[
https://issues.apache.org/jira/browse/CASSANDRA-5722?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13710268#comment-13710268
]
Jonathan Ellis commented on CASSANDRA-5722:
-------------------------------------------
Sorry, I was really fixating on trying to relate your comment to that commit.
I see what you mean, now.
That's actually the way it used to work, pre-CASSANDRA-4710 -- tldr, in
Daniel's workload, the decoration was indeed higher than the extra work from
scanning the extra rows.
If you think about it, you can see why that might be so -- we only have to scan
extra rows on a bloom filter false positive. The common case by the time we
start looping through rows is that the row we're looking for exists.
Now that we allow disabling BF entirely that might not always be the case, but
if you're disabling your BF I have to assume you know what you're doing and are
prepared for the consequences. :)
> Cleanup should skip sstables that don't contain data outside a nodes ranges
> ---------------------------------------------------------------------------
>
> Key: CASSANDRA-5722
> URL: https://issues.apache.org/jira/browse/CASSANDRA-5722
> Project: Cassandra
> Issue Type: Improvement
> Reporter: Nick Bailey
> Assignee: Tyler Hobbs
> Fix For: 2.0.1
>
> Attachments: 0001-Skip-cleanup-when-unneeded.patch
>
>
> Right now cleanup is optimized to simply delete sstables that *only* contain
> data that doesn't belong on the node, for all other sstables though, it will
> read them, check each row, and write out new sstables.
> Cleanup could be optimized to look at an sstable and determine that all data
> within the sstable does belong on a node, and therefore skip re-writing that
> sstable. This would make cleanup essentially a noop in the case where all
> data on a node belongs on that node.
--
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
