[
https://issues.apache.org/jira/browse/CASSANDRA-3411?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jonathan Ellis updated CASSANDRA-3411:
--------------------------------------
Attachment: 3411-v5.txt
v5 attached. minor changes like log lines to debug. also changed the
LogRecordAdder rotate check to >=.
I think we still need to solve the segment sizing problem though. If you don't
want to actually record the high water mark, how about recycling the segments
we just replayed?
Basically I don't want to inflict bad performance on people who have > 500MB of
unflushed data while CL "extends" itself. Once is acceptable, but every server
restart is not.
I think the logic around the size cap needs some work too. We should be
*strict* about keeping under the cap -- the idea is we could be given a
partition that size, so overflowing is Bad.
If we get stricter there then I think we can simplify the "high water mark"
tracking: you never discard segments (since you never violate the size
constraints) so the high water mark must be exactly the current number of
segments.
submitCreateFreshSegment is unused. Oversight?
Why does recycle bother creating a new Segment object? (Contrawise, I think
I'd prefer creating a new SequentialWriter object, to using seek here.)
Is the order in CL.shutdownBlocking correct? Seems to me it would be less
dangerous to leave the allocator active until after we're done processing
mutations.
> Pre-allocated, Recycled Commitlog Segment Files
> -----------------------------------------------
>
> Key: CASSANDRA-3411
> URL: https://issues.apache.org/jira/browse/CASSANDRA-3411
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Reporter: Rick Branson
> Priority: Minor
> Attachments:
> 001-pre-allocated-recycled-commitlog-segment-files-CASSANDRA-3411.patch,
> 003-pre-allocated-recycled-commitlog-segment-files-CASSANDRA-3411.patch,
> 004-pre-allocated-recycled-commitlog-segment-files-CASSANDRA-3411.patch,
> 3411-cleaned.txt, 3411-v5.txt
>
>
> An approach for improving commitlog performance is to pre-allocate the full
> 128MB segment files and reuse them once all the mutations have been flushed.
> Pre-allocation allows writes to be performed without modifying the file size
> metadata, and should (in theory) allow the filesystem to allocate a
> contiguous block of space for the file. Recycling the segment files prevents
> the overhead of pre-allocation from impacting overall performance.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
