[
https://issues.apache.org/jira/browse/CASSANDRA-1930?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12977126#action_12977126
]
Jonathan Ellis commented on CASSANDRA-1930:
-------------------------------------------
The reason we originally switched to a fair policy was, without it there is no
guarantee that the writer (the flushing thread) will ever get the lock, and we
were seeing OOMing from the memtable exhausting memory instead of getting
switched out.
Looks like we need a bigger change to fix both problems.
> db.Table flusherLock write lock fairness policy is sub-optimal
> --------------------------------------------------------------
>
> Key: CASSANDRA-1930
> URL: https://issues.apache.org/jira/browse/CASSANDRA-1930
> Project: Cassandra
> Issue Type: Improvement
> Components: Core
> Affects Versions: 0.3, 0.4, 0.5, 0.6, 0.6.1, 0.6.2, 0.6.3, 0.6.4, 0.6.5,
> 0.6.6, 0.6.7, 0.6.8, 0.6.9, 0.7 beta 1, 0.7 beta 2, 0.7 beta 3, 0.7.0 rc 1,
> 0.7.0 rc 2, 0.7.0 rc 3, 0.7.0, 0.7.1, 0.8
> Reporter: Kelvin Kakugawa
> Assignee: Kelvin Kakugawa
> Fix For: 0.8
>
> Attachments: CASSANDRA-1930.0001.patch
>
>
> h4. scenario:
> 1) high write throughput cluster
> 2) external services adding material cpu contention
> h4. symptoms:
> The row mutation stage falls *very* behind.
> h4. cause:
> ReentrantReadWriteLock's fair policy causes write lock acquisition / release
> to require a material amount of CPU time.
> h4. summary:
> When there are other services contending for the CPU, the RRW lock's fair
> policy causes write lock acquisition / release to take enough time to
> eventually put threads waiting for read lock acquisition very behind. We
> repro'd this scenario by reducing the scope of the write lock to: 1 boolean
> check, 1 boolean assignment, and 1 db.Memtable instantiation (itself, just: 2
> variable assignments) w/ the same performance. Modifying the fairness policy
> to be the default policy allowed the row mutation stage to keep up.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
