[
https://issues.apache.org/jira/browse/CASSANDRA-12876?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15655141#comment-15655141
]
Per Otterström edited comment on CASSANDRA-12876 at 11/10/16 9:04 PM:
----------------------------------------------------------------------
Thanks for the input [~alprema]!
So, it appears that overflow doesn't happen during rescale, but when mean value
is calculated. Triggered when all bucket values are added together. The sudden
stop of negative values happened when rescale was triggerd.
The attached patch will perform a rescale of the snapshot values when the
snapshot is created, thereby avoiding overflow. As a side effect, this will
make the max value to behave better. Before max value would continue to raise
and then get a sudden drop every 30 minutes when rescale happend. Now, max
value will appear to decay on a minute basis, whether value is peaking just
before or after rescale.
The patch includes a missing part from CASSANDRA-11752 that never were merged
into 2.2 branch. It seem to apply clean on 3.0.x as well.
was (Author: eperott):
Thanks for the input @alprema!
So, it appears that overflow doesn't happen during rescale, but when mean value
is calculated. Triggered when all bucket values are added together. The sudden
stop of negative values happened when rescale was triggerd.
The attached patch will perform a rescale of the snapshot values when the
snapshot is created, thereby avoiding overflow. As a side effect, this will
make the max value to behave better. Before max value would continue to raise
and then get a sudden drop every 30 minutes when rescale happend. Now, max
value will appear to decay on a minute basis, whether value is peaking just
before or after rescale.
The patch includes a missing part from CASSANDRA-11752 that never were merged
into 2.2 branch. It seem to apply clean on 3.0.x as well.
> Negative mean write latency
> ---------------------------
>
> Key: CASSANDRA-12876
> URL: https://issues.apache.org/jira/browse/CASSANDRA-12876
> Project: Cassandra
> Issue Type: Bug
> Components: Observability
> Reporter: Kévin LOVATO
> Assignee: Per Otterström
> Attachments: 12876-2.2.txt, negative_mean.png,
> negative_mean_details.PNG, negative_mean_periodicity.PNG
>
>
> The mean write latency returned by JMX turns negative every 30 minutes. As
> the attached screenshots show, the value turns negative every 30 minutes
> after the startup of the node.
> We did not experience this behavior in 2.1.16.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
