6.5 petabytes of production data in an RDBMS-based system they relied on a mere 96 nodes, and based on discussions with the original architects of the eBay system, to achieve comparable processing power through MR, many thousand nodes would be required.
5.5.3 Flexibility¶
MR paradigm treats a data set as a set of key-value pairs. It is structure-agnostic, leaving interpretation to user code and thus handling both poorly-structured and highly-complex data. Loose constraints on data allow users to get to data quicker, bypassing schema modeling, complicated performance tuning, and database administrators. In contrast, data in databases are structured strictly in records according to well-defined schemata.
While adjusting schema with ease is very appealing, in large scientific projects like LSST, the schema has to be carefully thought through to meet the needs of many scientific collaborations, each having a different set of requirements. The flexibility would be helpful during designing/debugging, however it is of lesser value for a science archive, compared to industry with rapidly changing requirements, and a strong focus on agility.
5.5.4 Cost¶
As of now, the most popular MR framework, Hadoop, is freely available as open source. In contrast, none of the freely available RDBMSes implements a shared-nothing MPP DBMS (to date), with the exception of SciDB, which can be considered only partially relational.
From the LSST perspective, plain MR does not meet project’s need, in particular the low-volume query short response time needs. Significant effort would be required to alleviate Hadoop’s high latency (today’s solution is to run idle MR daemons, and attach jobs to them, which pushes the complexity of starting/stopping jobs onto user code). Also, table joins, typically done in reduce stage, would have to be implemented as maps to avoid bringing data for joined tables to Reducer – in practice this would require implementing a clever data partitioning scheme. The main advantages of using MR as a base technology for the LSST system include scalability and fault-tolerance, although as alluded above, these features come at a high price: inefficient use of resources (full checkpointing between each Map and each Reduce step), and triple redundancy.
5.5.5 Summary¶
The key features of an ideal system, along with the comments for both Map/Reduce and RDBMS are given in the table below.
Feature | Map/Reduce | RDBMS |
---|---|---|
Shared nothing, MPP, columnar | Implements it. | Some implement it, but only as commercial, non open source to date, except not-yet-mature SciDB. |
Overlapping partitions, needed primarily for near-neighbor queries | Nobody implements this. | Only SciDB implements this to-date. |
Shared scans (primarily for complex queries that crunch through large sets of data) | This kind of logic would have to be implemented by us. | There is a lot of research about shared scans in databases. Implemented by Teradata. Some vendors, including SciDB are considering implementing it |
Efficient use of resources Catalog/schema | Very inefficient. Started adding support, e.g., Hive, HadoopDB |
Much better than MR. Much better than in MR. |
Indexes (primarily for simple queries from public that require real time response) | Started adding support, e.g., Hive, HadoopDB | Much better than in MR. |
Open source | Hadoop (although it is implemented in Java, not ideal from LSST point of view) | No shared-nothing MPP available as open source yet except still-immature SciDB. We expect there will be several by the time LSST needs it (SciDB, MonetDB, ParAccel and others) |
5.5.6 Convergence¶
Despite their differences, the database and MR communities are learning from each other and seem to be converging.
The MR community has recognized that their system lacks built-in operators. Although nearly anything can be implemented in successive MR stages, there may be more efficient methods, and those methods do not need to be reinvented constantly. MR developers have also explored the addition of indexes, schemas, and other database-ish features.[4] Some have even built a complete relational database system[5] on top of MR.
[4] | An example of that is Hive. |
[5] | An example of that is HadoopDB |
The database community has benefited from MR’s experience in two ways:
- Every parallel shared-nothing DBMS can use the MR execution style for internal processing – while often including more-efficient execution plans for certain types of queries. Though systems such as Teradata or IBM’s DB2 Parallel Edition have long supported this, a number of other vendors are building new shared-nothing-type systems.[6] It is worth noting that these databases typically use MR-style execution for aggregation queries.
- Databases such as Greenplum (part of EMC) and Aster Data (part of Teradata since March 2011) have begun to explicitly support the MR programming model with user-defined functions. DBMS experts have noted that supplying the MR programming model on top of an existing parallel flow engine is easy, but developing an efficient parallel flow engine is very hard. Hence it is easier for the DMBS community to build map/reduce than for the map/reduce community to add full DBMS functionality.
[6] | ParAccel, Vertica, Aster Data, Greenplum, DATAllegro (now part of Microsoft), Datapuia, Exasol and SciDB |
The fact MR community is rapidly adding database/SQL like features on top of their plain MR (Tenzing, Hive, HadoopDB, etc), confirms the need for database-like features (indexes, schemas, catalogs, sql).
As we continue monitoring the latest development in both RDBMS and MR communities and run more tests, we expect to re-evaluate our choices as new options become available.
6 Design Trade-offs¶
The LSST database design involves many architectural choices. Example of architectural decisions we faced include how to partition the tables, how many levels of partitioning is needed, where to use an index, how to normalize the tables, or how to support joins of the largest tables. This chapter covers the test we run to determine the optimal architecture of MySQL-based system.
6.1 Standalone Tests¶
6.1.1 Spatial join performance¶
This test was run to determine how quickly we can do a spatial self-join
(find objects within certain spatial distance of other objects) inside a
single table. Ultimately, in our architecture, a single table represents
a single partition (or sup-partition). The test involved trying various
options and optimizations such as using different indexes (clustered and
non clustered), precalculating various values (like COS(RADIANS(decl))
),
and reordering predicates. We run these tests for all reasonable table
sizes (using MySQL and PostgreSQL). We measured CPU and disk I/O to
estimate impact on hardware. In addition, we re-run these tests on the
lsst10 machine at NCSA to understand what performance we may expect
there for DC3b. These tests are documented at
http://dev.lsstcorp.org/trac/wiki/db/SpatialJoinPerf. We found that
PostgreSQL was 3.7x slower for spatial joins over a range of row counts,
and reducing the row-count per partition to less than 5k rows was
crucial in achieving lowering compute intensity, but that predicate
selectivity could compensate for a 2-4x greater row count.
6.1.2 Building sub-partitions¶
Based on the “spatial join performance” test we determined that in order to speed up self-joins within individual tables (partitions), these partitions need to be very small, \(O(\mathrm{few~} K)\) rows. However, if we partition large tables into a very large number of small tables, this will result in unmanageable number of tables (files). So, we determined we need a second level of partitioning, which we call sub-partition on the fly. This test included:
- sub-partitioning through queries:
- one query to generate one sub-partition
- relying on specially introduced column (
subPartitionId
).
- segregating data into sub-partitions in a client C++ program, including using a binary protocol.
We timed these tests. This test is described at http://dev.lsstcorp.org/trac/wiki/db/BuildSubPart. These tests showed that it was fastest to build the on-the-fly sub-partitions using SQL in the engine, rather than performing the task externally and loading the sub-partitions back into the engine.
6.1.3 Sub-partition overhead¶
We also run detailed tests to determine overhead of introducing
sub-partitions. For this test we used a 1 million row table, measured
cost of a full table scan of such table, and compared it against
scanning through a corresponding data set partitioned into
sub-partitioned. The tests involved comparing in-memory with
disk-based tables. We also tested the influence of introducing
“skinny” tables, as well as running sub-partitioning in a client C++
program, and inside a stored procedure. These tests are described at
http://dev.lsstcorp.org/trac/wiki/db/SubPartOverhead. The on-the-fly
overhead was measured to be 18% for select *
queries, but
3600% if only one column (the skinniest selection) was needed.
6.1.4 Avoiding materializing sub-partitions¶
We tried to run near neighbor query on a 1 million row table. A starting point is 1000 sec which is ~16 min 40 sec (based on earlier tests we determined it takes 1 sec to do near neighbor for 1K row table).
The testing included:
- Running near neighbor query by selecting rows with given subChunkId into in memory table and running near neighbor query there. It took 7 min 43 sec.
- Running near neighbor query by running neighbor once for each subChunkId, without building sub-chunks. It took 39 min 29 sec.
- Running near neighbor query by mini-near neighbor once for each subChunkId, without building sub-chunks, using in-memory table. It took 13 min 13 sec.
6.1.5 Billion row table / reference catalog¶
One of the catalogs we will need to support is the reference catalog, even in DC3b it is expected to contain about one billion rows. We have run tests with a table containing 1 billion rows catalog (containing USNO-B data) to determine how feasible it is to manage a billion row table without partitioning it. These tests are described in details at: http://dev.lsstcorp.org/trac/wiki/DbStoringRefCat This revealed that single 1-billion row table usage is adequate in loading and indexing, but query performance was only acceptable when the query predicates selectivity using an index was a small absolute number of rows (1% selectivity is too loose). Thus a large fraction of index-scans were unacceptably slow and the table join speed was also slow.
6.1.6 Compression¶
We have done extensive tests to determine whether it is cost effective to compress LSST databases. This included measuring how different data types and indexes compress, and performance of compressing and decompressing data. These tests are described in details at https://dev.lsstcorp.org/trac/wiki/MyIsamCompression. We found that table data compressed only 50%, but since indexes were not compressed, there was only about 20% space savings. Table scans are significantly slower due to CPU expense, but short, indexed queries were only impacted 40-50%.
6.1.7 Full table scan performance¶
To determine performance of full table scan, we measured:
- raw disk speed with
dd if=<large file> of=/dev/zero
and got 54.7 MB/sec (2,048,000,000 bytes read in 35.71 sec) - speed of
select count(*) from XX where muRA = 4.3
using a 1 billion row table. There was no index on muRA, so this forced a full table scan. Note that we did not doSELECT *
to avoid measuring speed of converting attributes. The scan of 72,117,127,716 bytes took 28:49.82 sec, which is 39.8 MB/sec.
So, based on this test the full table scan can be done at 73% of the raw disk speed (using MySQL MyISAM).
6.1.8 Low-volume queries¶
A typical low-volume queries to the best of our knowledge can be divided into two types:
analysis of a single object. This typically involves locating a small number of objects (typically just one) with given objectIds, for example find object with given id, select attributes of a given galaxy, extract time series for a given star, or select variable objects near known galaxy. Corresponding representative queries:
SELECT * from Object where objectId=<xx> SELECT * from Source where objectId =<xx>
analysis of objects meeting certain criteria in a small spatial region. This can be represented by a query that selects objects in a given small ra/dec bounding box, so e.g.:
SELECT * FROM Object WHERE ra BETWEEN :raMin AND :raMax AND decl BETWEEN :declMin AND :declMax AND zMag BETWEEN :zMin AND :zMax
Each such query will typically touch one or a few partitions (few if the needed area is near partition edge). In this test we measured speed for a single partition.
Proposed partitioning scheme will involve partitioning each large table into a “reasonable” number of partitions, typically measured in low tens of thousands. Details analysis are done in the storage spreadsheet ([LDM-141]). Should we need to, we can partition the largest tables into larger number of smaller partitions, which would reduce partition size. Given the hardware available and our time constraints, so far we have run tests with up to 10 million row partition size.
We determined that if we use our custom spatial index (“subChunkId”), we can extract 10K rows out of a 10 million row table in 30 sec. This is too long – low volume queries require under 10 sec response time. However, if we re-sort the table based on our spatial index, that same query will finish in under 0.33 sec.
We expect to have 50 low volume queries running at any given time. Based on details disk I/O estimates, we expect to have ~200 disk spindles available in DR1, many more later. Thus, it is likely majority of low volume queries will end up having a dedicated disk spindle, and for these that will end up sharing the same disk, caching will likely help.
Note that these tests were done on fairly old hardware (7 year old).
In summary, we demonstrated low-volume queries can be answered through an index (objectId or spatial) in well under 10 sec.
6.1.9 Solid state disks¶
We also run a series of tests with solid state disks to determine where it would be most cost-efficient to use solid state disks. The tests are described in details in [Docushare-11701]. We found that concurrent query execution is dominated by software inefficiencies when solid-state devices (SSDs) with fast random I/O are substituted for slow disks. Because the cost per byte is higher for SSDs, spinning disks are cheaper for bulk storage, as long as access is mostly sequential (which can be facilitated with shared scanning). However, because the cost per random I/O is much lower for SSDs than for spinning disks, using SSDs for serving indexes, exposure metadata, perhaps even the entire Object catalog, as well as perhaps for temporary storage is advised. This is true for the price/performance points of today’s SSDs. Yet even with high IOPS performance from SSDs, table-scan based selection is often faster than index-based selection: a table-scan is faster than an index scan when >9% of rows are selected (cutoff is >1% for spinning disk). The commonly used 30% cutoff does not apply for large tables for present storage technology.
7 Risk Analysis¶
7.1 Potential Key Risks¶
Insufficient database performance and scalability is one of the major risks [Docushare-7025].
We have a prototype system (Qserv) that will be turned into a production system. Given that a large fraction of its functionality is derived from two stable, production quality, open source components (MySQL and XRootD), turning it into production system is possible during the LSST construction phase.
A viable alternative might be to use an off-the-shelf system. In fact, an off-the-shelf solution could present significant support cost advantages over a production-ready Qserv, especially if it is a system well supported by a large user and developer community. It is likely that an open source, scalable solution will be available on the time scale needed by LSST (for the beginning of LSST construction a stable beta would suffice, beginning of production scalability approaching few hundred terabytes would be sufficient). Database systems larger than the largest single LSST data set have been successfully demonstrated in production today. For example, eBay manages a 10+ petabyte production database [dbms210] and expects to deploy a 36 petabyte system later in 2011. For comparison, the largest single LSST data set, including all indexes and overheads is expected to be below 10 petabytes in size, and will be produced ~20 years from now (the last Data Release).[9] The eBay system is based on an expensive commercial DBMS (Teradata), but there is a growing demand for large scale systems and growing competition in that area (Hadoop, SciDB, Greenplum, InfiniDB, MonetDB, Caché and others).
[9] | The numbers, both for eBay and LSST are for compressed data sets. |
Finally, a third alternative would be to use a closed-source, non free software, such as Caché, InfiniDB or Greenplum (Teradata is too expensive). Some of these systems, in particular Caché and InfiniDB are very reasonably priced. We believe the largest barrier preventing us from using an off-the-shelf DBMS such as InfiniDB is spherical geometry and spherical partitioning support.
Potential problems with off-the-shelf database software used, such as MySQL is another potential risk. MySQL has recently been purchased by Oracle, leading to doubts as to whether the MySQL project will be sufficiently supported in the long-term. Since the purchase, several independent forks of MySQL software have emerged, including MariaDB (supported by one of the MySQL founders), Drizzle (supported by key architects of MySQL), and Percona. Should MySQL disappear, these open-source, MySQL-compatible[10] systems are a solid alternative. Should we need to migrate to a different DBMS, we have taken multiple measures to minimize the impact:
- our schema does not contain any MySQL-specific elements and we have successfully demonstrating using it in other systems such as MonetDB and Microsoft’s SQL Server;
- we do not rely on any MySQL specific extensions, with the exception of MySQL Proxy, which can be made to work with non-MySQL systems if needed;
- we minimize the use of stored functions and stored procedures which tend to be DBMS-specific, and instead use user defined functions, which are easier to port (only the interface binding part needs to be migrated).
[10] | With the exception of Drizzle, which introduced major changes to the architecture. |
Complex data analysis. The most complex analysis we identified so far include spatial and temporal correlations which exhibit \(O(n^2)\) performance characteristics, searching for anomalies and rare events, as well as searching for unknown are a risk as well – in most cases industrial users deal with much simpler, well defined access patters. Also, some analysis will be ad-hoc, and access patterns might be different than these we are anticipating. Recently, large-scale industrial users started to express strong need for similar types of analyses; understanding and correlating user behavior (time-series of user clicks) run by web companies, searching for abnormal user behavior to detect fraud activities run by banks and web companies, analyzing genome sequencing data run by biotech companies, and what-if market analysis run by financial companies are just a few examples. Typically these analysis are ad-hoc and involve searching for unknowns, similar to scientific analyses. As the demand (by rich, industrial users) for this type of complex analyses grows, the solution providers are rapidly starting to add needed features into their systems.
The complete list of all database-related risks maintained in the LSST risk registry:
- DM-014: Database performance insufficient for planned load
- DM-015: Unexpected database access patterns from science users
- DM-016: Unexpected database access patterns from DM productions
- DM-032: LSST DM hardware architecture becomes antiquated
- DM-060: Dependencies on external software packages
- DM-061: Provenance capture inadequate
- DM-065: LSST DM software architecture incompatible with de-facto community standards
- DM-070: Archive sizing inadequate
- DM-074: LSST DM software architecture becomes antiquated
- DM-075: New SRD requirements require new DM functionality
7.2 Risks Mitigations¶
To mitigate the insufficient performance/scalability risk, we developed Qserv, and demonstrated scalability and performance. In addition, to increase chances an equivalent open-source, community supported, off-the-shelf database system becomes available in the next few years, we initiated the SciDB array-based scientific database project and work closely with its development team. We also closely collaborate with the MonetDB open source columnar database team – building on our Qserv lessons-learned, they are trying to add missing features and turn their software into a system capable of supporting LSST needs. A demonstration is expected in late 2011. Further, to stay current with the state-of-the-art in peta-scale data management and analysis, we continue a dialog with all relevant solution providers, both DBMS and Map/Reduce, as well as with data-intensive users, both industrial and scientific, through the XLDB conference and workshop series we lead, and beyond.
To understand query complexity and expected access patterns, we are working with LSST Science Collaborations and the LSST Science Council to understand the expected query load and query complexity. We have compiled a set of common queries [common-queries] and distilled this set into a smaller set of representative queries we use for various scalability tests–this set represents each major query type, ranging from trivial low volume, to complex correlations. [perftests]. We have also talked to scientists and database developers from other astronomical surveys, including SDSS, 2MASS, Gaia, DES, LOFAR and Pan-STARRS.
To deal with unpredictability of analysis, we will use shared scans. With shared scans, users will have access to all the data, all the columns, even these very infrequently used, at a predictable cost – with shared scans increasing complexity does not increase the expensive disk I/O needs, it only increases the CPU needs.
To keep query load under control, we will employ throttling to limit individual query loads.
8 Implementation of the Query Service (Qserv) Prototype¶
To demonstrate feasibility of running LSST queries without relying on expensive commercial solutions, and to mitigate risks of not having an off-the-shelf system in time for LSST construction, we built a prototype system for user query access, called Query Service (Qserv). The system relies on two production-quality components: MySQL and XRootD. The prototype closely follows the LSST baseline database architecture described in chapter 3
8.1 Components¶
8.1.1 MySQL¶
MySQL is used as an underlying SQL execution engine. To control the scope of effort, Qserv uses an existing SQL engine, MySQL, to perform as much query processing as possible. MySQL is a good choice because of its active development community, mature implementation, wide client software support, simple installation, lightweight execution, and low data overhead. MySQL’s large development and user community means that expertise is relatively common, which could be important during Qserv’s development or long-term maintenance in the years ahead. MySQL’s MyISAM storage engine is also lightweight and well-understood, giving predictable I/O access patterns without an advanced storage layout that may demand more capacity, bandwidth, and IOPS from a tightly constrained hardware budget.
