You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@cassandra.apache.org by "Jordan West (Jira)" <ji...@apache.org> on 2020/11/18 16:35:00 UTC
[jira] [Updated] (CASSANDRA-16285) Change Dynamic Snitch Default
Badness Threshold to 1.0
[ https://issues.apache.org/jira/browse/CASSANDRA-16285?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]
Jordan West updated CASSANDRA-16285:
------------------------------------
Change Category: Performance
Complexity: Low Hanging Fruit
Status: Open (was: Triage Needed)
> Change Dynamic Snitch Default Badness Threshold to 1.0
> ------------------------------------------------------
>
> Key: CASSANDRA-16285
> URL: https://issues.apache.org/jira/browse/CASSANDRA-16285
> Project: Cassandra
> Issue Type: Improvement
> Components: Consistency/Coordination
> Reporter: Jordan West
> Assignee: Jordan West
> Priority: Normal
> Attachments: readcount-0.1.png, readcount-1.0.png, readlatency-0.1.png, readlatency-1.0.png
>
>
> With the removal of compaction and IO from the DynamicEndpointSnitch score calculation, the default badness threshold of 10% (0.1) is too small of a margin from experience with production clusters. When compaction and IO values were included, the resulting scores were dominated by them and 10% was a much more noticeable difference. When relying solely on latency, the DynamicEndpointSnitch can rely on nodes that are performing only marginally better than their peers. This results in a lopsided request distribution among the replicas despite similar performance.
> Some graphs are attached from a production cluster showing the read count and latency among the nodes with the default of 0.1 and with the badness threshold set to 1.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@cassandra.apache.org
For additional commands, e-mail: commits-help@cassandra.apache.org