hbase-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From 崔建伟 <cuijian...@xiaomi.com>
Subject Themis : implements cross-row/corss-table transaction on HBase.
Date Tue, 08 Jul 2014 07:34:19 GMT
Hi everyone, I want to introduce our open-source project Themis which implements cross-row/corss-table
transaction on HBase.

Themis follows google's percolator algorithm(http://research.google.com/pubs/pub36726.html),
which provides ACID-compliant transaction and snapshot isolation. The cross-row transaction
is based on HBase's single-row atomic semantics and doesn't use a central transaction server,
so that supports linear-scalability.

Themis depends on a timestamp server to provides global strictly incremental timestamp to
define the order of transactions, which will be used to resolve the write-write and read-write
conflicts. The timestamp server is lightweight and could achieve hight throughput(500, 000
+ qps), and Themis will batch timestamp requests across transactions in one Rpc, so that it
won't become the bottleneck of the system even when processing billions of transactions every

Although Themis could be implemented totally in client-side, we adopt coprocessor framework
of HBase to achieve higher performance. Themis includes a client-side library to provides
transaction APIs, such as themisPut/themisGet/themisScan/themisDelete, and a coprocessor library
loaded on regionserver. Therefore, Themis could be used without changing the code and logic
of HBase.

We have been validating the correctness of Themis for a few months by a AccountTransfer simulation
program, which concurrently does cross-row transactions by transferring money among different
accounts(each account is a row in HBase) and verifies total money of all accounts doesn't
change in the simulation. We have also run Themis on our production environment.

We test the performance of Themis and get comparable result as percolator. The single-column
transaction represents the worst performance case for Themis compared with HBase, the result
1) For read, the performance of percolator is 90% of HBase;
2) For write, the performance of percolator is 23% of HBase.
The write performance drops a lot because Themis uses two-phase commit protocol to achieve
ACID of transaction. For multi-row write, we improve the performance by paralleling all writes
of pre-write phase. For single-row write, we are optimizing two-phase commit protocol to achieve
better performance and will update the result when it is ready. The details of performance
result could be found in github.

The repository and introduction of Themis include:
1. Themis github: https://github.com/XiaoMi/themis/. The source code, performance test result
and user guide could be found here.
2. Themis jira : https://issues.apache.org/jira/browse/HBASE-10999
3. Chronos github: https://github.com/XiaoMi/chronos. Chronos is our open-source high-availability,
high-performance timestamp server to provide global strictly incremental timestamp for Themis.

If you find Themis interesting, please leave us comment in the mail, jira or github.

  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message