It is worth noting, however, that Qserv’s design and implementation do not depend on specifics of MySQL beyond glue code facilitating results transmission. Loose coupling is maintained in order to allow the system to leverage a more advanced or more suitable database engine in the future.
8.1.2 XRootD¶
The XRootD distributed file system is used to provide a distributed, data-addressed, replicated, fault-tolerant communication facility to Qserv. Re-implementing these features would have been non-trivial, so we wanted to leverage an existing system. XRootD has provided scalability, fault-tolerance, performance, and efficiency for over 10 years of in the high-energy physics community, and its relatively flexible API enabled its use as a more general communication medium instead of a file system. Since it was designed to serve large data sets, we were confident that it could mediate not only query dispatch communication, but also bulk transfer of results.
A XRootD cluster is implemented as a set of data servers and a redirector(s). A client connects to the redirector, which acts as a caching namespace lookup service that redirects clients to appropriate data servers. In Qserv, XRootD data servers become Qserv workers by plugging custom code into XRootD as a custom file system implementation. The Qserv master dispatches work as an XRootD client to workers by writing to partition-addressed XRootD paths and reads results from hash-addressed XRootD paths.
While the primary purpose of XRootD is to provide a distributed clustered file system, it is implemented as a plug-in component based system that allows “file” to be replaced by any other resource object. Qserv makes use of this capability to cluster MySQL databases instead of files. Hence, we dispense with calling XRootD a distributed file system and simply call is a generic system for providing named communication paths, clustering, request scheduling, and error recovery.
8.2 Partitioning¶
In Qserv, large spatial tables are fragmented into spatial pieces in the two-level partitioning scheme. The partitioning space is a spherical space defined by two angles φ (right ascension/α) and θ (declination/δ). For example, the Object table is fragmented spatially, using a coordinate pair specified in two columns–right-ascension and declination. On worker nodes, these fragments are represented as tables named Object_CC and Object_CC_SS where CC is the “chunk id” (first-level fragment) and SS is the “sub-chunk id” (second-level fragment of the first larger fragment. Sub-chunk tables are built on-the-fly to optimize performance of spatial join queries. Large tables are partitioned on the same spatial boundaries where possible to enable joining between them.
8.3 Query Generation¶
Qserv is unusual (though not unique) in processing a user query into one or more queries that are subsequently executed on off-the-shelf single-node RDBMS software. This is done in the hopes of providing a distributed parallel query service while avoiding a full re-implementation of common database features. However, we have found that it is necessary to implement a query processing framework much like one found in a more standard database, with the exception that the resulting query plans contain SQL statements as the intermediate language.
A significant amount of query analysis not unlike a database query optimizer is required in order to generate a distributed execution plan that accurately and efficiently executes user queries. Incoming user queries are first parsed into an intermediate representation using a modified SQL92-compliant grammar (Lubos Vnuk’s SqlSQL2). The resulting query representation is equivalent to the original user query, and does not include any stateful interpretation, but may not completely reflect the original syntax. The purpose of this representation is to provide a semantic representation that may be operated upon by query analysis and transformation modules without the complexity of a parse tree containing every node in the original EBNF grammar.
Once the representation has been created, the query representation is processed by two sequences of modules. The first sequence operates on the query as a single statement. A transformation step occurs to split the single representation into a “plan” involving multiple phases of execution, one to be executed per-data-chunk, and a one to be executed to combine the distributed results into final user results. The second sequence is applied on this plan to apply the necessary transformations for an accurate result.
We have found that regular expressions and parse element handlers to be insufficient to analyze and manipulate queries for anything beyond the most basic query syntax constructions.
8.3.1 Processing modules¶
The processing modules perform most of the work in transforming the user query into statements that can produce a faithful result from a Qserv cluster. These include:
- Identify spatial indexing opportunities. This allows Qserv to dispatch spatially-restricted queries on only a subset of the available chunks constituting a table. Spatial restrictions given in Qserv-specific syntax are rewritten as boolean SQL clauses.
- Identify secondary index opportunities. Qserv databases designate one column (more are under consideration) as a key column where its values are guaranteed to exist in one spatial location. Identification allows Qserv to convert point queries on this column into spatial restrictions.
- Identify table joins and generate syntax to perform distributed join results. Qserv primarily supports “near-neighbor” spatial joins for limited distances defined in the partitioning coordinate space. Arbitrary joins between distributed tables are only supported using the key column. Classify queries according to data coverage and table scanning. By identifying tables scanned in a query, Qserv is able to mark queries for execution using shared scanning, which greatly increases efficiency.
8.3.2 Processing module overview¶
This figure illustrates the query preparation pipeline that generates physical queries from an input query string. User query strings are parsed (1) into a structured query representation that is passed through a sequence of processing modules (2) that operate on that representation in-place. Then, it is broken up (3) into pieces that are explicitly intended for parallel execution on table partitions and pieces intended to merge parallel results into user results. Another processing sequence (4) operates on this new representation, and then finally, concrete query strings are generated (5) for execution.
The two sequences of processing modules provide an extensible means to implement query analysis and manipulation. Earlier prototypes performed analysis and manipulation during parsing, but this led to a practically unmaintainable code base and the functionality has been ported the processing module model. Processing is split into two sequences to provide the flexibility to modules that manipulate the physical structures while offering the simpler single-query representation to modules that do not require the complexity. The clear separation between parsing, whose only goal is to provide a intelligible and modifiable query representation, and the qserv-specific analysis and manipulation is a key factor in the overall flexibility, maintainability, and extensibility of the system and should help the system adapt to current and future LSST needs.
8.4 Dispatch¶
The baseline Qserv uses XRootD as a distributed, highly-available communications system to allow Qserv frontends to communicate with data workers. Up until 2015, Qserv used a synchronous client API with named files as communication channels. The current baseline system utilizes a general two-way named-channeling system which eliminates explicit file abstractions in favor of generalized protocol messages that can be flexibly streamed. The scheme is called Scalable Service Interface (`SSI`_) and is built on top of XRootD. The interface was specifically designed to hide underlying XRootD dependencies. This allows switching the underlying implementation with minimal impact to Qserv.
8.4.1 Wire protocol¶
Qserv encodes query dispatches in ProtoBuf messages, which contain SQL statements to be executed by the worker and annotations that describe query dependencies and characteristics. Transmitting query characteristics allows Qserv workers to optimize query execution under changing CPU and disk loads as well as memory considerations. The worker need not re-analyze the query to discover these characteristics or guess at conditions that cannot be determined by query inspection.
Query results are also returned by ProtoBuf messages. Current versions transmit a MySQL dump file allowing the query results to be faithfully reproduced on the Qserv frontend, but the baseline system will transmit results directly. Initial implementations avoided logic to encode and decode data values, but experience with the prototype MonetDB worker backend proved that data encoding and marshalling were a contained problems whose solution could significantly improve overall query latency by avoiding mutating metadata operations on worker and frontend DBMS systems. Thus the baseline system will encode results in protobuf messages containing schema and row-by-row encoded data values. Streaming results directly from worker dbms instances into frontend dbms instances is a technique under consideration, as is a custom aggregation engine for results that would likely ease the implementation of providing partial query results to end users.
8.4.2 Frontend¶
ABH In 2012, a new XRootD client API was developed to address our concerns over the older version’s scalability (uncovered during a 150 node, 30TB scalability test). The new client API began production use for the broader XRootD community in late 2012. Subsequently, work began under our guidance towards an XRootD Qserv client API that was based on request-response interaction over named channels, instead of opening, reading, and writing files. A production version of this API, the Scalable Service Interface (`SSI`_) became available in early 2015 and Qserv has since been ported to use this interface. The port eliminated a significant body of code that maps dispatching and result-retrieving to file operations. The `SSI`_ API will reside in the Xroot code base, where it may be exercised by other projects.
The `SSI`_ API provides Qserv with a fully asynchronous interface that eliminates nearly all blocking threads used by the Qserv frontend to communicate with its workers. This eliminated one class of problems we have encountered during large-scale testing. The `SSI`_ API has defined interfaces that integrate smoothly with the Protobufs-encoded messages used by Qserv. Two novel features were specifically added to improve Qserv performance. The streaming response interface enables reduced buffering in transmitting query results from a worker mysqld to a the frontend, which lowers end-to-end query latency and reduces storage requirements on workers. The out-of-band meta-data response which arrives prior to the data results can be used to map out the Protobufs encoding and significantly simplify handling response memory buffers.
The fully asynchronous API is crucial on the master because of the large number of concurrent chunk queries in flight expected in normal operation. For example, with the sky split into 10k pieces, having 10 full-scanning queries running concurrently would have 100k concurrent chunk queries–too large a number of threads to allow on a single machine. Hence, an asynchronous API to XRootD is crucial. Threads are used to parallelize multiple CPU-bound tasks. While it does not seem to be important to parse/analyze/manipulate a single user query in parallel (and such a task would be a research topic), the retrieval and processing of results could be done in parallel if some portion of the aggregation/merging were done in Qserv code rather than loaded into the frontend’s MySQL instance and merged via SQL queries. Thus results processing should be parallelized among results from individual chunks, and query parsing/analysis/manipulation can be parallelized among independent user queries.
8.4.3 Worker¶
The Qserv worker uses both threads and asynchronous calls to provide concurrency and parallelism. To service incoming requests from the XRootD API, an asynchronous API is used to receive requests and enqueue them for action. Specifically, the Scalable Service Interface (`SSI`_) is used on Qserv workers as well. The interface provides a mirror image of the actions taken on the front-end making the logic relatively easy to follow and the implementation less error prone.
Threads are maintained in a thread pool to perform incoming queries and wait on calls into the DBMS’s API (currently, the apparently synchrnous MySQL C-API). Threads are allowed to run in observance of the amount of parallel resources available. The worker estimates the I/O dependency of each incoming chunk query in terms of the chunk tables involved and disk resources involved, and attempts to ensure that disk access is almost completely sequential. Thus if there are many queries that access the same table chunk, the worker allows as many of them to run as there are CPU cores in the system, but if it has many queries that involve different chunk tables, it allows fewer simultaneous chunk queries in order to ensure that only one table scan per disk spindle occurs. Further discussion of this “shared scanning” feature is described in 8.10 Shared Scans.
8.5 Threading Model¶
Nearly every portion of Qserv is written using a combination of threaded and asynchronous execution.
Qserv heavily relies on multi-threading to take advantage of all available CPU cores when executing queries, as an example, to complete one full table scan on a table consisting of 1,000 chunks, 1,000 queries (processes) will be executed. To efficiently handle large number of processes that are executed on each worker, we ended up rewriting the XRootD client and switching from thread-per-request model to a thread-pool model. The new client is completely asynchronous, with real call-backs.
mysqlproxy | Single-threaded Lua code |
Frontend-python | Single-threaded asynchronous reactor; blocking-thread per user query |
Frontend-C++ | Processing thread per user-query for preparation; Results-merging thread-per-user-query on-demand; |
Frontend-xrootd | Callback threads perform query transmission and results retrieval |
Frontend-xrootd internal | Threads for maintaining worker connections (< 1 per host) |
Xrootd, cmsd | Small thread pools for managing live network connections and performing lookups |
Worker-xrootd plugin | Small thread pool O(#cores) to make blocking mysql C-API calls into local mysqld; callback threads from XRootD perform admission/scheduling of tasks from frontend and transmission of results |
8.6 Aggregation¶
Qserv supports several SQL aggregation functions: AVG(), COUNT(), MAX(), MIN(), and SUM(), and should support SQL92 level GROUP BY.
8.7 Indexing¶
The secondary index utilizes one or more tables using the InnoDB storage engine on each czar to perform lookups on the database’s key colume (objectId). Performance tests (Figure fig-indexing_tests) on a single, dual-core host with 1 TB hard disk storage (not SSD) have shown that this configuration will support a full load of 40 billion rows in about 400,000 seconds (110 hours). A more realistic configuration with multiple cores and SSD storage is expected to meet the requirement of fully loading in less than 48 hours.
To improve the performance of the InnoDB storage engine for queries, the secondary index may be split across a small number (dozens) of tables, each containing a contiguous range of keys. This splitting, if done, will be independent of the partitioning of the database itself. The contiguity of key ranges will allow the secondary index service to identify the appropriate split table arithmetically via an in-memory lookup.
8.7.1 Secondary Index Structure¶
The secondary index consists of three columns: the key (objectId), the chunk where all data with that key are located (chunkId), and the subchunk within that chunk where data with the key are located (subChunkId). The objectId is assigned by the science pipelines as a 64-bit integer value; the chunkId and subChunkId are both 16-bit integers which identify spatial regions on the sky.
8.7.2 Secondary Index Loading¶
The InnoDB storage engine loads tables most efficiently if it is provided input data which has been presorted according to the table’s primary key. When the secondary index information is collected for loading (from each worker node handling a collection of chunks), it is sorted by objectId, and may be divided into roughly equal “splits”. Each of those splits is loaded into a table en masse.
To fully optimize the loading and table splitting, the entire index should be collected from all workers and pre-sorted in memory on the czar. This is not reasonable for 40 billion entries (requiring a minimum of 480 GB memory, plus overhead). Instead, the index data from a single worker can be assumed to be a “representative sample” from the full range of objectIds, so table splitting can be done using the first worker’s index data. The remaining workers will be split and loaded according to those defined tables.
8.8 Data Distribution¶
LSST will maintain the released data store both on tape media and on a database cluster. The tape archive is used for long-term archival. Three copies of the compressed catalog data will be kept. The database cluster will maintain 3 online copies of the data. Because computer clusters of reasonable size failure regularly, the cluster must maintain replicas in order to provide continuous data access. A replication factor of 3 (r=3) is needed in order to determine data integrity by majority rule when one replica is corrupt.
If periodic unplanned downtime is acceptable, an on-tape replica may function as one of the three. However, the use of tape dramatically increases the cost of recovering from a failure. This may be acceptable for some tables, particularly those that are large and lesser-used, although allowing service disruption may make it difficult to make progress on long-running analysis on those large tables.
8.8.1 Database data distribution¶
The baseline database system will provide access for two database releases: latest and previous . Data for each release will be spread out among all nodes in the cluster.
Data releases are partitioned spatially, and spatial pieces (chunks) are distributed in a round-robin fashion across all nodes. This means that area queries involving multiple chunks are almost guaranteed to involve resources on multiple nodes.
Each node should maintain at least 20% free space of its data storage allocation. The remaining free space is then available to be “borrowed” when another node fails. This will a temporary use of storage capacity until more server resources can be put online, until the 80% storage use is returned.
8.8.2 Failure and integrity maintenance¶
There will be failures in any large cluster of node, in the nodes themselves, in data storage volumes, in networks access and so on. These failures will remove access to data that is resident on those nodes, but this loss of data access should not affect that ability of scientists to analyze the dataset as a whole. We need to set a data availability time over 99.5% to ensure confidence of the community in the stability of the system. To ensure this level of data access, and to allow acceptable levels of node failures in a cluster, there will be replication of data on a table level throughout the cluster.
The replication level will be that each table in the database will exist 3 times, each on separate nodes. A monitoring layer to the system will check on the availability of each table every few hours, although this time will be tuned in practice. When this layer sees that a table has less than three replicas available, this will initiate a replication of that table to another nodes, not currently hosting that table. The times for the checking, and speed of replication will be tuned to the stability of the cluster, such that about 5% of all tables at any given time will only have 1 or 2 replicas. Three replicas will ensure that tables will be available even in cases of large failures, or when nodes need to be migrated to new hardware in bulk.
Should an entire node fail, replicating that data to another single node would be fairly expensive in terms of time. As of July 2013, a 3TB drive will have a write speed of 60/150/100 MB/s (min/max/avg), [http://www.legitreviews.com/article/2092/3/] and refilling this single drive would remove access to that replica of the data for about 8 hours. We plan on having free space on each node, and only fill local storage to 80%. The free space will be used for temporary storage of tables on failures, where replicas can take place in parallel between nodes into this free space. When new nodes with free storage are added to the cluster, then this data can be copied off this free space into the drive, taking the full 8 hours, but there will still be 3 replicas of data during this time. Once this is complete, this data will have 4 replicas for the short period of time while these tables can be removed from the temporary storage, returning each node to 80% usage.
8.9 Metadata¶
Qserv needs to track various metadata information, static (not changing or changing very infrequently), and dynamic (run-time) in order to maintain schema and data integrity and optimize the cluster usage.
8.9.1 Static metadata¶
Qserv typically works with databases and tables distributed across many different machines: it breaks individual large tables into smaller chunks and distributes them across many nodes. All chunks that belong to the same logical table must have the same schema and partitioning parameters. Different tables often need to be partitioned differently, for example some tables might be partitioned with overlap (such as the Object table), some might be partitioned with no overlap (for example the Source table), and some might not need partitioning at all (e.g., a tiny Filter table). Further, there might be different partitioning strategies, such as spherical-box based, or HTM-based. All this information about schema and partitioning for all Qserv-managed databases and tables needs to be tracked and kept consistent across the entire Qserv cluster.
Implementation of the static metadata in Qserv is based on hierarchical key-value storage which uses a regular MySQL database as a storage backend. This database is shared between multiple masters and it must be served by a fault-taulerant MySQL server instance, e.g. using a master-master replication solution like MariaDB Galera cluster. Database consistency is critical for metadata and it should be implemented using one of the transactional database engines in MySQL.
Static metadata may contain following information:
- Per-database and per-table partitioning and scan scheduling parameters.
- Table schema for each table, used to create database tables in all worker and master instances; the schema in the master MySQL instance can be used to obtain the same information when a table is already created.
- Database and table state information, used primarily by the process of database and table creation or deletion.
- Definitions for the set of worker and master nodes in a cluster including their availability status.
The main clients of the static metadata are:
- Administration tools (command-line utilities and modules) which allow one to define or modify metadata structures.
- Qserv master(s), mostly querying partitioning parameters but also allowed to modify table/database status when deleting/creating new tables and databases. Master(s) should not depend on node definitions in metadata, the xrootd facility is used to communicate with workers.
- Special “watcher” service which implements distributed process of database and table management.
- An initial implementation of the data loading application which will use the node definitions and will create/update database and table definitions. This initial implementation will eventually be replaced by a distributed loading mechanism which may be based on separate mechanisms.
8.9.2 Dynamic metadata¶
In addition to static metadata, a Qserv cluster also needs to track its current state, and keep various statistics about query execution. This sort of data is udated frequently, several times per query execution, and is called dynamic metadata.
Prototype implementation of the dynamic metadata is based on MySQL database. Like static metadata it needs to be shared between all master instances and will be served via a single fault-talerant MySQL instance which will be shared with static metadata database.
