cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sylvain Lebresne (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-4883) Optimize mostRecentTomstone vs maxTimestamp check in CollationController.collectAllData
Date Wed, 31 Oct 2012 09:53:13 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-4883?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13487644#comment-13487644
] 

Sylvain Lebresne commented on CASSANDRA-4883:
---------------------------------------------

As said in CASSANDRA-3306, I suggest using this issue to change the field {{sstables}} in
DataTracker.View from a list to a set (or at the very least removing the obsolete comment).
See the comments of CASSANDRA-3306 for more details.

I've targeted this to 1.2 out of conservatism (this is just an optimization after all, and
one that will affect a rather critical part of the code). 
                
> Optimize mostRecentTomstone vs maxTimestamp check in CollationController.collectAllData
> ---------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-4883
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-4883
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Sylvain Lebresne
>            Priority: Minor
>             Fix For: 1.2.0
>
>
> CollationController.collectAllData eliminates a sstable if we've already read a row tombstone
more recent that its maxTimestamp. This is however done in 2 passes and can be inefficient
(or rather, it's not as efficient as it could). More precisely, say we have 10 sstables s0,
... s9, where s0 is the most recent and s9 the least one (and their maxTimestamp reflect that)
and s0 has a row tombstone that is more recent than all of s1-s9 maxTimestamps. Now in collectAllData(),
we first iterate over sstables in a "random" order (because DataTracker keeps sstable in a
more or less random order). Meaning that we may iterate in the order s9, s8, ... s0. In that
case, we will end up reading the row header from all the sstable (hitting disk each time).
Then, and only then, the 2nd pass of collectAllData will eliminate s1 to s9.
> However, if we were to iterate sstable in maxTimestamps order (as we do in collectTimeOrdered),
we would only need one pass but more importantly we would minimize the number of row header
we read to perform that sstable eliminination. In my example, we would only ever read the
row tombstone from s0 and eliminate all other sstable directly, simply based on their maxTimestamp.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message