Hbase Nosql
•
45 likes•3,097 views
HBase is an open source, distributed, column-oriented database modeled after Google's Bigtable that runs on top of Hadoop. The presenter discusses HBase's architecture, performance improvements in version 0.20 including major gains from new file formats and compression, and Stumbleupon's extensive use of HBase including supporting over 9 billion rows in a single table with high import and read speeds.
Report
Share
More Related Content
What's hot
What's hot (20)
Viewers also liked
Viewers also liked (20)
Similar to Hbase Nosql
Similar to Hbase Nosql (20)
More from elliando dias
More from elliando dias (20)
Recently uploaded
Recently uploaded (20)
Hbase Nosql
- 1. HBase Ryan Rawson Sr Developer @ SU, HBase commi8er June 11th, NOSQL
- 2. Quick Backstory • Needed large data store @ SU • Started looking back in Jan ‘09 • Looked at the field of stores, tried: – Cassandra – Hypertable (fast) – HBase • Ended picking HBase NOSQL Meetup
- 3. Now • Personally rewri8en large porRons of HBase for 0.20 – Code easy to work with, understand, modify • Recently voted to commi8er status (thanks!) • Now giving presentaRons (hi!) NOSQL Meetup
- 4. Four Point Agenda • What is HBase? • Why HBase? • HBase 0.20 • HBase At Stumbleupon NOSQL Meetup
- 5. What is HBase? • Clone of Bigtable ‐ h8p://labs.google.com/papers/bigtable.html • Created originally at Powerset in 2007 • Hadoop‐subproject – The usual ASF things apply (license, JIRA, etc) NOSQL Meetup
- 6. What is HBase? • Column‐oriented semi‐structured data store • Distributed over many machines – Bigtable known to scale to >1000 nodes • Tolerant of machine failure • Layered over HDFS (& KFS) • Strong consistency (important) NOSQL Meetup
- 7. Table & Regions • Rows stored in byte‐lexographic sorted order • Table dynamically split into “regions” • Each region contains values [startKey, endKey) • Regions hosted on a regionserver NOSQL Meetup
- 8. Table & Regions NOSQL Meetup
- 9. Column Storage • In HBase, don’t think of a spreadsheet: All columns same ‘size’ and present (as NULL) NOSQL Meetup
- 10. Column Storage • Instead think of tags. Values any length, no predefined names or widths: Column names carry info (just like tags) NOSQL Meetup
- 11. Column Families • Table consists of 1+ “column families” • Column family is unit of performance tuning • Stored in separate set of files • Column names scoped like so: – “Family:qualifier” NOSQL Meetup
- 12. SorCng • Rows stored in byte‐lexographical order (row keys are raw bytes, not just strings) • Furthermore within a row, columns stored in sorted order • Fast, cheap easy to scan adjacent rows & columns NOSQL Meetup
- 13. SorCng (but there’s more!) • Not just scanning, but can do parRal‐key lookups • When combined with compound keys, has the same properRes as leading‐lel edge indexes in standard RDBMS – (Except your index is distributed of course) • Can use a second table to index a primary table. NOSQL Meetup
- 14. Values • Row id, column name, value all byte [] • Can store ascii, any binary or use serializaRon (eg: thril, protobuf) • Atomic increments available • SerializaRon good for structs that are always read in one unit (eg: Address book entry) NOSQL Meetup
- 15. Values & Versions • Each row id + column – stored with Rmestamp • HBase stores mulRple versions • Can be useful to recover data due to bugs! • Use to detect write conflicts/collisions NOSQL Meetup
- 16. API Example Scan scan = new Scan(startRow, endRow).addFamily(“family”); ResultScanner scanner = table.getScanner(scan); Result result; while ( (result=scanner.next()) != null) { EnRty e = new EnRty(); dser.deserialize(e, result.getValue("default”, “0”); } scanner.close(); NOSQL Meetup
- 17. Why HBase? • Community is highly acRve, diverse, helpful • User list Email acRvity for May: 78 threads • IRC Channel #hbase highly acRve • Helpful people in mulRple Rmezones, email answered all hours of the day/night/weekend. NOSQL Meetup
- 18. Why HBase? • Commi8er & contributor base broad: – PSet, Streamy, SU, Trend Micro, Openplaces, and more! • No monopoly on experts – deep knowledge at these companies and more! • (We’re really friendly… honest!) NOSQL Meetup
- 19. Why HBase? • Used in producRon at many companies • 12 companies listed on h8p://wiki.apache.org/hadoop/Hbase/ PoweredBy • Openplaces, Streamy, SU serve websites out of HBase • Lots of experience to draw upon! NOSQL Meetup