Dynamic metadata will contain the following information:
- Definition of every master instance in a Qserv cluster.
- Record of every SELECT-type query processed by cluster. This record includes query processing state and some statistical/timing information.
- Per-query list of table names used by the asynchronous queries, this information is used to delay table deletion while async queries are in progress.
- Per-query worker information, which includes chunk id and identifying information for the worker processing that chunk id. This information will be used to transparently restart the master or migrate query processing to a different master in case of master failure.
The most significant use of the dynamic metadata is to track execution of asyncronous queries. When an async query is submitted it is registered in dynamic metadata and its ID is returned to the user immediately. Later users can request status information for that query ID which is obtained from dynamic metadata. When query processing is finished users can request results from that query, and the master can obtain the location of the result data from dynamic metadata.
Additionally dynamic metadata can be used to collect statistical information about queries that were executed in the past which may be an important tool in understanding and improving system performance.
8.9.3 Architecture¶
The Qserv metadata system is implemented based on master/server architecture: the metadata is centrally managed by a Qserv Metadata Server (qms). The information kept on each worker is kept to a bare minimum: each worker only knows which databases it is supposed to handle, all remaining information can be fetched from the qms (through Qserv) as needed. This follows our philosophy of keeping the workers as simple as possible.
The real-time metadata is managed inside qms in in-memory tables, periodically synchronized with disk-based table. Such configuration allows reducing qms latency—important to avoid delaying query execution time. Should a qms failure occur, the in-flight queries for which the information was lost will be restarted. Since the synchronization to disk-based table will occur relatively frequently (eg. at least 1 per minute), the lost time is insignificant. To avoid overloading the qms with, only the high-level information available from Qserv-master is stored in qms; all worker-based information is cached in a scratch space locally to each worker in a simple, raw form (e.g, key-value, ASCII file), and can be fetched on demand as needed.
At the moment we use xml-rpc as a message protocol to communicate with qms. It was a natural choice given that this protocol is already in use by Qserv master.
8.9.4 Typical Data Flow¶
Static metadata:
- Parts of the static metadata known before data is partitioned/loaded are loaded by the administration scripts responsible for loading data into the database, then these scripts start data partitioner.
- The data partitioner reads static metadata loaded by the administration scripts, loads remaining information.
- When Qserv starts, it fetches all static metadata and caches it in memory in a special, in-memory optimized C++ structure.
- The contents of the in-memory metadata cache inside Qserv can be refreshed on demand if the static metadata changes (for example, when a new database or a table is added).
Dynamic-metadata:
- Master loads the information for each query (when it starts, when it completes).
- Detailed statistics are dumped by each worker into a scratch space kept locally. This information can be requested from each worker on demand. A typical use case: if all chunk-queries except one completed, qms would fetch statistics for the still-running chunk-query to estimate when the query might finish, whether to restart this query etc.
8.11 Level 3: User Tables, External Data¶
Level 3 tables including tables generated by users, and data catalogs brought from outside, depending on their type and size, will be either partitioned and distributed across the production database servers, or kept unpartitioned in one central location. While the partitioned and distributed Level 3 data will share the nodes with Level 2 data, it will be kept on dedicated disks, independent from the disks serving Level 2 data. This will simplify maintenance and recoverability from failures.
Level 3 tables will be tracked and managed through the Qserv Metadata System (qms), described in the Metadata chapter above. This includes both the static, as well as the dynamic metadata.
8.12 Cluster and Task Management¶
Qserv delegates management of cluster nodes to XRootD. The XRootD system manages cluster membership, node registration/de-registration, address lookup, replication, and communication. Its Scalable Service Interface (`SSI`_) API provides data-addressed communication channels to the rest of Qserv, hiding details like node count, the mapping of data to nodes, the existence of replicas, and node failure. The Qserv manager focuses on dispatching queries to endpoints and Qserv workers focus on receiving and executing queries on their local data.
Cluster management performed outside of XRootD does not directly affect query execution, but include coordinating data distribution, loading, nodes joining/leaving and is discussed in 8.15 Administration. The `SSI`_ API includes methods that allow dynamic updates to the data view of an XRootD cluster. So that when new tables appear or disappear, the XRootD system will incorporate that information for future scheduling decisions. Thus, clusters can dynamically change without the need to restart the XRootD system.
8.13 Fault Tolerance¶
Qserv approaches fault tolerance in several ways. The design exploits the immutability of the underlying data by replicating and distributing data chunks across a cluster such that in the event of a node failure, the problem can be isolated and all subsequent queries re-routed to nodes maintaining duplicate data. Moreover, this architecture is fundamental to Qserv’s incremental scalability and parallel performance. Within individual nodes, Qserv is highly modularized with minimal interdependence among its components, which are connected via narrow interfaces. Finally, individual components contain specialized logic for minimizing, handling, and recovering from errors.
The components that comprise Qserv include features that independently provide failure-prevention and failure-recovery capabilities. The MySQL proxy is designed to balance its load among several underlying MySQL servers and provide automatic fail-over in the event a server fails. The XRootD system provides multiple managers and highly redundant servers to provide high bandwidth, contend with high request rates, and cope with unreliable hardware. And the Qserv master itself contains logic that works in conjunction with XRootD to isolate and recover from worker-level failures.
A worker-level failure denotes any failure mode that can be confined to one or more worker nodes. In principle, all such failures are recoverable given the problem nodes are identified and alternative nodes containing duplicate data are available. Examples of such failures include a disk failure, a worker process or machine crashing, or network problems that render a worker unreachable.
Consider the event of a disk failure. Qserv’s worker logic is not equipped to manage such a failure on localized regions of disk and would behave as if a software fault had occurred. The worker process would therefore crash and all chunk queries belonging to that worker would be lost. The in-flight queries on its local mysqld would be cleaned up and have resources freed. The Qserv master’s requests to retrieve these chunk queries via XRootD would then return an error code. The master responds by re-initializing the chunk queries and re-submits them to XRootD. Ideally, duplicate data associated with the chunk queries exists on other nodes. In this case, XRootD silently re-routes the request(s) to the surviving node(s) and all associated queries are completed as usual. In the event that duplicate data does not exist for one or more chunk queries, XRootD would again return an error code. The master will re-initialize and re-submit a chunk query a fixed number of times (determined by a parameter within Qserv) before giving up, logging information about the failure, and returning an error message to the user in response to the associated query.
Error handling in the event that an arbitrary hardware or software bug (perhaps within the Qserv worker itself) causes a worker process or machine to crash proceeds in the same manner described above. The same is true in the event that network loss or transient sluggishness/overload has the limited effect of preventing XRootD from communicating with one or more worker nodes. As long as such failures are limited to a finite number of workers and do not extend to the Qserv master node, XRootD is designed to record the failure and return an error code. Moreover, if duplicate data exists on other nodes, this will be registered within XRootD, which will successfully route any subsequent chunk queries.
In the event of an unrecoverable error, the Qserv master is equipped with a status/error messaging mechanism designed to both log detailed information about the failure and to return a human-readable error message to the user. This mechanism includes C++ exception handling logic that encapsulates all of the master’s interactions with XRootD. If an unrecoverable exception occurs, the master gracefully terminates the query, frees associated resources, logs the event, and notifies the user. Qserv’s internal status/error messaging system also generates a status message and timestamp each time an individual chunk query achieves a milestone. Such milestones include: chunk query dispatch, written to XRootD, results read from XRootD, results merged, and query finalized. This real-time status information provides useful context in the event of an unrecoverable error.
Building upon the existing fault-tolerance and error handling features described above, future work includes introducing a heart-beat mechanism on worker nodes that periodically pings the worker process and will restart it in the event it becomes unresponsive. Similarly, a master monitoring process could periodically ping worker nodes and restart a worker machine if necessary. We are also considering managing failure at a per-disk level, but this would require research since application-level treatment of disk failure is relatively rare. It should also be possible to develop an interface for checking the real-time status of queries currently being processed by Qserv by leveraging its internally used status/error messaging mechanism.
8.14 Next-to-database Processing¶
We expect some data analyses will be very difficult, or even impossible to express through SQL language. This might be particularly useful for time-series analysis. For this type of analyses, we will allow users to execute their analysis algorithms in a procedural language, such as Python. To do that, we will allow users to run their own code on their own hardware resources co-located with production database servers. Users then run queries on the production database which stream rows directly from database cluster nodes to the user processing cluster, where arbitrary code may run without endangering the production database. This allows their incurred database I/O needs to be satisfied using the database system’s shared scanning infrastructure while providing the full flexibility of running arbitrary code.
8.15 Administration¶
8.15.1 Installation¶
Qserv as a service requires a number of components that all need to be running, and configured together. On the master node we require mysqld, mysql-proxy, XRootD, cmsd, qserv metadata service, and the qserv master process. On each of the worker nodes there will also be the mysqld, cmsd, and XRootD service. These major components come from the MySQL, XRootD, and Qserv distributions. But to get these to work together we will also require many more software package, such as protobuf, lua, expat, libevent, python, zope, boost, java, antlr, and so on. And many of these require more recent versions than you are provided in most system distributions. We have an installation layer, developed by SLAC, and LPC in Clermont-Ferrand, France in collaboration, which will determine the packages, configure, compile and install them in an automated process.
Currently, the Qserv installation procedure supports only the official LSST platform— RHEL6, and SL6 Linux distributions. Other UNIX-like systems will be supported in the future as needed. The Qserv package first can be downloaded from SLAC for install. In the initial README there are basic install procedures, which start with a bootstrap script, that will perform a yum install of needed packages distributed with RHEL6, where the versions will support the Qserv install. Once that is done an install script can be started. This will first download needed packages not shipped with RHEL6 from SLAC, and get those installed first. All software will be installed into a sandbox root path, and all installed by the production username. Along with this is an install of MySQL from source that will be configured for Qserv. These further packages will be configured to run together, and then Qserv will be complied and linked to these installed packages. All this runs without user interaction, and usually completes within 15 to 20 minutes, to provide a complete Qserv either master or worker node.
8.15.2 Data loading¶
As previously mentioned, Data Release Production will not write directly to the database. Instead, the DRP pipelines will produce binary FITS tables and image files that are reliably archived as they are produced. Data will be loaded into Qserv in bulk for every table, so that tables are either not available, or complete and immutable from the user query access perspective.
For replicated tables, these FITS files are converted to CSV (e.g. by harvesting FITS image header keyword value pairs, or by translating binary tables to ASCII), and the resulting CSV files are loaded directly into MySQL and indexed. For partitioned tables like Object and Source, FITS tables are fed to the Qserv partitioner, which assigns partitions based on sky coordinates and converts to CSV.
In particular, the partitioner divides the celestial sphere into latitude angle “stripes” of fixed height H. For each stripe, a width W is computed such that any two points in the stripe with longitudes separated by at least W have angular separation of at least H. The stripe is then broken into an integral number of chunks of width at least W, so that each stripe contains a varying number of chunks (e.g. polar stripes will contain just a single chunk). Chunk area varies by a factor of about pi over the sphere. The same procedure is used to obtain subchunks: each stripe is broken into a configurable number of equal-height “substripes”, and each substripe is broken into equal-width subchunks. This scheme is preferred over the Hierarchical Triangular Mesh for its speed (no trigonometry is required to locate the partition of a point given in spherical coordinates), simplicity of implementation, and the relatively fine control it offers over the area of chunks and sub-chunks.
The boundaries of subchunks constructed as described are boxes in longitude and latitude - the overlap region for a subchunk is defined as the spherical box containing all points outside the subchunk but within the overlap radius of its boundary.
The task of the partitioner is to find the IDs of the chunk and subchunk containing the partitioning position of each row, and to store each row in the output CSV file corresponding to its chunk. If the partitioning parameters include overlap, then the row’s partitioning position might additionally fall inside the overlap regions of one or more subchunks. In this case, a copy of the row is stored for each such subchunk (in overlap CSV files).
Tables that are partitioned in Qserv must be partitioned identically
within a Qserv database. This means that chunk tables in a database
share identical partition boundaries and identical mappings of chunk id
to spatial partition. In order to facilitate table joining, a single
table’s columns are chosen to define the partitioning space and all
partitioned tables (within a related set of tables) are either
partitioned according that pair of columns, or not partitioned at all.
Our current plan chooses the Object table’s ra_PS
and decl_PS
columns, meaning that rows in the Source and ForcedSource tables will be
partitioned according to the Objects they reference.
There is one exception: we allow for pre-computed spatial match tables. As an example, such a table might provide a many-to-many relationship between the LSST Object catalog and a reference catalog from another survey, listing all pairs of LSST Objects and reference objects separated by less than some fixed angle. The reference catalog cannot be partitioned by associated Object, as more than one Object might be matched to a reference object. Instead, the reference catalog must be partitioned by reference object position. This means that a row in the match table might refer to an Object and reference object assigned to different chunks stored on different Qserv worker nodes.
We avoid this complication by again exploiting overlap. We mandate (and verify at partitioning time) that no match pair is separated by more than the overlap radius. When partitioning match tables, we store a copy of each match in the chunk of both positions referenced by that match. When joining Objects to reference objects via the match table then, we are guaranteed to find all matches to Objects in chunk C by joining with all match records in C and all reference objects in C or in the overlap region of C.
All Qserv worker nodes will partition subsets of the pipeline output files in parallel – we expect partitioning to achieve similar aggregate I/O rates to those of full table scans for user query access, so that partitioning should complete in a low factor (2-3x) of the table scan time. Once it does, each Qserv worker will gather all output CSV files for its chunks and load them into MySQL. The structure of the resulting chunk tables is then optimized to maximize performance of user query access (chunk tables will likely be sorted, and will certainly be compressed), and appropriate indexes are built. Since chunks are sized to fit in memory, all of these steps can be performed using an in-memory file-system. I/O costs are incurred only when reading the CSV files during the load and when copying finalized tables (i.e. .MYD/.MYI files) to local disk.
The last phase of data loading is to replicate each chunk to one other Qserv worker node. We will rely on table checksum verification rather than a majority rule to determine whether a replica is corrupt or not.
The partitioner has been prototyped as a multi-threaded C++ program. It uses an in-memory map-reduce implementation internally to scale across cores, and can read blocks of one or more input CSV files in parallel. It does not currently understand FITS table files. CSV file writes are also parallelized - each output chunk is processed by a single reducer thread and can be written to in parallel with no application level locking. In preliminary testing, our partitioner was able to sustain several hundred MB/s of both read and write bandwidth when processing a CSV dump of the PT1.2 Source table.
We are investigating a pair of data loading optimizations. One is to have pipeline processes either integrate the partitioning code or feed data directly to the partitioner, rather than communicating via persistent storage. The other is to write out tables in the native database format (e.g. as .MYD files, ideally using the MySQL/MariaDB server code to do so), allowing the CSV database loading step to be bypassed.
8.15.3 Administrative scripts¶
The administration of the qserv cluster will require a set of scripts, all run from the one master machine, to control the large set of workers. The main admin script, qserv-admin, will supply the base needs, with starting all processes needed for the service, in order, and taking down all processes to stop the service. Also base monitoring of service is supplied here, to report on processes that are running, and responding to base queries, to check on MySQL or XRootD dying or locking up. Also is supplied is the updating of the configuration definitions from the master out to all workers, such that all machines need to have the same configurations for the services.
The base data loading onto the nodes tends to be a slightly detailed process, beyond the just the data preparation. Up to now, data preparation produces text files in csv format, and then these will be loaded into the MySQL layer as a MyISAM table. The schema for these tables will need to have added to them the fields for chunk and subchunk number needed for the Qserv service. The modification of the schema and the control of the loading of the data, which can take hours, is done with the qserv-load script. The loading of the data is also done without index creation, and then that is done after the data loading. We are also experimenting with the use of compressed read-only tables for the data serving, and this is an option.
Another needed setup for the data service in qserv, is the creation of the “emptyChunks” list. The data will be spatially partitioned into “chunks”, as previously described, but the for the complete service, the master process with need to know how many chunks exist in the data, and which of these chunks contain no data. In queries which will involve a complete table scan, which chunks to create query, or not, will need to be known. Once the data is loaded, there is a another script which will go out to nodes and see what chunks are there, and compile a list of all possible chunks and which chunks do not contain data, or the “emptyChunks” lists. This is loaded by the qserv master process at startup.
8.16 Result Correctness¶
To verify Qserv does not introduce any unexpectedly alter results (e.g., does not show the same object twice or does not miss any objects on the chunk boundaries), we developed an automated testbed, which allows us to run pre-set queries on pre-set data sets both through plain MySQL and through Qserv, and compare results.
8.17 Current Status and Future Plans¶
As of now (June 2013) we have implemented a basic version of the system end-to-end. Our prototype is capable of parsing a wide range of queries, including queries executed by our QA system, “PipeQA”, rewriting them into sub-queries, executing these sub-queries in parallel and returning results to the user. The implementation includes a generic parser, basic query scheduler, job executor, query result collector. We demonstrated running all query types (low, high, super-high such as large-area near-neighbor) including aggregations, scalably on a 150-node cluster using 30 TB data set; and a smaller subset of queries scalably on 300-node cluster (remaining tests in progress, expecting to complete in the next 2-3 weeks) we also demonstrated the system performs well enough to meet the LSST query response time requirements. We demonstrated the system can handle high-level of concurrency (10 concurrent queries simultaneously accessing 10,000 chunks each). We demonstrated the system can recover from a variety of faults, or at minimum gracefully fail if the error is unrecoverable. We extended SQL syntax coverage and ensured the system is capable of supporting all types of queries executed over the course of recent data challenges by PipeQA and users. We implemented a core foundation for the metadata, currently used for managing static metadata about Qserv-managed databases and tables, a set of administrative tools, and scalable data partitioner. We made the system easy to set up, resilient to typical failures and common user mistakes. We implemented automated test bed. We consider the current prototype to have a quality of a typical late-alpha / early-beta software.
Future work includes:
- extending metadata to support run-time statistics, implementing query management tools
- implementing support for Level 3 data
- completing initial shared scan implementation, testing and implementing concurrent and synchronized shared scans on multiple spindles
- demonstrating cross-match with external catalogs
- improving interfaces for users (eg hiding internal tables)
- re-examining and improving query coverage, including more advanced SQL syntax, such as sub-queries as needed
- improvements to administration scripts
- support for HTM partitioning in Qserv
- authentication and authorization
- resource management
- early partition results
- performance improvements
- partition granularity varying per table
- security
Extending metadata to support run-time statistics, implementing query management tools. Qserv currently does not maintain any explicit run-time system state. Keeping such state would simplify managing Qserv cluster, and building features such as query management: currently there are no tools for inspecting and managing queries in-flight, and there are no interfaces for halting queries except upon error detection. It is clear that users and administrators will need to list running queries, check query status and possibly abort queries.
