LLAP: Sub-Second Analytical Queries in Hive
•Télécharger en tant que PPTX, PDF•
7 j'aime•3,802 vues
We discuss the current state of LLAP (Live Long and Process) – the concurrent sub-second execution of analytical queries engine for Hive 2.0. LLAP is a hybrid execution model that enables performance improvement in and across queries, such as caching of columnar data with cache coherence and intelligent eviction for disaggregated storage models (like S3, Isilon, Azure), JIT-friendly operator pipelines, asynchronous I/O, data pre-fetching and multi-threaded processing. LLAP features robust machine and service failure tolerance achieved by building on top of the time-tested fault tolerant subsystems, as well as a concurrency-directed design that achieves high utilization with low latency via resource sharing, reducing overheads for multiple queries, and enabling the system to preempt tasks of lower priority without failing any query in-flight. The talk also aims to cover the novel deployment model required for hybrid execution. The elasticity demands of the system are served by a long-lived YARN service interacting with on-demand elastic containers serving as a tightly integrated DAG-based framework for query execution. We discuss the current state of the project, performance numbers, deployment and usage strategy, as well as future work, including how LLAP fits into a unified secure DataFrame access layer.
Recommandé
Recommandé
Contenu connexe
Tendances
Tendances (20)
Similaire à LLAP: Sub-Second Analytical Queries in Hive
Similaire à LLAP: Sub-Second Analytical Queries in Hive (20)
Plus de DataWorks Summit/Hadoop Summit
Plus de DataWorks Summit/Hadoop Summit (20)
Dernier
Dernier (20)
LLAP: Sub-Second Analytical Queries in Hive
- 1. Page1 © Hortonworks Inc. 2011 – 2017. All Rights Reserved LLAP: Sub-Second Analytical Queries in Hive Gunther Hagleitner
- 2. Page2 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Perf overview – a demo in a gif
- 3. Page3 © Hortonworks Inc. 2011 – 2017. All Rights Reserved
- 4. 4 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Hive 2 with LLAP: 5x Performance Boost at 10 TB Scale 0 2 4 6 8 10 12 14 16 18 0 500 1000 1500 2000 2500 Improvement(Factor) Querytime,s Hive 2 with LLAP: 5x Performance Improvement Across All Query Types (10 TB, 10 Nodes (256GB/32 cores), TPC-DS Queries) Hive 2 LLAP Improvement v Hive 1 (Factor)
- 5. 5 © Hortonworks Inc. 2011 – 2017. All Rights Reserved 0 5 10 15 20 25 30 35 40 45 50 0 50 100 150 200 250 Speedup(xFactor) QueryTime(s)(LowerisBetter) Hive 2 with LLAP averages 26x faster than Hive 1 Hive 1 / Tez Time (s) Hive 2 / LLAP Time(s) Speedup (x Factor) Hive 2 with LLAP: 25+x Performance Boost: Interactive / 1TB Scale
- 6. 6 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Apache Hive vs. Apache Impala at 10TB 10TB scale on 10 identical nodes. Hive and Impala showed similar times on most smaller queries. Hive scaled better, completing larger queries quicker with lower end-to- end time for all queries Highlights 1 10 100 1000 10000 Querytime,s Impala 2.8 Hive 2.1
- 7. 7 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Apache Hive vs. Presto on a partitioned 1TB dataset. Presto lacks basic performance optimizations like dynamic partition pruning. On a real dataset / workload Presto perform poorly without full re-writes. Example: Query 55 without re-writes = 185.17s, with re- writes = 16s. LLAP = 1.37s. Highlights
- 8. 8 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Hive LLAP: Stable Performance under High Concurrency 4x Queries, 2.8x Runtime Difference 5x Queries, 4.6x Runtime Difference Mark Concurrent Queries Average Runtime 5 7.76s 25 36.24s 100 102.89s
- 9. 9 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Hive with LLAP: Linear Scale Average and Maximum query times Mixed SQL workload of simple and complex queries 1TB Dataset Size 0 20 40 60 80 100 120 8 16 32 Time(s) Concurrent Users Average Query Time by Concurrency, 1TB Average Time: 8 Node Average Time: 16 Node 0 50 100 150 200 250 300 8 16 32 Time(s) Concurrent Users Maximum Query Time by Concurrency, 1TB Max Time: 8 Node Max Time: 16 Node Linear Scale: Double Nodes, Consistent Runtime
- 10. Page10 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Performance – cache on HDFS, 1Tb scale 20.2 18.3 17.6 16.2 16.2 16.8 14.5 11.3 0.00% 20.00% 40.00% 60.00% 80.00% 100.00% 0 5 10 15 20 25 24 26 28 30 32 34 36 38 40 42 Cachehitrate Queryruntime,s Cache size, Gb Query runtime, after unrelated workload Query runtime, after related workload Metadata hit rate Data hit rate
