[
https://issues.apache.org/jira/browse/AURORA-1211?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14371696#comment-14371696
]
Maxim Khutornenko commented on AURORA-1211:
-------------------------------------------
Additional point towards separating snapshot and backup writes is easing up
native log IO load. With the asynchronous backup processing, persisting
snapshot in a native log will overlay with writing a backup file. In large
clusters, this may create a significant IO hot spot and result in long (in
seconds) pauses in IO processing leading to ripple effects across other
components (including ZK session expiration and scheduler losing leadership).
> Spread snapshot creation and backup file write
> ----------------------------------------------
>
> Key: AURORA-1211
> URL: https://issues.apache.org/jira/browse/AURORA-1211
> Project: Aurora
> Issue Type: Task
> Components: Scheduler
> Reporter: Maxim Khutornenko
>
> Currently, backup files are written asynchronously and overlay with the peak
> of GC cleaning up after persisting snapshot in native log. We should
> generally avoid overlaying resource-heavy activities when possible to avoid
> troubleshooting complicated performance related issues.
> Consider delaying backup file write (e.g. for 1 minute) to clearly separate
> its impact from the snapshot GC.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