Implementing support for Level 3 data. Qserv will need to support level 3. That means users should be able to maintain their own tables to store their own data or results from previous queries. They should be able to create, drop, and update their own tables within the system.
Completing initial shared scan implementation, testing and implementing concurrent and synchronized shared scans on multiple spindles. The first prototype implementation of shared scanning is mostly complete, with the remaining work focused on basic analysis and characterization of incoming user queries to determine scanning tables and plumbing to convey the appropriate hints to worker nodes.
Demonstrating cross-match with external catalogs. One of the use cases involves cross matching with external catalogs. In case the catalogs to cross-match with is small, it will be treated as a small table and replicated as metadata tables will be. For cross-matching with larger catalogs, the catalog to cross-match with will need to be partitioned and distributed on the worker nodes.
Improving interfaces for users. Many admin-type commands such as
“list processes” or “explain” are not ported to the distributed Qserv
architecture, and thus will not show correct result. At the moment we
have disabled these commands. Additionally, commands such as listing
tables in a given database will have to be overloaded, for example, we
should show user a table “Object” (even though in practice such table
does not exist in the Qserv system), instead of all the chunk
Object_XXX
tables, that are internal, and should not be exposed to the
end-user.
Re-examining and improving query coverage, including more advanced SQL syntax, such as sub-queries as needed. We examined what queries users and production processes execute, however we realize this query set is far from the complete list of queries we will see in the future. All needed syntax needs to be understood and fully supported. Design and feasibility evaluation for sub-query support. Qserv does not support SQL sub-queries. Since there is evidence that such a capability might be useful to users, so we should formulate a few possible designs and understand how easy/difficult they would be to implement. Note that there are some alternative viable alternatives, such as splitting sub-queries into multiple queries, and/or using session variables. A naïve implementation that involves dumping all sub-query results to disk and then reading these results from disk, similarly to how multiple map/reduce stages are implemented, should be tractable to implement.
Improvements to administration scripts. To further automate common tasks related to database management, table management, partition management, data distribution, and others we need to implement many improvements to the administration scripts.
Support for HTM partitioning in Qserv. HTM is an alternative to the rectangular box form of spatial partitioning currently implemented in Qserv. Since HTM allows for more advanced indexing and optimization, it may eventually replace the current partitioning algorithm.
Authentication and authorization. The current Qserv does not implement any form of security or privileges. All access is full access. A production database system should provide some facility of user or role-based access so that usage can be controlled and resources can be shared. This is in particular needed for Level-3 data products.
Resource management. A production system should have some way to manage/restrict resource usage and provide quality-of-service controls. This includes a monitoring facility that can track each node’s load and per-user-query resource usage.
Early partition results. When performing interactive exploration of an observational data set, users frequently issue large-scale queries that produce undesired results, even after testing such queries on small subsets of the data. We can ameliorate this behavior by providing the investigator with early partial results from the query, allowing the user to recognize that the returned values are incorrect and permitting the query to be aborted without wasting additional resources. There are two mechanisms we will implement in Qserv for providing early results. First, for queries that retrieve a filtered set of rows, matching rows can be returned as their query fragments complete, well before all fragments finish. Second, for queries that group, sort, or aggregate information and therefore perform a global operation after any per-partition processing, the global operation can be applied to increasingly large subsets of the per-partition results, returning an early partial result each time.
Performance improvements. Significant performance gain can be obtained by improving scheduler. These improvements pose interesting state of the art computing challenges; more details are available in Appendix D. In addition, some parts of Qserv are inefficient since they were implemented under constraints of development time rather than efficiency, or maintainability – rewriting them would result in further performance gains. Caching results for future queries is another example of performance optimization that can yield significant speed improvements.
Partitioning granularity varying per table. Since large tables in LSST vary significantly in row count and row size, it may be worthwhile to support partitioning with multiple granularities. For execution management it is useful to have partitions sized so that query fragments have similar execution cost. To achieve this, partitions may need different spatial sizes.
Security. The system needs to be secure and resilient against denial of service attacks.
8.18 Open Issues¶
What follows is a (non-exhaustive) list of issues, technical and scientific, that are still being discussed and where changes are possible.
- Support for updates. Size of Level 1 catalog is relatively small, and the expected query access patterns are relatively non-challenging, thus currently do not envision any need to deploy scalable Qserv-like architecture for Alert Production. Should this change, we will need to support updates in Qserv, which will likely have some non-trivial impact on the architecture.
- Very large objects. Some objects (eg, large galaxies) are much larger than our overlap region, in some cases their footprint will span multiple chunks. Currently we are working with the object center, neglecting the actual footprint. While there are some science use cases that would benefit from a system that tracks objects based on their footprint, this is currently not a requirement. Potential solution would involve adding a custom index similar to the r-tree-based indexes such as the TOUCH [67].
- Very large results. Currently, the front-end that disptached the query is responsible for assembling the results. In general, this is not a scalable approach as the resources required to processes the results may be several orders of magnitude greater than those needed to dispatch the query. One solution is to replicate the front-end to the extent necessary to handle query results. Alternatively, the Scalable Service Interface can be augmented to allow running disconnected queries. That is, once a particular front-end dispatches a query it can get a handle to that query and disconnect from it. Another server can, using that handle, reconnect to the query and process the results. This is a more flexible model as it allows independent scaling of query dispatch and result processing. It also has the aded benefit of not cancelling in-progress queries dispatched by a particulr fron-end should that front-end die.
9 Large-scale Testing¶
9.1 Introduction¶
9.1.1 Ideal environment¶
Based on the detailed spreadsheet analysis, we expect the ultimate LSST production system will be composed of few hundred database servers [LDM-144], so a realistic test should include a cluster of at least 100 nodes.
Total database size of a single data release will vary from ~1.3 PB (DR1) to ~15 PB (DR11).[12] Realistic testing requires at least ~20-30 TB of storage (across all nodes).
[12] | These numbers are for single copy, data and indices, compressed when appropriate. |
Note that a lot of highly focused tests which are extremely useful to fine tune different aspects of the system can be done on a very small, 2-3 cluster, or even on a single machine. An example of that can be measuring the effect of table size on the performance of near-neighbor join: this type of join will be done per sub-partition, and sub-partitions will be small (few K rows), thus almost all tests involving a single sub-partition can be done on a single machine with very little disk storage.
A significant amount of testing should be done where the dataset size exceeds the system memory size by an order of magnitude. This testing is important to reveal system performance in the presence of disk performance characteristics.
It is essential to have at least two different types of catalogs: Object and Source. Of course this data needs to be correlated, that is, the objects should corresponds to the sources. Having these 2 tables will allow us to measure speed of joins. It is not necessary to have other types of source-like tables (DiaSource, ForcedSource) – the tests done with Source should be a good approximation.
The most important characteristic of the test data is its spatial distribution. The data should reflect realistic densities: presence of very crowded or very sparse regions have influence on how data is partitioned and on performance of certain queries (e.g., speed of near neighbor inside one partition). Other than realistic spatial distribution, we need several fields to be valid (e.g., magnitudes) in order to try some queries with predicates.
These tests are not only used to prove our system is capable of meeting the requirements, but also as a mean to stress the system and uncover potential problems and bottlenecks. In practice, whoever runs these tests should well understand the internals of the scalable architecture system and turning MySQL.
9.1.2 Schedule of testing¶
- Selecting the base technology – Q2 2009
- Determining the architecture – Q3 2009
- Pre-alpha tests focused on parallel distributed query execution engine, tests at small scale (~20 nodes) – Q4 2009
- Most major features in (except shared scan, user tables), performance tests on mid-size cluster (~100 nodes) – Q4 2010
- Scalability and performance improvements and tests on a large cluster (150-250 nodes) – Q4 2011
- Large scale tests, performance tests on a large cluster (250-300 nodes) – Q2 2013
- Shared scans – Q3 2013
- Fault tolerance – Q4 2013
9.1.3 Current status of tests¶
We have run several large scale tests
- (10/2009) A test with the “pre-alpha” version of our software written on top of MySQL, using the Tuson cluster at LLNL (160 nodes, each node: two Intel Xeon 2.4 GHz GPUs with 4 GB RAM and 1 local hard disk of 80 Gbs)
- (2010) several 100-node tests run at SLAC [58]. These tests helped us uncover many bottlenecks and prompted rewriting parts of our software, as well as implementing several missing features in XRootD.
- (4/2011) A 30 TB test on 150-node SLAC cluster using Qserv in 40/100/150 node configurations, using 2 billion row Object and 32 billion row Source tables, total of 30 TB data set.
- (12/2012) A 100-terabyte test on JHU’s Datascope 20-node cluster. Due to high instability of the cluster this test turned into testing resilience to faults of Qserv and the associated administrative tools.
- (05/2013) A 10,000-chunk test on a 120-node SLAC cluster to reproduce and understand subtle issues with concurrency at scale.
- (07/2013) A test on 300-node IN2P3 cluster to re-test scalability, performance, concurrency, and uncover unexpected issues and bottlenecks.
- (08/2013) A demonstration of shared scans.
The tests 3-6 listed above are described in further details below.
9.2 150-node Scalability Test¶
9.2.1 Hardware¶
We configured a cluster of 150 nodes interconnected via gigabit Ethernet. Each node had 2 quad-core Intel Xeon X5355 processors with 16GB memory and one 500GB 7200RPM SATA disk. Tests were conducted with Qserv SVN r21589, MySQL 5.1.45 and XRootD 3.0.2 with Qserv patches.
9.2.2 Data¶
We tested using a dataset synthesized by spatially replicating the dataset from the LSST data challenge (“PT1.1”). We used two tables: Object and Source.[13] These two tables are among the largest expected in LSST. Of these two, the Object table is expected to be the most frequently used. The Source table will have 50-200X the rows of the Object table, and its use is primarily confined to time series analyses that generally involve joins with the Object table.
[13] | The schema may be browsed online at http://lsst1.ncsa.uiuc.edu/schema/index.php?sVer=PT1_1 |
The PT1.1 dataset covers a spherical patch with right-ascension between 358˚ and 5˚ and declination between -7˚ and 7˚. This patch was treated as a spherical rectangle and replicated over the sky by transforming duplicate rows’ RA and declination columns, taking care to maintain spatial distance and density by a non-linear transformation of right-ascension as a function of declination. This resulted in an Object table of 1.7 billion rows (2TB) and a Source table of 55 billion rows (30 TB).[14] The Source table included only data between -54˚ and +54˚ in declination. The polar portions were clipped due to limited disk space on the test cluster. Partitioning was set for 85 stripes each with 12 sub-stripes giving a φ height of ~2.11˚ for stripes and 0.176˚ for sub-stripes. Each chunk thus spanned an area of approximately 4.5deg2, and each sub-chunk, 0.031deg2. This yielded 8,983 chunks. Overlap was set to 0.01667˚ (1 arc-minute).
[14] | Source for the duplicator is available at http://dev.lsstcorp.org/trac/browser/DMS/qserv/master/trunk/examples |
9.2.3 Queries¶
The current Qserv development focus is on features for scalability. We have chosen a set of test queries that demonstrate performance for both cheap queries (interactive latency), and expensive queries (hour, day latency). Runs of low volume queries ranged from 15 to 20 queries, while runs of high volume queries and super high volume queries consisted of only a few or even one query due to their expense. All reported query times are according to the command-line MySQL client (MySQL).
9.2.3.1 Low-volume 1 – object retrieval¶
SELECT * FROM Object WHERE objectId = <objId>
In Figure 6 we can see that performance of this query is roughly constant, taking about 4 seconds. Each run consisted of 20 queries. The slower performance of Runs 1 and 4, where each execution took 9 seconds, were probably the result of competing tasks in the cluster. We attribute the initial 8 second execution time in Run 5 and beyond to cold cache conditions (likely the objectId index) in the cluster.
9.2.3.2 Low-volume 2 – time series¶
SELECT taiMidPoint, fluxToAbMag(psfFlux), fluxToAbMag(psfFluxErr), ra, decl
FROM Source
WHERE objectId = <objId>
This query retrieves information from all detections of a particular astronomical object, effectively providing a time-series of measurements on a desired object. For testing, the objectId was randomized as for the Low Volume 1 query, which meant that null results were retrieved where the Source data was missing due to available space on the test cluster.
In Figure 7 we see that performance is roughly constant at about 4 seconds per query. Run 1 was done after Low Volume 1’s Run 1 and we discount its 9 second execution times similarly as anomalous.
9.2.3.3 Low-volume 3 – spatially-restricted filter¶
SELECT COUNT(*)
FROM Object
WHERE ra_PS BETWEEN 1 AND 2
AND decl_PS BETWEEN 3 AND 4
AND fluxToAbMag(zFlux_PS) BETWEEN 21 AND 21.5
AND fluxToAbMag(gFlux_PS)-fluxToAbMag(rFlux_PS)
BETWEEN 0.3 AND 0.4
AND fluxToAbMag(iFlux_PS)-fluxToAbMag(zFlux_PS)
BETWEEN 0.1 AND 0.12;
In Figure 8 we see the same 4 second performance that was seen for the other low volume queries. Again, the ~9 second performance in Run 2 could not be reproduced so we discount it as resulting from competing processes on the cluster.
9.2.3.5 High-volume 2 – full-sky filter¶
SELECT objectId, ra_PS, decl_PS, uFlux_PS, gFlux_PS,
rFlux_PS, iFlux_PS, zFlux_PS, yFlux_PS
FROM Object
WHERE fluxToAbMag(iFlux_PS) - fluxToAbMag(zFlux_PS) > 4
Using the on-disk data footprint (MySQL’s MyISAM .MYD, without indexes or metadata) of the Object table (1.824x1012 bytes), we can compute the aggregate effective table scanning bandwidth. Run 3’s 7 minute execution yields 4.0GB/s in aggregate, or 27MB/s per node, while the other runs yield approximately 11GB/s in aggregate, or 76MB/s per node. Since each node was configured to execute up to 4 queries in parallel, Run 3’s bandwidth is more realistic, given seek activity from competing queries and the disk manufacturer’s reported theoretical transfer rate of 98MB/s.
9.2.3.6 High-volume 3 – density¶
SELECT COUNT(*) AS n, AVG(ra_PS), AVG(decl_PS), chunkId
FROM Object
GROUP BY chunkId
This query computes statistics for table fragments (which are roughly equal in spatial area), giving a rough estimate of object density over the sky. It illustrates more complex aggregation query support in Qserv. This query is of similar complexity to High Volume 2, but Figure 11 illustrates measured times significantly faster, which is probably due to reduced results transmission time. As mentioned for HV2, cache behavior was not controlled, but the 4 minute time in Run 3 may be close.
9.2.3.7 Super-high-volume 1 – near neighbor¶
SELECT COUNT(*)
FROM Object o1, Object o2
WHERE qserv_areaspec_box(-5,-5,5,-5)
AND qserv_angSep(o1.ra_PS, o1.decl_PS,
o2.ra_PS, o2.decl_PS) < 0.1
This query finds pairs of objects within a specified spherical distance which lie within a particular part of the sky. Over two randomly selected 100 deg2 areas, the execution times were about 10 minutes (667.19 seconds and 660.25 seconds). The resultant row counts ranged between 3 to 5 billion. Since execution uses on-the-fly generated tables, the tables do not fit in memory, and Qserv does not yet implement caching, we expect caching effects to be negligible.
9.2.3.8 Super-high-volume 2 – sources not near objects¶
SELECT o.objectId, s.sourceId, s.ra, s.decl,
o.ra_PS, o.decl_PS
FROM Object o, Source s
WHERE qserv_areaspec_box(224.1, -7.5, 237.1, 5.5)
AND o.objectId = s.objectId
AND qserv_angSep(s.ra, s.decl, o.ra_PS, o.decl_PS) > 0.0045
This is an expensive query – an \(O(kn)\) join over 150 square degrees between a 2TB table and a 30TB table. Each objectId is unique in Object, but is shared by 41 rows (on average) in Source, so \(k \sim 41\). We recorded times of a few hours (5:20:38.00, 2:06:56.33, and 2:41:03.45). The variance is presumed to be caused by varying spatial object density over the three random areas selected.
9.2.4 Scaling¶
We tested Qserv’s scalability by measuring its performance while varying the number of nodes in the cluster. To simulate different cluster sizes, the frontend was configured to only dispatch queries for partitions belonging to the desired set of cluster nodes. This varies the overall data size proportionally without changing the data size per node (200-300GB). We measured performance at 40, 100, and 150 nodes to demonstrate weak scaling.
9.2.4.1 Scaling with small queries¶
From Figure 12 — Figure 14, we see that execution time is unaffected by node count given that the data per node is constant. The spike in the 40-node configuration in Figure 14 is caused by 2 slow queries (23s and 57s); the other 28 executed in times ranging from 4.09 to 4.11 seconds.
9.2.4.2 Scaling with expensive queries¶
High Volume
If Qserv scaled perfectly linearly, the execution time should be constant when the data per node is constant. In Figure 15 the times for high volume queries show a slight increase. HV1 is a primarily a test of dispatch and result collection overhead and its time increases linearly with the number of chunks since the front-end has a fixed amount of work to do per chunk. Since we varied the set of chunks in order to vary the cluster size, the execution time of HV1 should thus vary linearly with cluster size. HV3 seems to have a similar trend since due to cache effects – its result was cached so execution became more dominated by overhead.
The High Volume 2 query approximately exhibits the flat behavior that would indicate perfect scalability. Caching effects may have clouded the results, but they did not dominate. If the query results were perfectly cached, we expect the overall execution time to be dominated by overhead as in HV1, and this is clearly not the case.
Super High Volume
The tests on expensive queries did not show perfect scalability, but nevertheless, the measurements did show some amount of parallelism. It is unclear why execution in the 100-node configuration was the slowest for both SHV1 and SHV2. Our time-limited access to the cluster did not allow us to repeat executions of these expensive queries and study their performance in better detail.
9.2.5 Concurrency¶
We were able to test Qserv with multiple queries in flight. We ran 4 “streams” of queries: two parallel invocations of HV2, one of LV1, and one of LV2. Each low volume stream paused for 1 second between queries. Figure 12 illustrates concurrent performance. We see that the HV2 queries take about twice the time (5:53.75 and 5:53.71) as they would if running alone. This makes sense since each is a full table scan that is competing for resources and shared scanning has not been implemented. The first queries in the low volume streams execute in about 30 seconds, but each of their second queries seems to get “stuck” in queues. Later queries in the streams finish faster. Since the worker nodes maintain first- in-first-out queues for queries and do not implement any concept of query cost, long queries can easily hog the system. The slowness of low volume queries after the second queries may be curious at first glance, since they should be queued at the end on their assigned worker nodes and thus complete near the end of the HV2 queries. In that case, subsequent queries would land on workers with nearly empty queues and execute immediately. This slowness can be explained by query skew – short queries may land on workers that have or have not finished their work on the high volume queries.