- 11. Page11 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Performance – cache on cloud storage • Cloud storage, slower than on-prem HDFS • Large scan on 1Tb scale • Local HD/SDD cache • Massive improvement on slow storage with little memory cost 0 50 100 150 200 250 300 350 400 Cold cache Hot cache Runtime,s Entire query Scan portion
- 12. Page12 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Hive with LLAP on text data (CSV, logs, etc) - 5.00 10.00 15.00 20.00 25.00 30.00 35.00 40.00 45.00 50.00 0 200 400 600 800 1000 1200 1400 Text: No Cache Text: Cached Improvement (Factor)
- 13. Page13 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Hive with LLAP: Architecture Overview Deep Storage YARN Cluster LLAP Daemon Query Executors LLAP Daemon Query Executors LLAP Daemon Query Executors LLAP Daemon Query Executors Query Coordinators Coord- inator Coord- inator Coord- inator HiveServer2 (Query Endpoint) ODBC / JDBC SQL Queries In-Memory Cache (Shared Across All Users) HDFS and Compatible S3 WASB Isilon
- 14. Page14 © Hortonworks Inc. 2011 – 2017. All Rights Reserved I/OExecution LLAP: In-memory processing Decoder Distributed FS Fragment Hive operator Hive operator Vectorized processin g col1 col2 Low-level pushdown (e.g. filter)* SSD/HD cache Off-heap cache Compression codec Native data vectors Compact encoded data Compressed data col1 col2
- 15. Page15 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Parallel queries – priorities, preemption • Lower-priority fragments can be preempted • For example, a fragment can start running before its inputs are ready, for better pipelining; such fragments may be preempted • LLAP work queue examines the DAG parameters to give preference to interactive (BI) queries LLAP QueueExecutor Executor Interactive query map 1/3 … Interactive query map 3/3 Executor Interactive query map 2/3 Wide query reduce waiting Time ClusterUtilization Long-Running Query Short-Running Query
- 16. Page16 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Sidebar: Relational data node • LLAP can provide a "relational datanode" view of the data • Standard interface to read tables directly from LLAP nodes • Supports column-level security • LLAP handles ACID, schema-evolution • Optimized for fast access/scans • Read horizontal table partitions in parallel on compute nodes • Push projections, filters, aggregates into LLAP • Utilizes cache and LLAP IO subsystem
- 17. Page17 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Example - SparkSQL integration • Allowing direct access to Hive table data from Spark breaks security model for warehouses secured using Ranger or SQL Standard Authorization • Other features may not work correctly unless support added in Spark • Row/Cell level security (when implemented), ACID, Schema Evolution, … • SparkSQL has interfaces for external sources; they can be implemented to access Hive data via HS2/LLAP • SparkSQL Catalyst optimizer hooks can be used to push processing into LLAP (e.g. filters on the tables) • Some optimizations, like dynamic partition pruning, can be used by Hive even if SparkSQL doesn't support them; if execution pushdown is advanced enough
- 18. Page18 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Example - SparkSQL integration – execution flow Page 18 HadoopRDD.compute() LlapInputFormat.getRecordReader() - Open socket for incoming data - Send package/plan to LLAP Check permissions Generate splits w/LLAP locations Return securely signed splits HS2/Metastore Spark Executor HadoopRDD.getPartitions() Get Hive splits Cluster nodes Verify splits Run query plan Send data back LLAP Partitions/ Splits Request splits : : var llapContext = LlapContext.newInstance( sparkContext, jdbcUrl) var df: DataFrame = llapContext.sql("select * from tpch_text_5.region") DataFrame for Hive/LLAP data
- 19. Page19 © Hortonworks Inc. 2011 – 2017. All Rights Reserved LLAP Availability • First version shipped in Apache Hive 2.0 • Hive 2.0.1 contains important bug fixes to make it production ready • Hive 2.1 has new features and further perf improvements • HDP 2.5 includes beta version of LLAP • HDP 2.6 includes GA version of Apache Hive 2.1 w/ LLAP
- 20. Page20 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Setup: Ambari • Simple setup • Integrated into Hive configuration page • Launches HS2 instance and LLAP cluster • Handles configuration, slider, security, alerts and monitoring, … • Requires Ambari 2.5.x • Requires HDP 2.6.x stack (includes Apache Hive 2.1 w/ LLAP)
- 21. Page21 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Setup: Cloud • Spin up LLAP clusters in minutes • Exposes HS2 endpoint • JDBC/ODBC • Zeppelin/Hive view also options • RDS + S3 friendly • Use HD cache for fast S3 access • RDS for shared metastore instance • HDP 2.6 available now!
