Leveled, Riak and Hashtree rebuilds

classic Classic list List threaded Threaded
1 message Options
Reply | Threaded
Open this post in threaded view

Leveled, Riak and Hashtree rebuilds

Martin Sumner
The next phase of Riak and Leveled testing is now complete, with this phase focused on achieving stable throughput during hashtree rebuilds.  this scenario has been tested due to a history of issues with the cost of hashtree rebuilds in Riak production environments, and looks at improvements offered by both leveled and the new riak_kv_sweeper feature in the KV develop branch.

The test takes a scenario which should be fairly favourable to leveldb (using SSDs with sync_on_write disabled), although still with mid-size objects favourable to leveled (8KB size).  The test tries to push throughput with aae enabled with aggressive hashtree rebuilds.  

In the 6th hour of the test Riak/leveled was achieving 11,283 ops/s compared to 7,304 ops/s with Riak/leveldb.  As before volatility of throughput and tail latency is reduced significantly with leveled, but that improvement does come with some caveats explained in the link above.

Next stage will be to address the short-comings of compaction and anti-entropy in the value store (Journal) in leveled.  Further results should be available in the next 3 or 4 weeks.


riak-users mailing list
[hidden email]