9.2.6 Discussion¶
9.2.6.1 Latency¶
LSST’s data access needs include supporting both small, frequent, interactive queries and longer, hour/day-scale queries. We designed Qserv to operate efficiently in both cases to avoid needing multiple systems, which would be costly in development, maintenance, and hardware. Indexing was implemented in order to reduce latency for cheap queries that only touch a small part of the data.
The current Qserv implementation incurs significant overhead in dispatching queries and collecting results. In early development we decided to minimize the intelligence on each worker, so the front-end master became responsible for preparing the SQL queries so that workers did not need to perform parsing or variable substitution. Results collection is somewhat heavyweight as well. MySQL does not provide a method to transfer tables between server instances, so tables are dumped to SQL statements using mysqldump and reloaded on the front-end. This method was chosen to speed prototyping, but its costs in speed, disk, network, and database transactions are strong motivations to explore a more efficient method.
9.2.6.2 Solid-state storage¶
Some of Qserv’s design choices (e.g., shared scanning) are motivated by the need to work around poor seek performance characteristics of disks. Solid-state storage has now become a practical alternative to mechanical disk in many applications. While it may be useful for indexes, its current cost differential per unit capacity means that it is still impractical to store bulk data. In the case of flash storage, the most popular solid-state storage technology, shared scanning is still effective in optimizing performance since DRAM is much faster than flash storage and flash still has “seek” penalty characteristics (though it is much better than spinning disk).
9.2.6.3 Many core¶
We expect the performance to be I/O constrained, since the workload is data, not CPU performance limited. It is unlikely that many cores can be leveraged on a single node since they will be sized with only the number of disk spindles that saturate the north bridge, but shared scanning should increase CPU utilization efficiency.
9.2.6.4 Alternate partitioning¶
The rectangular fragmentation in right ascension and declination, while convenient to visualize physically for humans, is problematic due to severe distortion near the poles. We are exploring the use of a hierarchical scheme, such as the hierarchical triangular mesh [Kunszt] for partitioning and spatial indexing. These schemes can produce partitions with less variation in area, and map spherical points to integer identifiers encoding the points’ partitions at many subdivision levels. Interactive queries with very small spatial extent can then be rewritten to operate over a small set of fine partition IDs. If chunks are stored in partition ID order, this may allow I/O to occur at below sub-chunk granularity without incurring excessive seeks. Another bonus is that mature, well tested, and high-performance open source libraries exist for computing the partition IDs of points and mapping spherical regions to partition ID sets.
9.2.6.5 Distributed management¶
The Qserv system is implemented as a single master with many workers. This approach is reasonable and has performed adequately in testing, but the bottlenecks are clear. A Qserv instance at LSST’s planned scale may have a million fragment queries in flight, and while we have plans to optimize the query management code path, managing millions from a single point is likely to be problematic. The test data set described in this paper is partitioned into about 9,000 chunks, which means that a launch of even the most trivial full-sky query launches about 9,000 chunk queries.
One way to distribute the management load is to launch multiple master instances. This is simple and requires no code changes other than some logic in the MySQL Proxy to load-balance between different Qserv masters. Another way is to implement tree-based query management. Instead of managing individual chunk queries, the master would dispatch groups of them to lower-level masters which would could either subdivide and dispatch subgroups or manage the individual chunk queries themselves.
9.3 100-TB Scalability Test (JHU 20-node cluster)¶
In the fall of 2012, we were provided the opportunity to use a cluster of computers at John Hopkins University (JHU), that were large memory, multi-processor computers, each with very large storage attached, to setup as a Qserv service. There were 21 nodes provided for us, and the nodes had two processors with 12 cores each of Intel Xeon X5650 CPUs at 2.67GHz. But mounted as a data volume, were 22TB raid arrays, to provide a possible 450TB of storage. This would provide a high volume storage test, but with a low number of compute nodes, with one master node, and 20 worker nodes.
The data used would be produced on each node, starting with a test dataset called “pt12”. This test dataset was 220GB, but high density data in one particular spot in the sky, a few degrees wide. We chose a high density spot of this data, and then duplicated with across the whole sphere, to provide a high density dataset over the whole sky, yielding an estimated 100TB of data.
The production of this large amount of data proved to have problems. The production was rather slow, taking many weeks for a full production on each node. At first long processes were setup, and the stability of the cluster was an issue, with processes dying after days of running, and many smaller production processes were setup to get past stability issues. This produced over 100TB of csv text files, into about 7000 chunks worth of data. Once that was done, then this data was loaded into MySQL MyISAM tables, and with the large data sizes this also took days.
Over the course of this time, often nodes would go out with problems, and be down for some amount of time, before coming back. Often this would be with the data still on the mounted volumes, but the loss of computing nodes would set back the time until the data would be complete. But this was a small problem, than the problem of data stability. Once all the nodes were up and running, the data service was still having problems. This was found to be either loss or corruption of a few of the thousands of tables on the various nodes. With loss of data, either data would be re-created or just blocked for testing. Over the course of testing, dealing with data corruption was constant issue, but still a large percentage would be accessible at least. Also a problem was file corruption on the install software, and a few nodes needed to be reinstalled over the course of the testing. A full 100TB of data was generated, but only about 85TB could ever be served, before the resources had to be given back.
But some testing was able to get done on this cluster. A full table scan was performed on the Object data, although this was only on the aprox, 85TB of data was served, not the complete 100TB of data that was generated. The query was performed:
SELECT count(*) FROM Object
+------------+
| count(*) |
+------------+
| 2059335968 |
+------------+
1 row in set (19.26 sec)
Showing 2B objects in the table. The result here was after the query was performed a few times, and the cache had been stabilized. This is similar to the times found from previous testing.
The low volume test from access to a small portion of the sky was also performed. Using the query:
SELECT count(*)
FROM Object
WHERE qserv_areaspec_box(1,3,2,4)
AND scisql_fluxToAbMag(zFlux_PS) BETWEEN 21 AND 21.5
+----------+
| count(*) |
+----------+
| 748 |
+----------+
1 row in set (4.45 sec)
This time for access is also similar to previous testing, looking for the number of object in a small part of the sky of a certain color. The ~4.5 sec. overhead here is a baseline overhead for data access in this version of the qserv software.
A high volume data test was performed, looking for color information on records within a certain range of color. This will scan over all objects, to return certain number of records. The test query was:
SELECT objectId, ra_PS, decl_PS, uFlux_PS, gFlux_PS,
rFlux_PS,iFlux_PS, zFlux_PS, yFlux_PS
FROM Object
WHERE scisql_fluxToAbMag(iFlux_PS) -
scisql_fluxToAbMag(zFlux_PS) > 4
This query returned 15695 records in 6 min 33.50 sec. Again this query was performed a number of times, and this time is the average time after the caches had stabilized. This query was performed again, this time looking at a lower number of records, looking for the difference between i and z flux of 5 this time. This query returned 2967 records, in 6 min 14.0 sec. The time was a little lower this time, which was mostly the time to print the records to the screen, where the rest of the time was the over-head in scanning the available object data to return these records. The previous tests were done on 30TB of data, but using 150 nodes, although these nodes had many less cores. But this test would return in about 180 sec there, where here it is about 375 sec. The extra time here will come from the access of larger amounts of data per node, and amount of data in general, and the access rate of the data storage.
9.4 Concurrency Tests (SLAC 100,000 chunk-queries)¶
A previous version of the Qserv master code had dedicated two fixed size thread pools to each query, one for dispatching chunk queries, and the other for reading back results. The dispatch pool was sized at a quite high 500 threads for two reasons. Firstly, one goal was to dispatch work as quickly as possible, allowing the Qserv workers to prioritize as they know best. Secondly, the first query dispatch against a chunk takes ~5s, so that cluster cold start latency on a full table scan of ~10,000 chunks takes approximately 100s with this many dispatch threads. Subsequently, XRootD caching allows for near instantaneous dispatches in comparison.
The thread pool for result reads was given a much smaller size: just 20 threads per query. This is because the Qserv master process can only exploit limited amounts of parallelism when merging worker results for a single query. In fact, the main benefit of using a number as high as 20 threads is that it reduces result merge latency when chunk query execution times (and hence result availability times) are skewed.
An unfortunate consequence of this simple design was that running too many concurrent queries would cause thread creation failures in the Qserv master process. We therefore changed to a unified query dispatch and result reader thread pool model.
To test our ability to handle many concurrent full-table scan queries without running out of threads, we partitioned the PT1.2 Object table into ~8,000 chunks, and distributed them across an 80 node cluster at SLAC. The nodes in this cluster were quite old and had limited quantities of RAM, making them the perfect workhorses for this sort of test. In particular, asking for more than ~1000 threads would cause Qserv master failure on this platform. Using the new unified thread-pool design we were able to successfully run between 2 and 12 concurrent Object table scans each involving ~8,000 chunk queries, requiring a total execution time of 2 to 8 minutes, thus demonstrating that the Qserv master can handle loads of ~100,000 in-flight chunk queries, even on very old hardware.
Note that using a unified thread pool for result reads requires special measures to avoid query starvation. A single query can easily require 10,000 result reads and there will be far fewer total threads in the pool. As a result, we must be careful to avoid assigning all threads to a single query, or queries that should be interactive can easily become decidedly non-interactive as they wait for a table scan to finish. Our unified thread pool implementation therefore assigns available threads to the query using the fewest worker threads, and makes sure to create new threads when a new query is encountered (up to some hard limit).
To test this, we setup Qserv with a single master node and a single
worker node. The worker was configured with ~12,000 empty chunk tables.
We then submitted both full-table scans (SELECT COUNT(*) FROM
Object
), and an interactive query (SELECT COUNT(*) FROM Object WHERE
objectId IN (1)
) requiring just a single chunk query to answer. Though
we were able to demonstrate that the master immediately allocated
threads to dispatch the lone interactive chunk query and read back its
results, the execution time of the interactive query was still far
higher than it should have been. It turns out this is because the Qserv
worker uses a FIFO chunk query scheduling policy, and the single chunk
query corresponding to the interactive user query was being queued up
behind a multitude of chunk queries from the full table scan on the
worker side. We are currently working to address this deficiency as part
of ongoing work on shared scans.
9.5 300-node Scalability Test (IN2P3 300-node cluster)¶
The largest test we run to-date was run during July-September of 2013 on a 300-node cluster at the IN2P3 center in France. The main purpose of the test was to test Qserv scalability and performance beyond 150 nodes, and re-check concurrency at scale.
9.5.1 Hardware¶
Test machines were quad-core Intel(R) Xeon L5430 running at 2.66GHz speed, with small local spinning disks (120GB of usable space), and 16GB of RAM per machine. We received an allocation of 320 nodes, which was intended to allow for a 300 node test in spite of some failed nodes (and indeed, 17 nodes failed during the tests!)
9.5.2 Data¶
With only 120GB of available storage per node, only a limited amount of data could get produced. We tuned our data synthesizer to produce 50GB of table data per node, giving 15TB of aggregate data. 220GB of LSST PT1.2 data was synthesized into a dense stripe covering the declination range -9˚ to +9˚, setting the partitioning to 120 stripes and 9 substripes (1.5˚ x 1.5˚ chunks and 10 arc-minute-sided subchunks). This yielded 3,000 chunks of data, with 9 to 11 chunks of data on each node. The Object table had 0.4 billion rows, and the Source table had 14 billion rows. Data partitions for the Source table averaged 4GB.
Data synthesis took a couple hours for the Object table and overnight for the Source table. Recent work on the installation software enabled data loading ingest to happen within a couple hours (comparing to ~6 days for the 150-node test we run two years earlier.)
9.5.3 Software stability issues identified¶
The initial Qserv installation did not function for queries involving 300 nodes, even though subsets involving 10, 50, 100, and 150 nodes functioned properly. The first culprit was the use of an older XRootD release that was missing recent patches for a particular client race condition. Another culprit was instability exacerbated by excessive use of threads in the original threading model that the testing in section 9.5 was to address. This was addressed by re-tuning relevant threading constants. The new XRootD client has alleviated this problem.
9.5.4 Queries¶
9.5.4.1 Low volume – object retrieval¶
SELECT * FROM Object WHERE objectId = <id>
End-to-end user execution time averaged at 1.1 seconds.
9.5.4.2 Low volume – small area search¶
SELECT count(*)
FROM Object
WHERE qserv_areaspec_box(1,3,2,4) AND
scisql_fluxToAbMag(zFlux_PS) BETWEEN 21 AND 21.5
This average query response time was 1.3 sec. This is roughly the minimum end-to-end execution time for query that selects small region for this version of Qserv.
9.5.4.3 Low volume – small area search with additional conditions¶
SELECT count(*) FROM Object
WHERE qserv_areaspec_box(1,2,3,4)
AND scisql_fluxToAbMag(zFlux_PS) BETWEEN 21 AND 21.5
AND scisql_fluxToAbMag(gFlux_PS)-
scisql_fluxToAbMag(rFlux_PS) BETWEEN 0.3 AND 0.4
AND scisql_fluxToAbMag(iFlux_PS)-
scisql_fluxToAbMag(zFlux_PS) BETWEEN 0.1 AND 0.12
This average query response time was 1.3 sec. The extra CPU expense of the conditions was insignificant.
9.5.4.4 Low volume – simple join¶
SELECT s.ra, s.decl
FROM Object o
JOIN Source s USING (objectId)
WHERE o.objectId = 142367243760566706
AND o.latestObsTime = s.taiMidPoint
This returned in an average time of 11.2 sec.
9.5.4.5 High volume – object count¶
SELECT COUNT(*) FROM Object
End-to-end user execution time averaged 7.8 seconds. This is the minimum
overhead to dispatch queries for all 3,000 chunks to all 300 nodes and
retrieve their results. A condition-less COUNT(*)
is executed as a
metadata lookup by MySQL when using MyISAM tables, involving almost no
disk I/O.
Similar query executed on the Source table returned in 11.9 seconds.
9.5.4.6 High volume – full Object scans¶
SELECT COUNT(*) FROM LSST.Object WHERE gFlux_PS>1e-25
This query was repeated with different constants in the filtering
condition, and the execution time did not vary significantly – it
returned in an average time of 8.45 sec – or less than 1 second longer
than the condition-less COUNT(*)
query.
SELECT objectId, ra_PS, decl_PS, uFlux_PS, gFlux_PS,
rFlux_PS,iFlux_PS, zFlux_PS, yFlux_PS
FROM Object
WHERE scisql_fluxToAbMag(iFlux_PS)-
scisql_fluxToAbMag(zFlux_PS)>4
Varying the flux difference filter in a range of 4-5, the execution time ranged between 7-9 seconds.
SELECT objectId, ra_PS, decl_PS,
scisql_fluxToAbMag(zFlux_PS)
FROM LSST.Object
WHERE scisql_fluxToAbMag(zFlux_PS) BETWEEN 25 AND 26
End-to-end execution time ranged from 7.7 to 8.4 seconds.
9.5.4.7 High volume – Source table scan¶
SELECT objectId
FROM Source
JOIN Object USING(objectId)
WHERE qserv_areaspec_box(1,3,2,4)
This returned in 9 min 42.9 sec. Some portion of this time was spent printing the results to the screen (this test utilized a standard MySQL command-line client).
9.5.4.8 Super-high volume – near neighbor¶
SELECT COUNT(*) FROM Object o1, Object o2
WHERE qserv_areaspec_box(-5,-5,5,5)
AND scisql_angSep(o1.ra_PS, o1.decl_PS,
o2.ra_PS, o2.decl_PS) < 0.1
This query finds pairs of objects within a specified spherical distance which lie within a large part of the sky (100 deg2 area). The execution times was 4 min 50 sec. The resultant row counts was ~7 billion.
9.5.5 Scaling¶
We run a subset of the above queries on different number nodes (50, 100, 250, 200, 250, 300), in “week scaling” configuration, to determine how our software scales.
9.5.6 Discussion¶
We showed linear scalability of the dispatch – see Figure 18, achieving below 10 sec (12 for Source catalog) times when run on the entire, 300 node cluster. Queries that touch all chunks on all clusters are required to complete under an hour, so 10-12 sec overhead is very low. During previous large scale tests we run on 150 nodes 2 years ago, we were getting ~4 sec overhead. During this test, we measured 3.3 sec on 150-node configuration, which indicates we reduced the overhead, however since hardware used for these two tests was not the same, direct comparison would not be entirely fair.
We showed the overhead for simple, interactive queries was on the order of 1.8 sec when dispatching a query on one of the 300-nodes (see Figure 19). Yes, we can observe a non linearity starting from ~200 nodes, however that non-linearity is on the order of 0.03 second when going from 200 to 300 nodes. Since we are required to answer interactive queries under 10 sec, the <20% overhead is already acceptable, though we are planning to reduce it further in the future.
We were able to run all interactive-type queries well under required 10 second, with the exception of simple Object-Source join, which tool 11.2 sec. The longer-than-ideal time is attributed to unnecessary materialization of subchunks for every query that involves a join – this is expected to be optimized and alleviated in the near future.
More complex queries, such as a query that selects from a mid-size region showed linear scalability as well (Figure 20). The one time 6-sec “jump” between 100 and 150 node test is attributed to switching to different number of chunks: as we reduced the size of the cluster from 150 to 100 nodes, we excluded some chunks that were previously falling inside searched region.
We were also able to run complex queries, such as full table scans and near neighbor queries, and did not observe any anomalies.
It is important to note that due to the ratio of data size to RAM, a large fraction of the data, in particular for the “small” Object catalog was cached in memory. Such environment was particularly good for testing dispatch and result returning overheads, however it would be unfair to approximate observed performance to production-size data sets, especially given that we also had a smaller number of chunks (3,000 in the test vs expected 20,000 in production).
10 Other Demonstrations¶
10.2 Fault Tolerance¶
To prove Qserv can gracefully handle faults, we artificially triggered different error conditions, such as corrupting random parts of a internal MySQL files while Qserv is reading them, or corrupting data sent between various components of the Qserv (e.g., from the XRootD to the master process).
10.2.1 Worker failure¶
These tests are meant to simulate worker failure in general, including spontaneous termination of a worker process and/or inability to communicate with a worker node.
When a relevant worker (i.e. one managing relevant data) has failed prior to query execution, either 1) duplicate data exists on another worker node, in which case XRootD silently routes requests from the master to this other node, or 2) the data is unavailable elsewhere, in which case XRootD returns an error code in response to the master’s request to open for write. The former scenario has been successfully demonstrated during multi-node cluster tests. In the latter scenario, Qserv gracefully terminates the query and returns an error to the user. The error handling of the latter scenario involves recently developed logic and has been successfully demonstrated on a single-node, multi-worker process setup.
Worker failure during query execution can, in principle, have several manifestations.
- If XRootD returns an error to the Qserv master in response to a request to open for write, Qserv will repeat request for open a fixed number (e.g. 5) of times. This has been demonstrated.
