HBase Status Report - Hadoop Summit Europe 2014
- 1. 1 HBase 0.96+ A Report on the Current Status Lars George | EMEA Chief Architect
- 2. About Me • EMEA Chief Architect @ Cloudera • ConsulDng on Hadoop projects (everywhere) • Apache CommiLer • HBase and Whirr • O’Reilly Author • HBase – The DefiniDve Guide • Now in Japanese! • Contact • lars@cloudera.com • @larsgeorge 日本語版も出ました!
- 3. The Content... • Version History • Overview of new Features • Summary
- 4. CONFIDENTIAL -‐ RESTRICTED Version History A Timeline Overview
- 5. HBase Releases 5 URL: hLp://s.apache.org/hbase-‐releases
- 6. HBase Releases – Issues Closed (JIRA) 6 URL: hLp://s.apache.org/hbase-‐releases
- 7. HBase Releases – Issues Closed (DisDnct) 7 URL: hLp://s.apache.org/hbase-‐releases
- 8. HBase Book? I targeted 0.92.0 but… r1130336 | stack | 2011-06-02 00:52:45 +0200 ⤦ (Thu, 02 Jun 2011) | 1 line Add link to meet up ... r1234894 | stack | 2012-01-23 17:50:43 +0100 ⤦ (Mon, 23 Jan 2012) | 1 line Move version on past 0.92.0 to 0.92.1-SNAPSHOT $ svn log -r 1130336:1234894 | grep "^r" | wc -l 807 8
- 9. HBase Book? I am trailing 0.92.0 by 800+ commits, including for example r1153634 | tedyu | 2011-08-03 21:59:48 +0200 ⤦ (Wed, 03 Aug 2011) | 2 lines HBASE-3857 Change the HFile Format (Mikhail & Liyin) …which is not “unimportant”. J I am working on an update! 9
- 10. 10 Coprocessors and more… HBase 0.92
- 11. HBase 0.92 -‐ Highlights • 682 issues addressed • 811 issues total in 0.92.x line • New logo! (HBASE-‐4312) • HFile v2 (HBASE-‐3857) • Distributed Log Splilng (HBASE-‐1364) • Enhanced Master UI • Major compacDon progress (HBASE-‐3900) • Regions in transiDon (HBASE-‐4291) • Tasks (HBASE-‐3839) • Slow Query Metrics (HBASE-‐4117) 11
- 12. HBase 0.92 -‐ Highlights • Coprocessors (HBASE-‐2000) • Oneap cache (HBASE-‐4027) • Online Table Schema Change (HBASE-‐1730) • Regions Size from 256MB to 1GB (HBASE-‐4374) • Hadoop 1 Support (HBASE-‐5125) • Snappy Support (HBASE-‐3691) • Keep last version with TTL (HBASE-‐4071) • MulDthreaded CompacDons (HBASE-‐4572) 12
- 13. HFile v1 – HBase 0.90 13 • Previously the file layout was data blocks, meta blocks and then file metadata like indexes. • Each data block held a magic header and then the actual data sequenDally.
