[ 
https://issues.apache.org/jira/browse/CASSANDRA-16161?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17224671#comment-17224671
 ] 

Alexander Dejanovski commented on CASSANDRA-16161:
--------------------------------------------------

My initial preference while reading this ticket was to use different throttles 
as well. As mentioned by [~mck], validation compactions put a slightly 
different type of pressure on nodes and folks might want to unthrottle 
validation compactions (current behavior) and keep a throttle on compactions.
But on the other hand, disks are now way faster than they used to be when 
compaction throttle was introduced and heap pressure is mostly what we're 
protecting the clusters from I guess. We used to consider at TLP that over 
45MB/s Cassandra couldn't keep up anyway in compaction sustained throughput 
because of heap pressure (it would only allow bursts). This will indeed change 
in 4.0 with all the improvements made to lower compaction heap pressure.
I also think that repair should be throttled in order to lighten its impact on 
clusters and force folks to investigate if it's not going fast enough, rather 
than harm clusters when using defaults to make it go fast.
Also, adding a new hidden configuration setting just for 3.0/3.x this close to 
4.0 going GA doesn't seem like the best thing to do.

TL;DR: +1 on using {{compaction_throughput_mb_per_sec}} to throttle validation 
compactions as well as standard compactions.

> Validation Compactions causing Java GC pressure
> -----------------------------------------------
>
>                 Key: CASSANDRA-16161
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-16161
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Local/Compaction, Local/Config, Tool/nodetool
>            Reporter: Cameron Zemek
>            Assignee: Stefan Miklosovic
>            Priority: Normal
>             Fix For: 3.11.x, 3.11.8
>
>         Attachments: 16161.patch
>
>
> Validation Compactions are not rate limited which can cause Java GC pressure 
> and result in spikes in latency.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

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

Reply via email to