[
https://issues.apache.org/jira/browse/CASSANDRA-11301?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15186596#comment-15186596
]
Stefania commented on CASSANDRA-11301:
--------------------------------------
I can see what the problem described by [~benedict] is, and from code
inspection it should only affect 2.2 since prior to that the throttled
compressed reader was created without an owner and in 3.0+ the reader's pool
has been removed. The patch below should fix the issue in principle, but I have
not been able to reproduce the problem locally.
I tried comparing {{cassandra-stress}} read rates for 1M entries using 50 or
100 threads before and after a repair with very low compaction throughput. I
also tried a verify operation. I was expecting the rate to drop but I saw no
difference - presumably there are non-rate-limited readers also available in
the cache.
Does anyone have a reliable way to reproduce this or a safe environment where
to test the patch?
|[patch|https://github.com/stef1927/cassandra/commits/11301-2.2]|[testall|http://cassci.datastax.com/view/Dev/view/stef1927/job/stef1927-11301-2.2-testall/]|[dtest|http://cassci.datastax.com/view/Dev/view/stef1927/job/stef1927-11301-2.2-dtest/]|
> Non-obsoleting compaction operations over compressed files can impose rate
> limit on normal reads
> ------------------------------------------------------------------------------------------------
>
> Key: CASSANDRA-11301
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11301
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Benedict
> Assignee: Stefania
> Fix For: 2.2.6
>
>
> Broken by CASSANDRA-9240; the rate limiting reader passes the ICompressedFile
> interface to its parent, which uses this to attach an "owner" - which means
> the reader gets recycled on close, i.e. pooled, for normal use. If the
> compaction were to replace the sstable there would be no problem, which is
> presumably why this hasn't been encountered frequently. However validation
> compactions on long lived sstables would permit these rate limited readers to
> accumulate.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