- 14. HFile v2 – HBase 0.92+ The 2nd version of HFile splits the indexes and Bloom filters up into a hierarchy and interleaves those with data blocks. 14 The data block header now holds addiDonal info on the block itself. Source: hLp://blog.cloudera.com/blog/2012/06/hbase-‐io-‐hfile-‐input-‐output/
- 16. Coprocessors: RPC Calls 16
- 17. Slab Cache – Off-‐heap Block Cache 17 hLp://blog.cloudera.com/blog/2012/01/caching-‐in-‐hbase-‐slabcache/ • The off-‐heap cache uses Java NIO’s Direct ByteBuffer structures • Uses its on slab allocaDon handling • Does copy-‐on-‐read during access of data • Uses L2 cache and replaces OS buffer cache
- 18. 18 Performance Tuning… HBase 0.94
- 19. HBase 0.94 -‐ Highlights • 420 issues addressed • 1394 issues total in 0.94.x line • Read Caching Improvements (HBASE-‐5074) • Seek OpDmizaDon • Bloom Filter for Delete Family (HBASE-‐4532) • Lazy Seeks (HBASE-‐4465) • Write to WAL OpDmizaDons • WAL Compression (HBASE-‐4608) • Data Block Encoding of KeyValues (HBASE-‐4218) • Improved HBaseFsck (HBASE-‐5128) 19
- 20. HBase 0.94 -‐ Highlights • Simplified Region Sizing (HBASE-‐4365) • Smarter TransacDon SemanDcs • Atomic Put&Delete in One Call (HBASE-‐3584) • Snapshots (0.94.6) (HBASE-‐7360) • Atomic Appends (HBASE-‐4102) • MulD Increment and Append (HBASE-‐2947) • More Aggressive Off-‐Peak CompacDons (HBASE-‐4463) 20
- 21. HBase 0.94 -‐ Highlights • Per Column Family Metrics (HBASE-‐4219) • MulD-‐row local transacDons (HBASE-‐5229) • Pluggable Split Key Policy (HBASE-‐5304) • Load balance regions by table (HBASE-‐3373) • Also backported to 0.92.1 • Make CompacDon Code Pluggable (HBASE-‐6427) • Deprecate HTablePool (0.94.11) (HBASE-‐6580) • Canary Test Tool (HBASE-‐4393) 21
- 22. Block Encoding • Allows to reduce data footprint in memory • Only encodes the key porDon of a key/value pair • Encoded keys stay encoded also during flushes • Compression on top of encoding takes care of the values and remainder of key data Example: • Key length: 90B • Value length: 8B 22 Type Ra0o Key Compression 92% Total Compression 85% LZO on same data 85% LZO axer encoding 91% hLps://issues.apache.org/jira/browse/HBASE-‐4218
- 23. Block Encoding: None 23 Source: hLp://blog.cloudera.com/blog/2012/06/hbase-‐io-‐hfile-‐input-‐output/ • With no encoding the Key/Value structures are stored verbaDm (with some overhead for lengths) • In the past you were advised to keep the “keys” short for that reason
- 24. Block Encoding: Prefix Encoding 24 Source: hLp://blog.cloudera.com/blog/2012/06/hbase-‐io-‐hfile-‐input-‐output/ • The encoding patch added a new Cell abstracDon that allows for extra fields in a Key/Value • The fields are used to track necessary details for the encoding
- 25. Block Encoding: Diff Encoding 25 Source: hLp://blog.cloudera.com/blog/2012/06/hbase-‐io-‐hfile-‐input-‐output/ • Apart from the prefix encoding there are other ways of encoding the keys • The diff encoding is one of such approaches
- 26. Block Encoding • Advantage of block encoding is faster decompression/ decoding • 20-‐80% faster than LZO • Also it allows to seek data sDll, which is not possible with compressed data • Penalty is a slightly slower read performance compared to non-‐encoded keys • Important is to watch the sizes and repeDDon of key data, encoding might not be useful for random data 26 hLps://issues.apache.org/jira/browse/HBASE-‐4218
- 27. 27 The Singularity HBase 0.96
