[
https://issues.apache.org/jira/browse/CASSANDRA-11035?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15154467#comment-15154467
]
Marcus Eriksson commented on CASSANDRA-11035:
---------------------------------------------
been running a few more benchmarks locally;
* as long as compaction keeps up, it is identical to the current STCS (atleast
in my short benchmarks) - we basically always compact the 4 similarly sized
sstables that we find
* once compaction can't keep up (there is a bunch of sstables to pick from) we
see big improvements
I think we should do this for STCS-in-L0 as well since the the most common (I
think) way to get behind in L0 is from streaming/repair where we drop in many
sstables in L0 from other nodes. This should result in a bunch of subsets of
sstables that are highly overlapping while not overlapping with the other
subsets at all. For example, say we have 4 levels in LCS and stream range (10,
20], then we would end up with 4 sstables with that range, and compacting those
4 sstables together first should produce good results.
> Use cardinality estimation to pick better compaction candidates for STCS
> (SizeTieredCompactionStrategy)
> -------------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-11035
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11035
> Project: Cassandra
> Issue Type: Improvement
> Components: Compaction
> Reporter: Wei Deng
> Assignee: Marcus Eriksson
>
> This was initially mentioned in this blog post
> http://www.datastax.com/dev/blog/improving-compaction-in-cassandra-with-cardinality-estimation
> but I couldn't find any existing JIRA for it. As stated by [~jbellis],
> "Potentially even more useful would be using cardinality estimation to pick
> better compaction candidates. Instead of blindly merging sstables of a
> similar size a la SizeTieredCompactionStrategy." The L0 STCS in LCS should
> benefit as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