- If XRootD returns an error to the Qserv master in response to a write, Qserv immediately terminates the query gracefully and returns an error to the user. This has been demonstrated. Note that this may be considered acceptable behavior (as opposed to attempting to recover from the error) since it is an unlikely failure-mode.
- If XRootD returns an error to the Qserv master in response to a request to open for read, Qserv will attempt to recover by re-initializing the associated chunk query in preparation for a subsequent write. This is considered the most likely manifestation of worker failure and has been successfully demonstrated on a single-node, multi-worker process setup.
- If XRootD returns an error to the Qserv master in response to a read, Qserv immediately terminates the query gracefully and returns an error to the user. This has been demonstrated. Note that this may be considered acceptable behavior (as opposed to attempting to recover from the error) since it is an unlikely failure-mode.
10.2.2 Data corruption¶
These tests are meant to simulate data corruption that might occur on disk, during disk I/O, or during communication over the network. We simulate these scenarios in one of two ways. 1) Truncate data read via XRootD by the Qserv master to an arbitrary length. 2) Randomly choose a single byte within a data stream read via XRootD and change it to a random value. The first test necessarily triggers an exception within Qserv. Qserv responds by gracefully terminating the query and returning an error message to the user indicating the point of failure (e.g. failed while merging query results). The second test intermittently triggers an exception depending on which portion of the query result is corrupted. This is to be expected since Qserv verifies the format but not the content of query results. Importantly, for all tests, regardless of which portion of the query result was corrupted, the error was isolated to the present query and Qserv remained stable.
10.2.3 Future tests¶
Much of the Qserv-specific fault tolerance logic was recently developed and requires additional testing. In particular, all worker failure simulations described above must be replicated within a multi-cluster setup.
10.3 Multiple Qserv Installations on a Single Machine¶
Once in operations, it will be important to allow multiple qserv instances to coexist on a single machine. This will be necessary when deploying new Data Release, or for testing new version of the software (e.g., MySQL, or Qserv). In the short term, it is useful for shared code development and testing on a limited number of development machines we have access to. We have successfully demonstrated Qserv have no architectural issues or hardcoded values such as ports or paths that would prevent us from running multiple instances on a single machine.
11 References¶
[LDM-141] | (1, 2, 3, 4) LSST Data Management Storage Sizing and I/O Model, Docushare LDM-141 (spreadsheet) and LDM-139 (explanation). |
[common-queries] | (1, 2) Common User Queries, http://dev.lsstcorp.org/trac/wiki/dbQueries |
[Dean04] | (1, 2) MapReduce: Simplified Data Processing on Large Clusters – Google, Jeffrey Dean, Sanjay Ghemawat, OSDI‘04 |
[Bigtable06] | (1, 2) Bigtable: A Distributed Storage System for Structured Data, http://labs.google.com/papers/bigtable-osdi06.pdf |
[Chubby] | (1, 2) Google Chubby: http://labs.google.com/papers/chubby.html |
[Pike] | Interpreting the Data: Parallel Analysis with Sawzall, Rob Pike, Sean Dorward, Robert Griesemer, Sean Quinlan, Scientific Programming Journal |
[Dremel] | (1, 2, 3) Dremel: Interactive Analysis of Web-Scale Datasets. http://research.google.com/pubs/archive/36632.pdf |
[Greenplum] | http://www.greenplum.com/resources/mapreduce/ |
[Chattopadhyay11] | Tenzing: A SQL Implementation On the MapReduce Framework, Bishwapesh Chattopadhyay at al, to be released at VLDB 2011 |
[Yahoo] | http://pressroom.yahoo.net/pr/ycorp/yahoo-and-benchmark-capital-introduce-hortonworks.aspx |
[Szalay08] | Szalay, A., Bell, B., Vandenberg, J., Wonders, A., Burns, R., Fay, D., Heasley, J, Hey, T., Nieto-Santisteban, M., Thakar, A., Catharine van Ingen, Wilton, R., GrayWulf: Scalable Clustered Architecture for Data Intensive Computing, Microsoft Technical Report MSR-TR-2008-187, 2008. |
[Simmhan09] | Simmhan, Y., Barga, R., Catharine van Ingen, Nieto-Santisteban, M., Dobos, L., Li, N., Shipway, M., Szalay, A., Werner, S., Heasley, J., GrayWulf: Scalable Software Architecture for Data Intensive Computing, Hawaii International Conference on System Sciences, pp. 1-10, 42nd Hawaii International Conference on System Sciences, 2009. |
[Jedicke06] | Jedicke R., Magnier E. A., Kaiser N., Chambers K. C. The next decade of Solar System discovery with Pan-STARRS. In Proceedings of the International Astronomical Union, pp 341-352, 2006 |
[Gray07] | Gray J., Nieto-Santisteban M., Szalay A., The Zones Algorithm for Finding Points-Near-a-Point or Cross-Matching Spatial Datasets, Microsoft Technical Report MSR TR 2006 52, 2007 |
[dbms209] | (1, 2) eBay’s two enormous data warehouses. http://www.dbms2.com/2009/04/30/ebays-two-enormous-data-warehouses/ |
[dbms210] | (1, 2) eBay followup — Greenplum out, Teradata > 10 petabytes, Hadoop has some value, and more http://www.dbms2.com/2010/10/06/ebay-followup-greenplum-out-teradata-10-petabytes-hadoop-has-some-value-and-more/ |
[eweek04] | At Wal-Mart, Worlds Largest Retail Data Warehouse Gets Even Larger http://www.eweek.com/c/a/Enterprise-Applications/At-WalMart-Worlds-Largest-Retail-Data-Warehouse-Gets-Even-Larger/ |
[scidb] | History of SciDB, http://www.scidb.org/about/history.php |
[Becla05] | Lessons Learned from Managing a Petabyte, Jacek Becla, Daniel Wang, CIDR Conference, Asilomar, CA, USA, January 2005 |
[Intersystems13] | European Space Agency Chooses InterSystems CACHE Database, http://www.intersystems.com/casestudies/cache/esa.html |
[ODBMS11] | Object in Space, http://www.odbms.org/blog/2011/02/objects-in-space/ |
[Stonebaker05] | C-store: a column-oriented DBMS , Stonebraker, M., Abadi, D. J., Batkin, A., Chen, X., Cherniack, M., Ferreira, M., Lau, E., Lin, A., Madden, S., O’Neil, E., O’Neil, P., Rasin, A., Tran, N., and Zdonik, S. In Proceedings of the 31st international Conference on Very Large Data Bases (Trondheim, Norway, August 30 - September 02, 2005). Very Large Data Bases. VLDB Endowment, 553-564., 2005 |
[Pavlo09] | A Comparison of Approaches to Large-Scale Data Analysis, A. Pavlo, E. Paulson, A. Rasin, D. J. Abadi, D. J. Dewitt, S. Madden, and M. Stonebraker. In SIGMOD ‘09: Proceedings of the 2009 ACM SIGMOD International Conference, 2009 |
[Docushare-11701] | Solid state disks tests, Docushare Document-11701 |
[Docushare-7025] | DM Risk Register, Docushare Document-7025, latest version at https://www.lsstcorp.org/sweeneyroot/riskmanagement/risks_01.php |
[perftests] | Query Set for Performance Tests, http://dev.lsstcorp.org/trac/wiki/dbQueriesForPerfTest |
[LDM-144] | LSST Data Management Infrastructure Costing, Docushare LDM-144 (spreadsheet) and LDM-143 (explanation). |
[Kunszt] | The Hierarchical Triangular Mesh, Peter Kunszt, Alexander Szalay, Aniruddha Thakar |
[databasecolumn] | Mapreduce: a major step back, http://www.databasecolumn.com/2008/01/mapreduce-a-major-step-back.html, http://www.databasecolumn.com/2008/01/mapreduce-continued.html. |
[hadoop-users] | Hadoop users: http://wiki.apache.org/hadoop/PoweredBy |
[nosqlpedia] | http://nosqlpedia.com/wiki/Facebook_Messaging_-_HBase_Comes_of_Age |
[Dryad07] | Distributed Data-Paralle Programs from Sequential Building Blocks, http://research.microsoft.com/en-us/projects/dryad/eurosys07.pdf |
[Calpont08] | http://www.mysqlconf.com/mysql2009/public/schedule/detail/8997 |
[Calpont09] | Calpont Launches Open Source Analytics Database Offering, http://www.calpont.com/pr/95-calpont-launches-open-source-analytics-database-offering |
[HighScalability08] | Skype Plans for PostgreSQL to scale to 1 billion users, http://highscalability.com/skype-plans-postgresql-scale-1-billion-users |
[PostGIS] | http://postgis.refractions.net/ |
[Vertica] | Using Vertica as a Structured Data Repositories for Apache Hadoop, http://www.vertica.com/MapReduce |
[slac-test] | 100-node scalability test run at SLAC, http://dev.lsstcorp.org/trac/wiki/dbQservTesting |
[Cache] | Using Caché’s Globals, http://docs.intersystems.com/documentation/cache/20082/pdfs/GGBL.pdf |
[Becla07] | Organizing the Extremely Large LSST Database for Real-Time Astronomical Processing, J. Becla, K-T Lim, S. Monkewitz, M. Nieto-Santisteban, A. Thakar, Astronomical Data Analysis Software & Systems XVII, London, UK, September 2007 |
[TokuDB] | TokuDB: Scalable High Performance for MySQL and MariaDB Databases, http://www.tokutek.com/wp-content/uploads/2013/04/Tokutek-White-Paper.pdf |
[CitusDB] | Overview of CitusDB, http://www.citusdata.com/docs |
[Ulin13] | Driving MySQL Innovation, Tom Ulin, Oracle, http://www.percona.com/live/mysql-conference-2013/sessions/keynote-driving-mysql-innovation |
[Actian] | Ingres Becomes Actian to Take Action on Big Data, http://www.actian.com/ingres-becomes-actian |
[Zdnet] | (1, 2) Microsoft drops `Dryad`_; puts its big-data bets on Hadoop*, http://www.zdnet.com/blog/microsoft/microsoft-drops-dryad-puts-its-big-data-bets-on-hadoop/11226 |
[Nobari13] | TOUCH: In-Memory Spatial Join by Hierarchical Data-Oriented Partitioning, Sadegh Nobari, Farhan Tauheed, Thomas Heinis, Panagiotis Karras, Stéphane Bressan, Anastasia Ailamaki, SIGMOD, 2013 |
12 Appendix A – Map/Reduce Solutions¶
12.1 Hadoop¶
Hadoop is a Lucene sub-project hosted by Apache. It is open source. It tries to re-create the Google MR technology [Dean04] to provide a framework in which parallel searches/projections/transformations (the Map phase) and aggregations/groupings/sorts/joins (the Reduce phase) using key-value pairs can be reliably performed over extremely large amounts of data. The framework is written in Java though the actual tasks executing the map and reduce phases can be written in any language as these are scheduled external jobs. The framework is currently supported for GNU/Linux platforms though there is on-going work for Windows support. It requires that ssh be uniformly available in order to provide daemon control.
Hadoop consists of over 550 Java classes implementing multiple components used in the framework:
- The Hadoop Distributed File System (HDFS), a custom POSIX-like file system that is geared for a write-once-read-many access model. HDFS is used to distribute blocks of a file, optionally replicated, across multiple nodes. HDFS is implemented with a single Namenode that maintains all of the meta-data (i.e., file paths, block maps, etc.) managed by one or more Datanodes (i.e., a data server running on each compute node). The Namenode is responsible for all meta-data operations (e.g., renames and deletes) as well as file allocations. It uses a rather complicated distribution algorithm to maximize the probability that all of the data is available should hardware failures occur. In general, HDFS always tries to satisfy read requests with data blocks that are closest to the reader. To that extent, HDFS also provides mechanisms, used by the framework, to co-locate jobs and data. The HDFS file space is layered on top of any existing native file system.
- A single JobTracker, essentially a job scheduler responsible for submitting and tracking map/reduce jobs across all of the nodes.
- A TaskTracker co-located with each HDFS DataNode daemon which is responsible for actually running a job on a node and reporting its status.
- DistributedCache to distribute program images as well as other required read-only files to all nodes that will run a map/reduce program.
- A client API consisting of JobConf, JobClient, Partitioner, OutputCollector, Reporter, InputFormat, OutputFormat among others that is used to submit and run map/reduce jobs and retrieve the output.
Hadoop is optimized for applications that perform a streaming search over large amounts of data. By splitting the search across multiple nodes, co-locating each with the relevant data, wherever possible, and executing all the sub-tasks in parallel, results can be obtained (relatively) quickly. However, such co-ordination comes with a price. Job setup is a rather lengthy process and the authors recommend that the map phase take at least a minute to execute to prevent job-setup from dominating the process. Since all of the output is scattered across many nodes, the map phase must also be careful to not produce so much output as to overwhelm the network during the reduce phase, though the framework provides controls for load balancing this operation and has a library of generally useful mappers and reducers to simplify the task. Even so, running ad hoc map/reduce jobs can be problematic. The latest workaround used by many Hadoop users involves running Hadoop services continuously (and jobs are attached to these services very fast). By default, joining tables in MR involves transferring data for all the joined tables into the reducer, and performing the join in the reduce stage, which can easily overwhelm the network. To avoid this, data must be partitioned, and data chunked joined together must be placed together (on the same node), in order to allow performing the join in the map stage.
Today’s implementation of Hadoop requires full data scan even for simplest queries. To avoid this, indices are needed. Implementing indices has been planned by the community for several years, and according to the latest estimates they will be implemented in one or two years. In the meantime, those who need indices must implement and maintain them themselves, the index files can be stored e.g. as files in the Hadoop File System (HDFS).
One of the “features” of MR systems is lack of official catalog (schema); instead, knowledge about schema in part of the code. While this dramatically improves flexibility and speeds up prototyping, it makes it harder to manage such data store in the long term, in particular if multi-decade projects with large number of developers are involved.
Lack of some features that are at the core of every database system should not be a surprise – MR systems are simply built with different needs in mind, and even the Hadoop website officially states that *Hadoop is not a substitute for a database*. Nethertheless, many have attempted to compare Hadoop performance with databases. According to some publications and feedback from Hadoop users we talked to, Hadoop is about an order of magnitude more wasteful of hardware than a e.g. DB2 [databasecolumn].
Hadoop has a large community supporting it; e.g., over 300 people attended the first Hadoop summit (in 2008). It is used in production by many organizations, including Facebook, Yahoo!, and Amazon Facebook [hadoop-users]. It is also commercially supported by Cloudera. Hadoop Summit 2011 was attended by more than 1,600 people from more than 400 companies.
We evaluated Hadoop in 2008. The evaluation included discussions with key developers, including Eric Baldeschwieler from Yahoo!, Jeff Hammerbacher from Facebook, and later Cloudera, discussions with users present at the 1st Hadoop Summit, and a meeting with the Cloudera team in September of 2010.
12.2 Hive¶
Hive is a data warehouse infrastructure developed by Facebook on top of Hadoop; it puts structures on the data, and defines SQL-like query language. It inherits Hadoop’s deficiencies including high latency and expensive joins. Hive works on static data, it particular it can’t deal with changing data, as row-level updates are not supported. Although it does support some database features, it is a “state where databases were ~1980: there are almost no optimizations” (based on Cloudera, meeting at SLAC Sept 2010). Typical solutions involve implementing missing pieces in user code, for example once can build their own indexes and interact directly with HDFS (and skip the Hadoop layer).
12.3 HBase¶
HBase is a column-oriented structured storage modeled after Google’s Bigtable [Bigtable06], and built on top of the Hadoop HDFS. It is good at incremental updates and column key lookups, however, similarly to plain MR, it offers no mechanism to do joins – a typical solution used by most users is to denormalize data. HBase is becoming increasingly more popular at Facebook [nosqlpedia]. It is supported commercially by Cloudera, Datameer and Hadapt.
12.4 Pig Latin¶
Pig Latin is a procedural data flow language for expressing data analysis programs. It provides many useful primitives including filter, foreach ... generate, group, join, cogroup, union, sort and distinct, which greatly simplify writing Map/Reduce programs or gluing multiple Map/Reduce programs together. It is targeted at large-scale summarization of datasets that typically require full table scans, not fast lookups of small numbers of rows. We have talked to the key developer of Pig Latin – Chris Olston.
12.6 Dryad¶
Dryad [Dryad07] is a system developed by Microsoft Research for executing distributed computations. It supports a more general computation model than MR in that it can execute graphs of operations, using so called Directed Acyclic Graph (DAG). It is somewhat analogous to the MR model in that it can model MR itself, among others, more complex flows. The graphs are similar to the query plans in a relational database. The graph execution is optimized to take advantage of data locality if possible, with computation moving to the data. If non-local data is needed, it is transferred over the network.
Dryad currently works on flat files. It is similar to Hadoop in this way.
The core execution engine in Dryad has been used in production for several years but not heavily. There are several integration pieces we might want (loading data from databases instead of files, tracking replicas of data) that do not yet exist.
Beyond the execution engine, Dryad also incorporates a simple per-node task scheduler inherited from elsewhere in Microsoft. It runs prioritized jobs from a queue. Dryad places tasks on nodes based on the data available on the node and the state of the task queue on the node. A centralized scheduler might improve things, particularly when multiple simultaneous jobs are running; that is an area that is being investigated.
Dryad requires that the localization or partitioning of data be exposed to it. It uses a relatively generic interface to obtain this metadata from an underlying filesystem, enabling it to talk to either a proprietary GFS-like filesystem or local disks.
Dryad runs only on Windows .NET at present. Building the system outside of Microsoft is difficult because of dependencies on internal libraries; this situation is similar to the one with Google’s GFS and Map/Reduce. The core execution engine could conceivably be implemented within Hadoop or another system, as its logic is not supposed to be terribly complicated. The performance-critical aspect of the system is the transfer of data between nodes, a task that Windows and Unix filesystems have not been optimized for and which Dryad therefore provides.
Dryad has been released as open source to academics/researchers in July 2009. This release however does not include any distributed filesystem for use with the system. Internally, Microsoft uses the Cosmos file system, but it is not available in the academic release. Instead there are bindings for NTFS and SQL Server.
Microsoft dropped supporting Dryad back in late 2011 [Zdnet].
12.7 Dremel¶
Dremel is a scalable, interactive ad-hoc query system for analysis of read-only data, implemented as an internal project at Google [Dremel]. Information about Dremel has been made available in July 2010. Dremel’s architecture is in many ways similar to our baseline architecture (executing query in parallel on many nodes in shared nothing architecture, auto fail over, replicating hot spots). Having said that, we do not have access to the source code, even though Google is an LSST collaborator, and there is no corresponding open source alternative to date.
12.8 Tenzing¶
Tenzing is an SQL implementation on the MapReduce Framework [Chattopadhyay11] We managed to obtain access to pre-published paper from Google through our XLDB channels several months before planned publication at the VLDB 2011 conference.
