[jira] [Commented] (CASSANDRA-15080) Paxos tables should use a 4KB chunk length

["Jeff Jirsa (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CASSANDRA-15080?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16899160#comment-16899160 ] 

Jeff Jirsa commented on CASSANDRA-15080:
----------------------------------------

Mentioned this in slack, but I think hard coding this is probably wrong. There are some real and valid use cases where payloads are much larger than 4k, so those would see a meaningful regression here. There may be a middle ground, or it may make sense to be somewhat configurable.

 

> Paxos tables should use a 4KB chunk length
> ------------------------------------------
>
>                 Key: CASSANDRA-15080
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-15080
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Feature/Lightweight Transactions
>            Reporter: Jon Haddad
>            Assignee: Jon Haddad
>            Priority: Normal
>              Labels: performance
>         Attachments: flamegraph-bad-perf.svg
>
>
> In doing some research for a client on LWTs, I found that once we start pushing a node hard enough, with enough LWTs, decompressing the data in the paxos table takes up quite a bit of time.  I've attached an SVG flame graph showing about 10% of time is spent in LZ4_decompress_fast in queries hitting the paxos table.  We should be able to get a nice little performance bump from changing this to 4KB chunk length or disabling it completely.



--
This message was sent by Atlassian JIRA
(v7.6.14#76016)

---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org