[jira] [Commented] (CASSANDRA-10571) ClusteringIndexNamesFilter::shouldInclude is not implemented, SinglePartitionNamesCommand not discarding the sstables it could

["Sylvain Lebresne (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-10571?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14989320#comment-14989320 ] 

Sylvain Lebresne commented on CASSANDRA-10571:
----------------------------------------------

The problem was that we should not exclude sstables if they query static columns as those are not accounted by the sstable min/max clustering values. I've pushed an addition patch to [my branch|https://github.com/pcmanus/cassandra/commits/10571] for that and test now looks roughly on par with the cassandra-3.0 branch: [utest|http://cassci.datastax.com/job/pcmanus-10571-testall/2/] and [dtests|http://cassci.datastax.com/job/pcmanus-10571-dtest/1/]. There is a bit of jitter so I've rebased the branch as of this morning and restarted the tests to make sure, but the additional commit is ready to review.

> ClusteringIndexNamesFilter::shouldInclude is not implemented, SinglePartitionNamesCommand not discarding the sstables it could
> ------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-10571
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-10571
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Aleksey Yeschenko
>            Assignee: Sylvain Lebresne
>             Fix For: 3.0.0
>
>
> Now that we use {{SinglePartitionNamesCommand}} in more places - where we'd previously use what is now {{SinglePartitionSliceCommand}} - not being able to skip sstables with non-overlapping clusterings is actually a performance regression.
> {{SinglePartitionNamesCommand::queryMemtableAndDiskInternal}} should prune sstables based on {{ClusteringIndexNamesFilter::shouldInclude}} output, and the latter must be replaced with an actual implementation instead of a {{TODO}}.
> This is also a potentially a big regression in performance for counter writes (say, with DTCS), since before 3.0, the read-before-write code would use {{collectAllData}}, and that *was* pruning sstable with non-overlapping clusterings.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)