Tenzing is a query engine built on top of MR for ad hoc analysis of Google data. It supports a mostly complete SQL implementation (with several extensions) combined with several key characteristics such as heterogeneity, high performance, scalability, reliability, metadata awareness, low latency support for columnar storage and structured data, and easy extensibility.
The Tenzing project underscores importance and need of database-like features in any large scale system.
12.9 “NoSQL”¶
The popular term NoSQL originally refered to systems that do not expose SQL interface to the user, and it recently evolved and refers to structured systems such as key-value stores or document stores. These systems tend to provide high availability at the cost of relaxed consistency (“eventual” consistency). Today’s key players include Cassandra and MongoDB.
While a key/value store might come handy in several places in LSST, these systems do not address many key needs of the project. Still, a scalable distributed key-value store may be appropriate to integrate as an indexing solution within a larger solution.
13 Appendix B – Database Solutions¶
In alphabetical order.
13.1 Actian¶
Actian, formerly known as Ingres [Actian] provides analytical services through Vectorwise, acquired from CWI in 2010. Primary speed ups rely on exploiting data level parallelism (rather than tuple-at-a-time processing). Main disadvantage from LSST perspective: it is a single-node system.
13.2 Caché¶
InterSystems Caché is a shared-nothing object database system, released as an embedded engine since 1972. Internally it stores data as multi-dimensional arrays, and interestingly, supports overlaps. We are in discussions with the company—we have been discussing Caché with Stephen Angevine since early 2007, and met with Steven McGlothlin in June 2011. We also discussed Caché with William O’Mullane from the ESA’s Gaia mission, an astronomical survey that selected Caché as their underlying database store in 2010 [25, 26]). InterSystems offers free licensing for all development and research, for academic and non-profit research, plus support contracts with competitive pricing. However, their system does not support compression and stores data in strings, which may not be efficient for LSST catalog data.
A large fraction of the code is already available as open source for academia and non-profit organizations under the name “Globals” [Cache].
13.3 CitusDB¶
CitusDB is a new commercial distributed database built on top on PostgreSQL. It supports joins between one large and multiple small tables [CitusDB] (star schema) – this is insufficient for LSST.
13.4 DB2¶
IBM’s DB2 “parallel edition” implements a shared-nothing architecture since mid-1990. Based on discussions with IBM representatives including Guy Lohman (e.g., at the meeting in June 2007) as well as based on various news, it appears that IBM’s main focus is on supporting unstructured data (XML), not large scale analytics. All their majors projects announced in the last few years seem to confirm them, including Viper, Viper2 and Cobra (XML) and pureScale (OLTP).
13.5 Drizzle¶
Drizzle is a fork from the MySQL Database, the fork was done shortly after the announcement of the acquisition of MySQL by Oracle (April 2008). Drizzle is lighter than MySQL: most advanced features such as partitioning, triggers and many others have been removed (the code base was trimmed from over a million lines down to some 300K, it has also been well modularized). Drizzle‘s main focus is on the cloud market. It runs on a single server, and there are no plans to implement shared-nothing architecture. To achieve shared-nothing architecture, Drizzle has hooks for an opaque sharding key to be passed through client, proxy, server, and storage layers, but this feature is still under development, and might be limited to hash-based sharding.
Default engine is InnoDB. MyISAM engine is not part of Drizzle, it is likely MariaDB engine will become a replacement for MyISAM.
Drizzle’s first GA release occurred in March 2011.
We have discussed the details of Drizzle with key Drizzle architects and developers, including Brian Aker (the chief architect), and most developers and users present at the Drizzle developers meeting in April 2008.
13.6 Greenplum¶
Greenplum is a commercial parallelization extension of PostgreSQL. It utilizes a shared-nothing, MPP architecture. A single Greenplum database image consists of an array of individual databases which can run on different commodity machines. It uses a single Master as an entry point. Failover is possible through mirroring database segments. According to some, it works well with simple queries but has issues with more complex queries. Things to watch out for: distributed transaction manager, allegedly there are some issues with it.
Up until recently, Greenplum powered one of the largest (if not the largest) database setups: eBay was using it to manage 6.5 petabytes of data on a 96-node cluster [dbms209]. We are in close contact with the key eBay developers of this system, including Oliver Ratzesberger.
We are in contact with the Greenplum CTO: Luke Lonergan.
08/28/2008: Greenplum announced supporting MapReduce [Greenplum].
Acquired by EMC in July 2010.
13.7 GridSQL¶
GridSQL is an open source project sponsored by EnterpriseDB. GridSQL is a thin layer written on top of postgres that implemented shared-nothing clustered database system targeted at data warehousing. This system initially looked very promising, so we evaluated it in more details, including installing it on our 3-node cluster and testing its capabilities. We determined that currently in GridSQL, the only way to distribute a table across multiple nodes is via hash partitioning. We can’t simply hash partition individual objects, as this would totally destroy data locality, which is essential for spatial joins. A reasonable workaround is to hash partition entire declination zones (hash partition on zoneId), this will insure all objects for a particular zone end up on the same node. Further, we can “chunk” each zone into smaller pieces by using a regular postgres range partitioning (sub-tables) on each node.
The main unsolved problems are:
- near neighbor queries. Even though it is possible to slice a large table into pieces and distribute across multiple nodes, it is not possible to optimize a near neighbor query by taking advantage of data locality – GridSQL will still need to do n2 correlations to complete the query. In practice a special layer on top of GridSQL is still needed to optimize near neighbor queries.
- shared scans.
Another issue is stability, and lack of good documentation.
Also since GridSQL is based on PostgreSQL, it inherits the postgres “cons”, such as the slow performance (comparing to MySQL) and having to reload all data every year.
The above reasons greatly reduce the attractiveness of GridSQL.
We have discussed in details the GridSQL architecture with their key developer, Mason Sharp, who confirmed the issues we identified are unlikely to be fixed/implemented any time soon.
13.7.1 Gridsql Tests¶
We installed GridSQL on a 3 node cluster at SLAC and run tests aimed to uncover potential bottlenecks, scalability issues and understand performance. For these tests we used simulated data generated by the software built for LSST by the UW team.
Note that GridSQL uses PostgreSQL underneath, so these tests included installing and testing PostgreSQL as well.
For these tests we used the USNO-B catalog. We run a set of representative queries, ranging from low volume queries (selecting a single row for a large catalog, a cone search), to high-volume queries (such as near-neighbor search).
Our timing tests showed acceptable overheads in performance compared to PostgreSQL standalone.
We examined all data partitioning options available in GridSQL. After reading documentation, interacting with GridSQL developers, we determined that currently in GridSQL, the only way to distribute a table across multiple nodes is via hash partitioning. We can’t simply hash partition individual objects, as this would totally destroy data locality, which is essential for spatial joins. A reasonable workaround we found is to hash partition entire declination zones (hash partition on zoneId), this will insure all objects for a particular zone end up on the same node. Further, we can “chunk” each zone into smaller pieces by using a regular PostgreSQL range partitioning (sub-tables) on each node.
We were unable to find a clean solution for the near neighbor queries. Even though it is possible to slice a large table into pieces and distribute across multiple nodes, it is not possible to optimize a near neighbor query by taking advantage of data locality, so in practice GridSQL will still need to do n2 correlations to complete the query. In practice a special layer on top of GridSQL is still needed to optimize near neighbor queries. So in practice, we are not gaining much (if anything) by introducing GridSQL into our architecture.
During the tests we uncovered various stability issues, and lack of good documentation.
In addition, GridSQL is based on PostgreSQL, so it inherits the PostgreSQL “cons”, such as the slow performance (comparing to MySQL) and having to reload all data every year, described separately.
13.8 InfiniDB¶
InfiniDB is an open source, columnar DBMS consisting of a MySQL front end and a columnar storage engine, build and supported by Calpont. Calpont introduced their system at the MySQL 2008 User Conference [Calpont08], and more officially announced it in late Oct 2009 [Calpont09]. It implements true MPP, shared nothing (or shared-all, depending how it is configured) DBMS. It allows data to be range-based horizontal partitioning, partitions can be distributed across many nodes (overlapping partitions are not supported though). It allows to run distributed scans, filter aggregations and hash joins, and offers both intra- and inter- server parallelism. During cross-server joins: no direct communication is needed between workers. Instead, they build 2 separate hash maps, distribute smaller one, or if too expensive to distribute they can put it on the “user” node.
A single-server version of InfiniDB software is available through free community edition. Multi-node, MPP version of InfiniDB is only available through commercial, non-free edition, and is closed source.
We are in contact with Jim Tommaney, CTO of the Calpont Corporation since April 2008. In late 2010 we run the most demanding query – the near neighbor tests using Calpont. Details of these tests are covered in 14 Appendix C: Tests with InfiniDB.
13.9 LucidDB¶
LucidDB is an open source columnar DBMS. Early startup (version 0.8 as of March 2009). They have no plans to do shared-nothing (at least there is no mention of it, and on their main page they mention “great performance using only a single off-the-shelf Linux or Windows server.”). Written mostly in java.
13.10 MySQL¶
MySQL utilizes a shared-memory architecture. It is attractive primarily because it is a well supported, open source database with a large company (now Oracle) behind it and a big community supporting it. (Note, however, that much of that community uses it for OLTP purposes that differ from LSST’s needs.) MySQL’s optimizer used be below-average, however it is slowly catching up, especially the MariaDB version.
We have run many, many performance tests with MySQL. These are documented in trac in various places, many of them on these four pages:
- http://dev.lsstcorp.org/trac/wiki/dbSpatialJoinPerf
- http://dev.lsstcorp.org/trac/wiki/dbBuildSubPart
- http://dev.lsstcorp.org/trac/wiki/dbSubPartOverhead
- http://dev.lsstcorp.org/trac/wiki/DbStoringRefCat
We are well plugged into the MySQL community, we attended all MySQL User Conferences in the past 5 years, and talked to many MySQL developers, including director of architecture (Brian Aker), the founders (Monty Widenius, David Axmark), and theMySQL optimizer gurus.
There are several notable open-source forks of MySQL:
- The main one, supported by Oracle. After initial period when Oracle was pushing most new functionality into commercial version of MySQL only [Error: Reference source not found], the company now appears fully committed to support MySQL, arguing MySQL is a critical component of web companies and it is one of the components of the full stack of products they offer. Oracle has doubled the number of MySQL engineers and tripled the number of MySQL QA staff over the past year [Ulin13], and the community seems to believe Oracle is truly committed now to support MySQL. The main “problem” from LSST perspective is that Oracle is putting all the effort into InnoDB engine only (the engine used by web companies including Facebook and Google), while the MyISAM engine, the engine of choice for LSST, selected because of vastly different access pattern characteristics, remains neglected and Oracle currently has no plans to change that.
- MontyProgram and SkySQL used to support two separate forks of MySQL, in April 2013 they joint efforts; the two founders of MySQL stand behind these two companies. MontyProgram is supporting a viable alternative to InnoDB, called MariaDB, and puts lots of efforts into improving and optimizing MyISAM. As an example, the mutli-core performance issues present in all MySQL engines in the past were fixed by Oracle for InnoDB, and in Aria, the MontyProgram’s version of MyISAM by MontyProgram.
- Percona, which focuses on multi-core scaling
- Drizzle, which is aslimmed-down version, rewriten from scratch and no longer compatible with MySQL. Based on discussions with the users, the Drizzle effort has not picked up, and is slowly dying.
Spatial indexes / GIS. As of version 5.6.1, MySQL has rewritten spatial support, added support for spatial indexes (for MyISAM only) and functions using the OpenGIS geometry model. We have not yet tested this portion of MySQL, and have preferred using geometry functionality from SciSQL, a MySQL plug-in written inhouse..
13.10.1 MySQL – Columnar Engines¶
13.10.1.1 KickFire¶
KickFire is a hardware appliance built for MySQL. It runs a proprietary database kernel (a columnar data store with aggressive compression) with operations performed on a custom dataflow SQL chip. An individual box can handle up to a few terabytes of data. There are several factors that reduce the attractiveness of this approach:
- it is a proprietary “black box”, which makes it hard to debug, plus it locks us into a particular technology
- it is an appliance, and custom hardware tends to get obsolete fairly rapidly
- it can handle low numbers of terabytes; another level is still needed (on top?) to handle petabytes
- there is no apparent way to extend it (not open source, all-in-one “black box”)
We have been in contact with the key people since April of 2007, when the team gave us a demo of their appliance under an NDA.
13.10.1.2 InfoBright¶
Infobright is a proprietary columnar engine for MySQL. Infobright Community Edition is open-source, but lacks many features, like parallelism and DML (INSERT, UPDATE, DELETE, etc). Infobright Enterprise Edition is closed-source, but supports concurrent queries and DML. Infobright’s solution emphasizes single-node performance without discussing distributed operation (except for data ingestion in the enterprise edition).
13.10.2 TokuDB¶
Tokutek built a specialized engine, called TokuDB. The engine relies on new indexing method, called Fractal Tree indexes [TokuDB], this new type of an index primarily increases speed of inserts and data replication. While its benefits are not obvious for our data access center, rapid inserts might be useful for Level 1 data sets (Alert Production). We have been in touch with the Tokutek team for several years, the key designers of the Fractal Tree index gave a detailed tutorial at the XLDB-2012 conference we organized.
The engine was made open source in Apr 2013.
13.11 Netezza¶
Netezza Performance Server (NPS) is a proprietary, network attached, streaming data warehousing appliance that can run in a shared-nothing environment. It is built on PostgreSQL.
The architecture of NPS consists of two tiers: a SMP host and hundreds of massively parallel blades, called Snippet Processing Units (SPU). Each SPU consists of a CPU, memory, disk drive and an FPGA chip that filters records as they stream off the disk. See http://www.netezza.com/products/index.cfm for more information.
According to some rumours, see eg http://www.dbms2.com/2009/09/03/teradata-and-netezza-are-doing-mapreduce-too/, Netezza is planning to support map/reduce.
Pros:
- It is a good, scalable solution
- It has good price/performance ratio.
Cons:
- it is an appliance, and custom hardware tends to get obsolete fairly rapidly
- high hardware cost
- proprietary
Purchased by IBM.
13.12 Oracle¶
Oracle provides scalability through Oracle Real Application Clusters (RAC). It implements a shared-storage architecture.
Cons: proprietary, expensive. It ties users into specialized (expensive) hardware (Oracle Clusterware) in the form of storage area networks to provide sufficient disk bandwidth to the cluster nodes; the cluster nodes themselves are often expensive shared-memory machines as well. It is very expensive to scale to very large data sets, partly due to the licensing model. Also, the software is very monolithic, it is therefore changing very, very slowly.
We have been approached several times by Oracle representatives, however given we believe Oracle is not a good fit for LSST, we decided not to invest our time in detailed investigation.
13.13 ParAccel¶
ParAccel Analytic Database is a proprietary RDBMS with a shared-nothing MPP architecture using columnar data storage. They are big on extensibility and are planning to support user-defined types, table functions, user-defined indexes, user-defined operators, user-defined compression algorithms, parallel stored procedures and more.
When we talked to ParAccel representatives (Rick Glick, in May 2008), the company was still in startup mode.
13.14 PostgreSQL¶
PostgreSQL is an open source RDBMS running in a shared-memory architecture.
PostgreSQL permits horizontal partitioning of tables. Some large-scale PostgreSQL-based applications use that feature to scale. It works well if cross-partition communication is not required.
The largest PostgreSQL setup we are aware of is AOL’s 300 TB installation (as of late 2007). Skype is planning to use PostgreSQL to scale up to billions of users, by introducing a layer of proxy servers which will hash SQL requests to an appropriate PostgreSQL database server, but this is an OLTP usage that supports immense volumes of small queries [HighScalability08].
PostgreSQL also offers good GIS support [PostGIS]. We are collaborating with the main authors of this extension.
One of the main weaknesses of PostgreSQL is a less-developed support system. The companies that provide support contracts are less well-established than Sun/MySQL. Unlike MySQL, but more like Hadoop, the community is self-organized with no single central organization representing the whole community, defining the roadmap and providing long term support. Instead, mailing lists and multiple contributors (people and organizations) manage the software development.
PostgreSQL is more amenable to modification than MySQL, which may be one reason why it has been used as the basis for many other products, including several mentioned below.
Based on the tests we run, PostgreSQL performance is 3.7x worse than MySQL. We realize the difference is partly due to very different characteristics of the engines used in these tests (fully ACID-compliant PostgreSQL vs non-transactional MyISAM), however the non-transactional solution is perfectly fine, and actually preferred for our immutable data sets.
We are in touch with few most active PostgreSQL developers, including the authors of Q3C mentioned above, and Josh Berkus.
13.14.1 Tests¶
We have run various performance tests with PostgreSQL to compare its performance with MySQL. These tests are described in details in the “Baseline architecture related” section below. Based on these tests we determined PostgreSQL is significantly (3.7x) slower than MySQL for most common LSST queries.
We have also tried various partitioning schemes available in PostgreSQL. In that respect, we determined PostgreSQL is much more advanced than MySQL.
Also, during these tests we uncovered that PostgreSQL requires dump/reload of all tables for each major data release (once per year), see http://www.postgresql.org/support/versioning. The PostgreSQL community believes this is unlikely to change in the near future. This is probably the main show-stopper preventing us from adapting PostgreSQL.
13.15 SciDB¶
SciDB is a new open source system inspired by the needs of LSST15 and built for scientific analytics. SciDB implements a shared nothing, columnar MPP array database, user defined functions, overlapping partitions, and many other features important for LSST. SciDB Release 11.06, the first production release, was published on June 15, 2011. We are in the process of testing this release.
13.16 SQLServer¶
Microsoft’s SQLServer’s architecture is shared-memory. The largest SQLServer based setup we are aware of is the SDSS database (6 TB), and the Pan-STARRS database.
In 2008 Microsoft bought DATAllegro and began an effort, codenamed “Project Madison,” to integrate it into SQLServer. Madison relies on shared nothing computing. Control servers are connected to compute nodes via dual Infiniband links, and compute servers are connected to a large SAN via dual Fiber Channel links. Fault tolerance relies on (expensive) hardware redundancy. For example, servers tend to have dual power supplies. However, servers are unable to recover from storage node failures, thought a different replica may be used. The only way to distribute data across nodes is by hashing; the system relies on replicating dimension tables. [the above is based on the talk we attended: http://wiki.esi.ac.uk/w/files/5/5c/Dyke-Details_of_Project_Madison-1.pdf]
Cons: It is proprietary, relies on expensive hardware (appliance), and it ties users to the Microsoft OS.
About DATAllegro. DATAllegro was a company specializing in data warehousing server appliances that are pre-configured with a version of the Ingres database optimized to handle relatively large data sets (allegedly up to hundreds of terabytes). The optimizations reduce search space during joins by forcing hash joins. The appliances rely on high speed interconnect(Infiniband).