- 22. Page22 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Monitoring • Simple UI for monitoring • JMX endpoint with more data • Logs and jstack endpoints
- 23. Page23 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Monitoring • HDP ships with number of default Grafana dashboards for LLAP • Show wealth of metrics for cache, compute and scheduling • Helps to identify bottlenecks and issues • Helps with capacity planning • Easy to extend or integrate with other systems
- 24. Page24 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Watching queries – Tez UI integration
- 25. Page25 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Future work • Fine-grained workload management • Reservations & priorities • Configurable guardrails • Maximum permissible scan range, runtime, cost • Integrated with CBO’s cost model • Admin tools • Identify hotspots & data layout issues • Easy reporting on health, performance and utilization
- 26. Page26 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Summary • LLAP is a new execution substrate for fast, concurrent analytical workloads, harnessing Hive vectorized SQL engine and efficient in-memory caching layer • Provides secure relational view of the data through a simple API • Available in Hive 2, integrated with ecosystem (Ambari, Cloud, Grafana, jmx, Hive & Tez views)
- 27. Page27 © Hortonworks Inc. 2011 – 2017. All Rights Reserved Questions? ? Interested? Stop by the Hortonworks booth to learn more
Notes de l'éditeur
- Introduce myself Set stage for demo
- Llap off -> 10s Llap on -> < 1s
- Observations: -> same hive, same interface (only ‘mode’ difference) -> huge speed up, esp significant when working online (tableau, ad hoc) -> always on (+ cache, memory) v on demand -> why containers? Throughput, shared cluster Rest of presentation: More details about performance and behavior, then technical details
- TPCDS (decision support: ad-hoc, reporting, iterative OLAP, data mining) TPCDS, 10TB, Hive 2 (llap) v Hive 1 (containers) (availability: hive 2.0+ has llap) Benchmark on 10 nodes (There are new techniques to bring down the long running queries, not included (optimizer v runtime)) Things to note: -> red line is factor improvement (< 15x, average 5x), tremendous improvement across the board -> analytical queries (ad hoc, drill down, etc) very fast (< 5s, 10s) -> heavy lifting still takes time (other ways to handle) Average is 5x
- This time: Choose interactive (adhoc, iterative) queries, smaller data set. Avg is 25x, Blue hive 2 times hard to see on this graph. All hive 2 queries are low seconds again. -> Good for working w/ data set.
- HDP 2.5 ORC/ZLib, CDH 5.8 Parquet/Snappy
- Hive version = Hive 2.1 Presto version = Presto 0.163
- Additional details: Queries are 20 interactive queries taken from TPC-DS using 1 TB of data from TPC-DS schema. Queries run at random using a jmeter test harness Average moves linearly System tries to handle short queries first, w/o starving longer ones
- -> cache only speed up on hdfs: llap benefits always there, cache size is variable -> HDFS - this is speed up over: locality + buffer cache -> unrelated workload: query stays the same -> Metadata cache is immediately useful, even if cache is small -> Cache size doesn’t have to be as big as whole data set (pushdown proj + filter, hot data set) -> 2x speed up when all is cached
- Story is different when there is no locality
- - Text is a horrible format for analytical engines: verbose, low information density, row major, inefficient serialization of data - Typically: ETL step to transform text data to columnar format - LLAP: use it directly, first access will transform and cache the data Cache can be HD or SD (large cache on slower medium still preferable) Steady state: Similar performance to ORC
- Start left: No change to clients, HS2 still enpoint for ODBC, JDBC Number of coordinators, each one handles one query at a time (limits concurrency), run in yarn These coordinators break the query into “fragments (part of the plan + part of the data)” and schedule the work out to the daemons Coos let HS2 know the status of the query execution LLAP daemons consist of 2 parts IO subsystem: Cache + elevator, cache is shared Executors: Multi threaded query execution Storage can be either local (hdfs) or remote (s3, wasb, etc) -> Perf: Daemons + coos run permanently, no spin up cost Daemons + coos run fixed set of code: JIT gains Executors are multi threaded, more use of cluster Cache avoids reload of data Multi-threaded IO
- Left side: Fragment (part of plan + part of data) Operators need to work on data: -> setup will initialize operators + IO -> IO scan (table, range, filters, projections) -> pool of column vectors (consumer/producer) Pushdown: I/O has metadata cache (column index, min/max, bloom) Will jump over unnecessary data Cache: Read through + coherent Trade-off: Encoding (decoder + cpu v space), dictionary + runlength Off heap (GC), ssd/hd Eviction: lrfu (don’t swap entire table for 1 scan) Cache locality (multiple v wait) Text conversion! Performance: Threading: No more ping/pong, independent IO/proc Multiple fragments in same process JIT Push down Metadata cache: Don’t need to read headers/footers/index Primary data cache: Don’t touch disk
- Scheduling: Each llap daemon has queue of fragments Coordinators submit these fragments These fragments can be pre-empted even after they start running Might be running but waiting for input Short running query comes in to interrupt long running (to keep system responsive)
- - Think distributed result set or relational datanode - Granular security
- Quick way to setup on premise Pre-emption is required for guaranteed startup Pick queue, #nodes, #concurrent queries Done! Under the cover computes: Memory sizes, cache size, #executors, jvm parameters, etc etc All of that can be fine tuned (later or on advanced page) Will use slider to spin up llap, check RM UI and monitoring pages for more info
- Hortonworks data cloud Get it on the amazon marketplace, check hwx.com for details HDC allows for transient + permanent clusters, multiple use cases, single view on data EDW-Analytics based on LLAP