- 28. HBase 0.96 -‐ Highlights • 1219 issues addressed • 2243 issues total in 0.96.x line • Improved Stability (HBASE-‐6241/6201) • ZK based Read/Write locks for table operaDons (HBASE-‐7305) • Scalability Improvements (HBASE-‐8877) • Schema Storage (HBASE-‐8778) • Log Cleaner for ReplicaDon Speed Up (HBASE-‐9208) • Mean-‐Time-‐To-‐Recovery (MTTR) Improvements (HBASE-‐5844/5926) • Distributed Log Replay (HBASE-‐7006) • Dedicated WAL for System Table (HBASE-‐7213/8631) 28
- 29. HBase 0.96 -‐ Highlights • Operability Improvements • Hooks for Health Scripts (HBASE-‐7399/7406) • Trace Lagging Calls with HTrace (HBASE-‐9121) • Versioned RPCs and Metadata (Protobufs) (HBASE-‐3505) • Parallel Seeks in Stores (HBASE-‐7495) • Hadoop 1 and 2 Support • Secure Short Circuit Reads on H2 (HBASE-‐6783) • Namespaces Support (HBASE-‐8015) • New Metrics v2 (HBASE-‐4050) • Cell Interface vs KeyValue (HBASE-‐7162) 29
- 30. HBase 0.96 -‐ Highlights • No more ROOT table (HBASE-‐3171) • Remove HFile v1 (HBASE-‐7660) • Trie Data Block Encoding (HBASE-‐4676) • Remove Client-‐side Row Locks (HBASE-‐7263/7315) • CompacDon and Flush Improvements (HBASE-‐7516/7763/6466/7678) (HBASE-‐7667/7110/7603/7519/7842) • Improved Default ConfiguraDon (HBASE-‐4657?) • Client-‐side Type Library (HBASE-‐8089) 30
- 31. HBase 0.96 -‐ Highlights • Online Region Merging (HBASE-‐7403/8219) • Bucket Cache Support (HBASE-‐7404) • Remove older ICV Calls (HBASE-‐7032) • New “Bootstrap” based UIs! (HBASE-‐6135) • Remove Client-‐side Row Locks (HBASE-‐7263/7315) • CompacDon and Flush Improvements (HBASE-‐7516/7763/6466/7678) (HBASE-‐7667/7110/7603/7519/7842) 31
- 32. 32 — Michael Stack, HBase PMC Chair
- 33. Mean-‐Time-‐To-‐Recovery (MTTR) • Lot‘s of effort put into improve how long data might not be accessible during a region move • The offline period is made up of phases: • a detecDon phase, • a repair phase, • reassignment, and finally, • clients noDcing the data available in its new locaDon • Improvements in many of those areas • Faster detecDon, efficient repair, parallel replay • Dedicated WAL for system tables 33 hLps://blog.cloudera.com/blog/2013/10/hbase-‐0-‐96-‐0-‐released/
- 34. 34 Cell Level Security HBase 0.98
- 35. HBase 0.98 -‐ Highlights • 1303 issues addressed • 1458 issues total in 0.98.x line • Cell Level Security (HBASE-‐6222/7663/7662) • Server-‐side EncrypDon (HBASE-‐7544) • WAL Throughput Improvements (HBASE-‐8755) • Reverse Scanner (HBASE-‐4811) • MapReduce over Snapshot Files (HBASE-‐8369) • Striped CompacDons (HBASE-‐7667) • ThroLle ReplicaDon (HBASE-‐9501) 35
- 36. Cell Level Security • Added HFile v3 which can store arbitrary metadata in a cell, called tags • Also extended ACL checks to apply to cell levels • With this visibility labels can be stored in tags • An API and CLI tools are provided that are akin to Accumulo’s, axer which it is modeled • AddiDonal encrypDon of data at rest ensures further security of sensiDve data 36 hLps://blogs.apache.org/hbase/entry/hbase_cell_security
- 37. Visibility Labels The API allows to set visibility by using expressions with “&”, “|”, and “!”, as well as “(“ and “)”, e.g. label set of { confidenDal, secret, topsecret, probaDonary } could be combined as ( secret | topsecret ) & !probationary At runDme the expressions are evaluated against a user and then applied to each cell. 37
- 38. 38 The Future… HBase 0.??
- 39. HBase Future • Not much is wriDng in stone yet • Master gets rewriLen and also META table handling • Build in consensus (HBASE-‐10296) • Co-‐locate Master and META (HBASE-‐10569) • MTTR is further extended into interesDng areas • Read replicas (HBASE-‐10070) It has to be seen when 1.0.0 is released and what it contains. Your opinion counts! 39