13.17 Sybase IQ¶
Sybase IQ is a commercial columnar database product by Sybase Corp. Sybase IQ utilizes a “shared-everything” approach that designed to provide graceful load-balancing. We heard opinions that most of the good talent has left the company; thus it is unlikely it will be a major database player.
Cons: proprietary.
13.18 Teradata¶
Teradata implements a shared-nothing architecture. The two largest customers include eBay and WalMart. Ebay is managing multi petabyte Teradata-based database.
The main disadvantage of Teradata is very high cost.
We are in close contact with Steve Brobst, acting as Teradata CTO, and key database developers at eBay.
13.19 Vertica¶
The Vertica Analytics Platform is a commercial product based on the open source C-store column-oriented database, and now owned by HP. It utilizes a shared-nothing architecture. Its implementation is quite innovative, but involves signficant complexity underneath.
It is built for star/snowflake schemas. It currently can not join multiple fact tables; e.g. self-joins are not supported though this will be fixed in future releases. Star joins in the MPP environment are made possible by replicating dimension tables and partitioning the fact table.
In 2009, a Vertica Hadoop connector was implemented. This allows Hadoop developers to push down map operators to Vertica database, stream Reduce operations into Vertica [Vertica], and move data between the two environments.
Cons:
- lack of support of self-joins
- proprietary..
13.20 Others¶
In addition to map/reduce and RDBMS systems, we also evaluation several other software packages which could be used as part of our custom software written on top of MySQL. The components needed include SQL parser, cluster management and task management.
13.20.1 Cluster and task and management¶
Two primary candidates to use as cluster and task management we identified so far are Gearman and XRootD. Cluster management involves keeping track of available nodes, allowing nodes to be added/removed dynamically. Task management involves executing tasks on the nodes.
Detailed requirements what we need are captured at: http://dev.lsstcorp.org/trac/wiki/dbDistributedFrameworkRequirements
13.20.1.1 Gearman¶
Gearman is a distributed job execution system, available as open source. It provides task management functions, e.g., cluster management is left out to be handled in application code.
During a meeting setup in June 2009 with Eric Day, the key developer working on integration of Drizzle with Gearman, who also wrote the C++ version of Gearman, we discussed details of Gearman architecture and its applicability for LSST.
Gearman manages workers as resources that provide RPC execution capabilities. It is designed to provide scalable access to many resources that can provide similar functionality (e.g., compress an image, retrieve a file, perform some expensive computation). While we could imagine a scheme to use Gearman’s dispatch system, its design did not match LSST’s needs well. One problem was its store-and-forward approach to arguments and results, which would mean that the query service would need to implement its own side transmission channel or potentially flood the Gearman coordinator with bulky results.
14 Appendix C: Tests with InfiniDB¶
In late 2010 we collaborated with the Calpont team on testing their InfiniDB product. Testing involved executing the most complex queries such as near neigbor on 1 billion row USNOB catalog. The tests were run by Jim Tommaney, the final results are pasted below.
Thank you for the chance to evaluate InfiniDB against the stellar data set and the near neighbor problem. Towards that end I installed our 2.0 version on a Dell 610 server with 16GB memory, 8 physical cores (16 Hyper-Threaded Intel virtual cores), and a 4 disk raid 0 data mount point with 7200 RPM disk drives.
As you know, the N-squared search space becomes problematic at scale, so part of the solution involved a specialized query and the addition of 4 additional columns as shown below. These new columns defined two overlapping grids on top of the search space such that any given object existed in 2 grids. Note that these are abstract grids represented by additional columns and the table is a single table with our standard vertical + horizontal partitioning that happens with the basic create table statement. So, this virtual ‘gridding’ doesn’t change other characteristics of the table or prevent other such extensions.
Column additions:
alter table object add column ra_r2 decimal(5,2);
alter table object add column decl_r2 decimal(4,2);
alter table object add column ra_d2 decimal(5,2);
alter table object add column decl_d2 decimal(4,2);
Example update statements:
update object set ra_r2 = round(ra,2) where ra < 10;
update object set decl_r2 = round(decl,2) where ra < 10;
update object set ra_d2 = truncate(ra,2) where ra < 10;
update object set decl_d2 = truncate(decl,2) where ra < 10;
The query itself consist of 4 parts, the sum of which is the count of near neighbors.
Search within Grid D defined by
ra_d2
,decl_d2
.Search within Grid R defined by
ra_r2
,decl_r2
, adding predicates to only include matches that span two D Grids.and (( o1.ra_d2 <> o2.ra_d2 ) or (o1.decl_d2 <> o2.decl_d2))
There is the additional condition where a given pair of neighbors span both Grid D and Grid R. For this subset of the data, the neighboring objects share Grid R coordinates for RA, and Grid D coordinates for Decl.
FROM object o1 join object o2 on (o1.ra_r2 = o2.ra_r2 and o1.decl_d2 = o2.decl_d2)
The last case covers the same basic condition as 3, but includes a join that covers neighboring objects that share Grid D coordinates for RA, and Grid R coordinates for Decl.
FROM object o1 join object o2 on (o1.ra_d2 = o2.ra_d2 and o1.decl_r2 = o2.decl_r2)
Anyway, the results appear very promising, and indicate that it may
satisfy arbitrarily large search spaces. I executed the query
against the full range of declination, and searched with the range
of RA between 0 and ra_limit
. I then scaled ra_limit
between 0.01
through 20 and charted the results below, trending search space vs.
rows processed per second. The baseline numbers you provided appear
to avg. about 1000 rows/second, and capped out at about 80k search
space. With InfiniDB, the search rate is relatively flat after a
ramp-up of a couple seconds, running at about ~800 K rows processed
per second through a search space of about 32 M x 32 M objects. At
32M objects x 32M objects the query consumed about 6GB for the hash
structures, however extending the query logic above would allow for
running something like 33 of these queries serially to search
through a 1B x 1B space. Running the 4 sections serially would
reduce the memory requirements if desired.
There are a number of variations on the near neighbor problem that
provide a filter on one of the object tables, i.e. search for white
dwarf that I would characterize as M x N problems where M << N. To
profile those queries I selected an arbitrary filter ( o1.bMag > 24.9
)
that restricted 1 of the 2 sides of the join to at most ~330 K
objects. I then executed 12 queries with ra between 0 and ra_limit
,
varying ra_limit
from 30 to 360. Each query was executed 3 times
sequentially following a flush of the data buffer cache, and the average
of the three values charted.
With a bounded M, the processing rate went up significantly, approaching 4.5 M rows per second when the second and third executions of a query were satisfied from cache, and running at nearly 3 M rows per second for larger queries that did not fit in the data buffer cache ( which was configured at 8 GB ). These queries only used about 6% of memory for temporary space and could be run against an arbitrarily large N as desired.
There are more details regarding the load rate, options on other grid sizes, limitation of this style grid analysis for larger definitions of ‘near’, etc. that can be shared and reviewed as desired, and I am more than happy to profile additional queries as desired. For example, I can take a look at getting an exact time for finding all of the near-neighbors within a 1B x 1B search space if that is interesting (should be something like 23-25 minutes), it is just a matter of tweaking the query restrictions to allow proper handling of objects on each side of these larger query boundary.
There are definitely some significant differences between InfiniDB and MySQL in terms of best practices for a number of items. For example, our fastest load capability is via cpimport rather than load data infile. The near neighbors problem appears to be one example of many where we handle large data analysis significantly better than MySQL, although there are plenty of examples where MySQL shines relative to InfinDB (individual row insertion, individual record access via index, etc). Any external application that relies on individual row lookups with an expected latency in the microseconds will run significantly slower with InfiniDB.
select sum(cnt) from (
SELECT count(*) cnt
FROM object o1 join object o2 using(ra_d2, decl_d2)
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND ABS(o1.decl -
o2.decl) < 0.00083 and o1.objectid <> o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit and o1.bMag > 24.9
and o2.ra >= 0 and o2.ra < @ra_limit
union all
SELECT count(*) cnt
FROM object o1 join object o2 using(ra_r2, decl_r2)
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl
and ABS(o1.decl - o2.decl) < 0.00083
and o1.objectid <> o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit and o1.bMag > 24.9
and o2.ra >= 0 and o2.ra < @ra_limit
and (( o1.ra_d2 <> o2.ra_d2 ) or (o1.decl_d2 <> o2.decl_d2))
union all
select count(*) cnt
FROM object o1 join object o2 on (o1.ra_r2 = o2.ra_r2 and
o1.decl_d2 = o2.decl_d2 )
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND
ABS(o1.decl - o2.decl) < 0.00083
and o1.ra_d2 <> o2.ra_d2
and o1.decl_r2 <> o2.decl_r2
and abs(o1.ra - o1.ra_r2) * o1.cosRadDecl < 0.00083
and abs(o2.ra - o2.ra_r2) * o2.cosRadDecl < 0.00083
and abs(o1.decl - (o1.decl_d2 + 0.005)) < 0.00083
and abs(o2.decl - (o2.decl_d2 + 0.005)) < 0.00083
and o1.objectid <> o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit and o1.bMag > 24.9
and o2.ra >= 0 and o2.ra < @ra_limit
union all
select count(*) cnt
FROM object o1 join object o2 on (o1.ra_d2 = o2.ra_d2 and
o1.decl_r2 = o2.decl_r2 )
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND
ABS(o1.decl - o2.decl) < 0.00083
and o1.ra_r2 <> o2.ra_r2
and o1.decl_d2 <> o2.decl_d2
and abs(o1.ra - (o1.ra_d2 + 0.005)) * o1.cosRadDecl < 0.00083
and abs(o2.ra - (o2.ra_d2 + 0.005)) * o2.cosRadDecl < 0.00083
and abs(o1.decl - o1.decl_r2 ) < 0.00083
and abs(o2.decl - o2.decl_r2 ) < 0.00083
and o1.objectid <> o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit and o1.bMag > 24.9
and o2.ra >= 0 and o2.ra < @ra_limit
) a;
mysql> set @ra_limit:= 0.01;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 16652 | 1 | 8643 |
+----------+------------------------+--------------------------+
1 row in set (0.07 sec)
+----------+
| sum(cnt) |
+----------+
| 2834 |
+----------+
1 row in set (0.60 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-0; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 0.1;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 167632 | 10 | 17048 |
+----------+------------------------+--------------------------+
1 row in set (0.08 sec)
+----------+
| sum(cnt) |
+----------+
| 31757 |
+----------+
1 row in set (0.95 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-6; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 0.5;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 833849 | 50 | 17812 |
+----------+------------------------+--------------------------+
1 row in set (0.16 sec)
+----------+
| sum(cnt) |
+----------+
| 155065 |
+----------+
1 row in set (2.07 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-6; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 1;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 1638966 | 100 | 17891 |
+----------+------------------------+--------------------------+
1 row in set (0.21 sec)
+----------+
| sum(cnt) |
+----------+
| 290972 |
+----------+
1 row in set (2.22 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-7; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 2;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 3153437 | 200 | 17947 |
+----------+------------------------+--------------------------+
1 row in set (0.25 sec)
+----------+
| sum(cnt) |
+----------+
| 516670 |
+----------+
1 row in set (3.79 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-8; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 3;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 4675828 | 300 | 17961 |
+----------+------------------------+--------------------------+
1 row in set (0.28 sec)
+----------+
| sum(cnt) |
+----------+
| 734540 |
+----------+
1 row in set (6.51 sec)
+----------------------------------------------------------------------------------------------------+
| idb() |
+----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-9; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+----------------------------------------------------------------------------------------------------+
1 row in set (0.01 sec)
mysql> set @ra_limit:= 4;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 6356011 | 400 | 17967 |
+----------+------------------------+--------------------------+
1 row in set (0.40 sec)
+----------+
| sum(cnt) |
+----------+
| 1037556 |
+----------+
1 row in set (7.61 sec)
+-----------------------------------------------------------------------------------------------------+
| idb() |
+-----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-13; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+-----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 5;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 7989071 | 500 | 17975 |
+----------+------------------------+--------------------------+
1 row in set (0.46 sec)
+----------+
| sum(cnt) |
+----------+
| 1326087 |
+----------+
1 row in set (9.38 sec)
+-----------------------------------------------------------------------------------------------------+
| idb() |
+-----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-14; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+-----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
mysql> set @ra_limit:= 10;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 15896387 | 1000 | 17986 |
+----------+------------------------+--------------------------+
1 row in set (0.92 sec)
+----------+
| sum(cnt) |
+----------+
| 2609059 |
+----------+
1 row in set (19.71 sec)
+-----------------------------------------------------------------------------------------------------+
| idb() |
+-----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-18; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+-----------------------------------------------------------------------------------------------------+
1 row in set (0.01 sec)
mysql> set @ra_limit:= 15;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 24045205 | 1500 | 17993 |
+----------+------------------------+--------------------------+
1 row in set (1.34 sec)
+----------+
| sum(cnt) |
+----------+
| 3949531 |
+----------+
1 row in set (32.46 sec)
+-----------------------------------------------------------------------------------------------------+
| idb() |
+-----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-28; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+-----------------------------------------------------------------------------------------------------+
1 row in set (0.01 sec)
mysql> set @ra_limit:= 20;
Query OK, 0 rows affected (0.00 sec)
mysql> . near_neighbors.sql
+----------+------------------------+--------------------------+
| count(*) | count(distinct(ra_d2)) | count(distinct(decl_d2)) |
+----------+------------------------+--------------------------+
| 31841849 | 2000 | 17996 |
+----------+------------------------+--------------------------+
1 row in set (1.74 sec)
+----------+
| sum(cnt) |
+----------+
| 5247760 |
+----------+
1 row in set (40.48 sec)
+-----------------------------------------------------------------------------------------------------+
| idb() |
+-----------------------------------------------------------------------------------------------------+
| Query Stats: MaxMemPct-37; ApproxPhyI/O-0; CacheI/O-0; BlocksTouched-0; PartitionBlocksEliminated-0 |
+-----------------------------------------------------------------------------------------------------+
1 row in set (0.00 sec)
lsst_near_neighbors.sql
:
select sum(cnt) from (
SELECT count(*) cnt
FROM object o1 join object o2 using(ra_d2, decl_d2)
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND ABS(o1.decl -
o2.decl) < 0.00083 and o1.objectid < o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit
and o2.ra >= 0 and o2.ra < @ra_limit
union all
SELECT count(*) cnt
FROM object o1 join object o2 using(ra_r2, decl_r2)
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl
and ABS(o1.decl - o2.decl) < 0.00083
and o1.objectid < o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit
and o2.ra >= 0 and o2.ra < @ra_limit
and (( o1.ra_d2 <> o2.ra_d2 ) or (o1.decl_d2 <> o2.decl_d2))
union all
select count(*) cnt
FROM object o1 join object o2 on (o1.ra_r2 = o2.ra_r2 and
o1.decl_d2 = o2.decl_d2 )
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND
ABS(o1.decl - o2.decl) < 0.00083
and o1.ra_d2 <> o2.ra_d2
and o1.decl_r2 <> o2.decl_r2
and abs(o1.ra - o1.ra_r2) * o1.cosRadDecl < 0.00083
and abs(o2.ra - o2.ra_r2) * o2.cosRadDecl < 0.00083
and abs(o1.decl - (o1.decl_d2 + 0.005)) < 0.00083
and abs(o2.decl - (o2.decl_d2 + 0.005)) < 0.00083
and o1.objectid < o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit
and o2.ra >= 0 and o2.ra < @ra_limit
union all
select count(*) cnt
FROM object o1 join object o2 on (o1.ra_d2 = o2.ra_d2 and
o1.decl_r2 = o2.decl_r2 )
WHERE ABS(o1.ra - o2.ra) < 0.00083 / o2.cosRadDecl AND
ABS(o1.decl - o2.decl) < 0.00083
and o1.ra_r2 <> o2.ra_r2
and o1.decl_d2 <> o2.decl_d2
and abs(o1.ra - (o1.ra_d2 + 0.005)) * o1.cosRadDecl < 0.00083
and abs(o2.ra - (o2.ra_d2 + 0.005)) * o2.cosRadDecl < 0.00083
and abs(o1.decl - o1.decl_r2 ) < 0.00083
and abs(o2.decl - o2.decl_r2 ) < 0.00083
and o1.objectid < o2.objectid
and o1.ra >= 0 and o1.ra < @ra_limit
and o2.ra >= 0 and o2.ra < @ra_limit
) a;
16 Appendix E: People/Communities We Talked To¶
Solution providers of considered products:
- Map/Reduce – key developers from Google
- Hadoop – key developers from Yahoo!, founders and key developers behind Cloudera and Hortonworks, companyies supporting enterprise edition of Hadoop
- Hive – key developers from Facebook.
- Dryad – key developers from Microsoft (Dryad is Microsofts’s version of map/reduce), including Michael Isard
- Gearman – key developers (gearman is a system which allows to run MySQL in a distributed fashion)
- representatives from all major database vendors, including Teradata, Oracle, IBM/DB2, Greenplum, Postgres, MySQL, MonetDB, SciDB
- representatives from promising startups including HadoopDB, ParAcell, EnterpriseDB, Calpont, Kickfire
- Intersystem’s Cache—Stephen Angevine, Steven McGlothlin
- Metamarkets
User communities:
- Web companies, including Google, Yahoo, eBay, AOL
- Social networks companies, including Facebook, LinkedIn, Twitter, Zynga and Quora
- Retail companies, including Amazon, eBay and Sears,
- Drug discovery (Novartis)
- Oil & gas companies (Chevron, Exxon)
- telecom (Nokia, Comcast, ATT)
- science users from HEP (LHC), astronomy (SDSS, Gaia, 2MASS, DES, Pan-STARRS, LOFAR), geoscience, biology
Leading database researchers
- M Stonebraker
- D DeWitt
- S Zdonik
- D Maier
- M Kersten
17 Change Record¶
Version | Date | Description | Owner |
---|---|---|---|
1.0 | 6/15/2009 | Initial version. | Jacek Becla |
2.0 | 7/12/2011 | Most sections rewritten, added scalability test section | Jacek Becla |
2.1 | 8/12/2011 | Refreshed future-plans and schedule of testing sections, added section about fault tolerance. | Jacek Becla, Daniel Wang |
3.0 | 8/2/2013 | Synchronized with latest changes to the requirements (LSE-163). Rewrote most of the “Implementation” chapter. Documented new tests, refreshed all other chapters. | Jackek Becla, Daniel Wang, Serge Mokewitz, Kian-Tat Lim, Douglas Smith, Bill Chickering |
3.1 | 10/10/2013 | Refreshed numbers based on latest LDM-141. Updated shared scans (implementation) and 300-node test sections, added section about shared scans demonstration | Jacek Becla Daniel Wang |
3.2 | 10/10/2013 | TCT approved | R Allsman |