[ https://issues.apache.org/jira/browse/CASSANDRA-6916?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13976895#comment-13976895 ]
Benedict commented on CASSANDRA-6916: ------------------------------------- bq. when doing anticompaction, should we not clean up old readers? (repairedSSTableWriter.finish(false, repairedAt) We have two writers open with the *same* readers here, so we only close the readers when we finish the second writer. It's a bit clunky, I know, but it's not a common occurrence to be rewriting into two places. I'll add comments. bq. do we need to move starts back in SSTableRewriter.resetAndTruncate()? If we resetAndTruncate right after doing early opening, i think we could create a gap between the start in the compacting file and the end in the written one We always open before performing any append, and open with exclusive upper bounds, i.e. we should only ever truncate back to a position we're still safe at. That said, this is definitely worthy of a comment or two. I've copied your tweaks and also added another fix: wasn't dealing with marking compacted correctly everywhere. Need to mark the compaction finished before deleting the old files, and mark the preemptively opened reader as compacting before adding it to the live set to avoid it being compacted before it's written (this was in the earlier patches but dropped out somewhere along the way) Also, I left the new HashSet() inside of the Rewriter out, as I want the provided set to be mutable, so that the users of the rewriter can have access to the currently extant versions. It's not actually necessary anywhere, but I think it will prevent future surprises. I've uploaded to the repository, but intend to add some comments to the places we mentioned. The functional stuff should be static now, if you're happy with it. > Preemptive opening of compaction result > --------------------------------------- > > Key: CASSANDRA-6916 > URL: https://issues.apache.org/jira/browse/CASSANDRA-6916 > Project: Cassandra > Issue Type: Bug > Components: Core > Reporter: Benedict > Assignee: Benedict > Priority: Minor > Labels: performance > Fix For: 2.1 beta2 > > Attachments: 6916-stock2_1.mixed.cache_tweaks.tar.gz, > 6916-stock2_1.mixed.logs.tar.gz, 6916v3-preempive-open-compact.logs.gz, > 6916v3-preempive-open-compact.mixed.2.logs.tar.gz, > 6916v3-premptive-open-compact.mixed.cache_tweaks.2.tar.gz > > > Related to CASSANDRA-6812, but a little simpler: when compacting, we mess > quite badly with the page cache. One thing we can do to mitigate this problem > is to use the sstable we're writing before we've finished writing it, and to > drop the regions from the old sstables from the page cache as soon as the new > sstables have them (even if they're only written to the page cache). This > should minimise any page cache churn, as the old sstables must be larger than > the new sstable, and since both will be in memory, dropping the old sstables > is at least as good as dropping the new. > The approach is quite straight-forward. Every X MB written: > # grab flushed length of index file; > # grab second to last index summary record, after excluding those that point > to positions after the flushed length; > # open index file, and check that our last record doesn't occur outside of > the flushed length of the data file (pretty unlikely) > # Open the sstable with the calculated upper bound > Some complications: > # must keep running copy of compression metadata for reopening with > # we need to be able to replace an sstable with itself but a different lower > bound > # we need to drop the old page cache only when readers have finished -- This message was sent by Atlassian JIRA (v6.2#6252)