On Nov 5, 2011, at 05:37 , lars hofhansl wrote: Cool stuff Daniel,
Hi Lars, Thanks for the good points. Was looking through the code a bit. Seems like you make a best effort to push as much of the filtering of KVs of uncommitted transactions to HBase and then do some filtering on the client not a bad approach. (I hope I didn't misunderstand the approach, only looked through the code for 1/2 hour or so). Putting it more accurately, the uncommitted KVs are stored at HBase, but it is the client's job to filter them using the commit information that it has received from the status oracle. According to snapshot isolation guarantee, all the versions that are inserted with a timestamp larger than the transaction start timestamp must be ignored, which is done by setting the time range on the client's get request sent to HBase. Since the uncommitted changes of the aborted transactions are eventually removed from HBase, the client rarely needs to fetch more than a version to reach a KV that is committed before the transaction starts (the first property of snapshot isolation). One thing I was wondering: Why bookkeeper? Why not store the WAL itself in HBase? That way you might not even need a separate server. Did you see: HBaseSI (http://www.cs.uwaterloo.ca/~c15zhang/HBaseSI.pdf), they also do MVCC on top of unaltered HBase/schema, although from reading that paper I get the impression that it would not scale to scans touching many rows (which is where your client side filtering comes in). Thanks for the link. We had seen the other paper of the same authors (Grid2010) that shares the same bottlenecks with the recent work. As you pointed out correctly, the question is about performance. You could see the scalability bottleneck of 400 TPS in the evaluation section of this paper. Our approach, however, provides snapshot isolation with a negligible overhead on region servers, and could scale up to tens of thousands write transactions per second. If you are interested, a summary of techniques that we used to achieve this performance is published at SOSP'11, poster section. http://sigops.org/sosp/sosp11/posters/summaries/sosp11-final12.pdf -- Lars ----- Original Message ----- From: Daniel Gómez Ferro <[email protected]<mailto:[email protected]>> To: "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>>; "[email protected]<mailto:[email protected]>" <[email protected]<mailto:[email protected]>> Cc: Maysam Yabandeh <[email protected]<mailto:[email protected]>>; Flavio Junqueira <[email protected]<mailto:[email protected]>>; Benjamin Reed <[email protected]<mailto:[email protected]>>; Ivan Kelly <[email protected]<mailto:[email protected]>> Sent: Friday, November 4, 2011 4:24 AM Subject: Omid: Transactional Support for HBase (I apologize for resending but I forgot to add the user list.) Hi all, It is my pleasure to announce the open source release of Omid, a project whose goal is to add lock-free transactional support on top of HBase. The current release includes CrSO, a client-replicated status oracle that detects the write-write conflicts to provide Snapshot Isolation. CrSO has the following appealing properties: 1) It does not need any modification into the HBase code nor the table scheme. 2) The overhead on HBase DataNodes is negligible (only after an abort) 3) It scales up to 50,000 write transactions per second (TPS) and a thousand of client connections. We have setup a github project: https://github.com/dgomezferro/omid More information is available at the wiki: https://github.com/dgomezferro/omid/wiki If you are interested, installation and running instructions are available on the README: https://github.com/dgomezferro/omid/blob/master/README.md Please do not hesitate to contact us in the case of any question. Best Regards, Daniel Gómez Ferro