- 20. Why HBase? (Features) • Full web management/monitoring UI (master & regionservers) • Push metrics to log files & Ganglia • Rolling upgrades possible! (Including master!) • Non‐SQL shell – re‐enforces the non‐SQL‐ness of HBase NOSQL Meetup
- 21. HBase Features • Easy integraRon with Hadoop MR – table input and output formats ship • Cascading connectors for input and output • Other ancillary open source acRviRes around the edges (ORM, schema management, etc) NOSQL Meetup
- 22. Why HBase? • But… HBase is slow! • That metabrew/last.fm blog post said so! – (Also other people too…) • “It’s much more than a KV store, but latency is too great to serve data to the website.” • Answer: 0.20 NOSQL Meetup
- 23. HBase 0.20 • Two major and exciRng themes: • #1: Performance • #2: ZooKeeper integraRon, mulRple masters NOSQL Meetup
- 24. HBase 0.20 vs 0.19 0.19 0.20 Master Single master – if it fails, so Master elecRon and does the cluster membership via ZK Compression Not really GZ, LZO Memory usage Small values cause big New file‐format limits indexes and OOM index size (800kB for 10m entries) Scan Speed 300‐600ms per 500 rows 20‐30ms per 500 rows NOSQL Meetup
- 25. Zookeeper? • A highly available configuraRon storage system • Set up in a 2N+1 quorum • Hadoop subproject NOSQL Meetup
- 26. Master & Zookeeper • Store membership info in ZK • Detect dead servers (via ephemeral nodes) • Master elecRon and recovery • Can kill master and cluster conRnues • New master determines state and conRnues NOSQL Meetup
- 27. Performance • Significant performance gains in 0.20 • New file format with 0‐copy infrastructure • Scan and get improvements • LZO compression • Block caching • Speed increases as much as 30x! NOSQL Meetup
- 28. Performance • 0.20 is not the final word on performance: • Other RPC‐related performance improvements • Other Java‐related improvements (G1?, 1.7?) NOSQL Meetup
- 29. Performance Numbers • 1m rows, 1 column per row (~16 bytes) – SequenRal insert: 24s, .024ms/row – Random read: 1.42ms/row (avg) – Full Scan: 11s, (117ms/10k rows) • Performance under cache is very high: – 1ms to get single row – 20 ms to read 550 rows – 75ms to get 5500 rows NOSQL Meetup
- 30. HBase at Stumbleupon • Strong commitment to HBase @ SU • Supports a HBase commi8er • Looking to hire more HBase hackers NOSQL Meetup
- 31. Big accomplishments @ SU • Over 9b small rows in single table – Sustained import performance – 3‐4 days to import 9b rows (mysql limiRng speed) • 1.2m row reads/sec on 19 nodes (!!) – That is 60‐100k reads/sec/node sustained, 2hrs – Scalable with more nodes – HBase has been improved since then NOSQL Meetup
- 32. Fast accomplishments @ SU • Extremely high speed increments and writes • Supports su.pr analyRcs • Su.pr reads from HBase with no intervening caches • Integrated with PHP NOSQL Meetup
- 33. HBase & PHP @ SU • PHP access via Thril gateway • Easy (PHP) deployment with Thril • App developers like sol‐schema, easy querying and wriRng • Want to use HBase for more features and applicaRons! NOSQL Meetup
- 34. HBase deployment trivia • Nodes are 8x16 w/2TB (best price point) – Don’t use RAID1. Use RAID0 or JBOD support • Ganglia allows overall cluster performance monitoring • Clusters won’t span datacenters – We want fully duplicate data for DR anyways • Update master with code & config – Rsync to other nodes (1 dir, very easy) – Controlled restart for rolling upgrade NOSQL Meetup
- 35. HBase deployment trivia • HDFS – set xciever limit to 2048, Xmx2000m – Never get HDFS problems even under heavy load • For 9b row import, randomized key insert order gives substanRal speedup • Give HBase enough ram, you wouldn’t starve mysql! • Import speeds of 200k ops/sec on 19 machines possible! – Hard to provide a SQL‐based source fast enough – 100k ops/sec typical for sustained NOSQL Meetup
- 36. HBase deployment trivia • Consider dual writes or logs to get HBase up to date but without actually moving your data • Duplicate data in indexes (already done in mysql) • Have to think about read pa8erns when designing table key order! NOSQL Meetup
- 37. HBase future @ SU • Latency sensiRve cluster • Batch/analyRcs cluster • Use replicaRon to keep la8er up to date • Allows batch jobs to go full thro8le against reasonably up to date data without risking the website NOSQL Meetup
- 38. Q&A • QuesRons? • Stumbleupon is hiring awesome HBase hackers! NOSQL Meetup