Ocean Base Taobao

1 . OceanBase taobao-obdevelop@list.alibaba- inc.com 1 

2 . The Challenge of Big Data  Operating Data @Taobao, 2010  370 millions registered users  60 millions independent visitors per day  more than 2 billions page view per day  more than 800 millions items on line  more than 800 items sold per second  GMV：>1 billion per day  The amount of data may increase several times, or even dozens of times in the next few years.  RDBMS Sharding may be unpractical Reference：http://www.alibuybuy.com/posts/52702.html 2 

3 . Comparison of existing solutions Bigtable HBase scalability Trillions of Percolator records Dynamo (PB) 100 billions of Cassandra OceanBase records (hundreds of TB) Billions of records (TB) RDBMS 10 millions of Transaction & records consistency (hundreds of GB) Eventually Single-row Cross-table  NoSQL system consistent transaction transaction  Big data、scalability、fault-tolerant  No cross-table transaction、weak consistency model 3 

4 . Storage requirement decomposition  Oceanbase: online storage system with big data and complex data relation  Key feature of online storage: large data scale but recent write is relatively small  Separating incremental data with historical data  Historical data：large scale, distributed, SAS or SSD；  Incremental data：relatively small，centralized, memory or SSD； 4 

5 . OceanBase Architecture RootServer/ RootServer/ UpdateServer UpdateServer (Primary) (backup) Client (Java/C++) ChunkServer/ ChunkServer/ ChunkServer/ ChunkServer/ MergeServer MergeServer MergeServer MergeServer  RootServer(Master): Primary/backup，meta table / cluster management /schema…  UpdateServer(Incremental data): Primary/backup，real time write (Memory+SSD)  ChunkServer(Historical data): Distributed，static data (SAS or SSD)  MergeServer：Distributed，Merge historical and incremental data => final result  Daily Merge：historical data + incremental data => new historical data 5 

6 . OceanBase Data Distribution Incremental Data Regions （COW B+ Tree） （Meta Table） Rootserver Updateserver Chunkserver 1 Chunkserver 2 Chunkserver 3 Chunkserver 4 6/26 

7 . Oceanbase Timeline Historical data Incremental data (Chunkserver) (Updateserver) 7/26 

8 . Scalability & Reliability  Scalability  ChunkServer  Automatically add and remove machine  UpdateServer (Performance)  Memory+SSD, multiple NICs, 10Gb NIC  Reading from several backups  Reliability  ChunkServer  Replication, default 3 copies  UpdateServer & RootServer  Commit log + RAID 1  Synchronous real-time backup locally  Almost real-time backup remotely 8 

9 . Transaction and Consistency  Transaction  Centralized write transaction and distributed read transaction  Supporting cross-row and cross-table transaction  Consistency  Strong Consistency: Successful mutation should be applied at both primary and backup 9 

10 . Load balance & Lock elimination  Automatic load balance  Coordinated by RootServer  Load factor: memory and disk usage, read/write load, etc.  Data migration will not disturb regular read/write service  Lock elimination  ChunkServer: read only, no lock needed  UpdateServer: Copy-on-write B+ tree, no lock for read 10 

11 . Other Features  Other Features  Online schema change  Abandoned random disk write, well suited for SSD  Built-in data compression  Non-stop system upgrade 11 

12 . Oceanbase Performance  Oceanbase Performance  4 ChunkServer, 2 * E5520 @2.27HZ, 10 * 300GB SAS, 16GB  2 billions records, 3 * 160GB (compressed by LZO), 10KB block, random read, cache is closed  Theoretical value：4 * 180 * 10 = 7200；OceanBase：5100 12 

13 . UpdateServer Performance  UpdateServer Performance  2 * E5520 @ 2.27HZ, 24G, 1Gb NIC Size(byte) 20 100 1024 2048 QPS 78000 76000 70000 55000 Context Switch 26W 25W 21W 13W  Optimization Point  The first policy (Reduce malloc/free ): QPS > 10W  The second policy (Reduce context switch): QPS > 20W  Conclusion：UPS won’t be bottleneck 13 

14 . Online Application @Taobao  Collect  Mysql 16*2=> Oceanbase 12+2  the load of machine decreased dramatically  CTU: 2.5 billions records (2.5TB) per day  MongoDB => Oceanbase  5 instance，500GB per instance  SNS feed index：Cassandra => Oceanbase  Shop decoration system  More than 10 billions records, several hundreds of millions of real time read and write operations  stably running on line for more then 6 monthes 14 

15 . Open Source  Community  http://code.taobao.org/project/view/587/  http://oceanbase.taobao.org/  Open source time：2011/08/31  Licence: GPL  Google search: more than 90 thousands results in one month after open source  Technical Speaking  Industry: Database Technology Conference China, System Architect Conference China , Taobao Open Develop Conference  Community: OpenSourceCamp  Dependency: Pacemaker, Linux operating system 15 

16 . Oceanbase Future Plan  Supports SQL-Like query language  Supports distributed index  Column-oriented storage  Supports multi-machine parallel computing  Supports distributed UpdateServer  Integrates Hadoop MapReduce  Sustained performance optimization  Full open source  … 16