[
https://issues.apache.org/jira/browse/AURORA-1321?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14546407#comment-14546407
]
Maxim Khutornenko commented on AURORA-1321:
-------------------------------------------
I was unable to repro this problem on master. In fact, the batch version of the
api behaved worse than the original version due to dynamic SQL. Digging
further, I figured this change obliterated the perf problem completely:
https://reviews.apache.org/r/34046/. Specifically, setting the "MV_STORE=false"
turned off the H2 MVCC, which is infamous for its non-linear insert
performance:
http://h2-database.66688.n3.nabble.com/Non-linear-insert-times-in-table-with-FK-td4024729.html.
With the MVCC off, inserting 200K instance events takes only about 4.5 seconds.
> Create bulk version of the saveJobInstanceUpdateEvent db store method
> ---------------------------------------------------------------------
>
> Key: AURORA-1321
> URL: https://issues.apache.org/jira/browse/AURORA-1321
> Project: Aurora
> Issue Type: Bug
> Components: Scheduler
> Reporter: Maxim Khutornenko
> Assignee: Maxim Khutornenko
>
> Perf analysis revealed unsatisfactory performance of the job update instance
> insertion implementation during snapshot recovery. It may take up to 5
> minutes to process a total of ~100K instance events during scheduler restart.
> Implement and benchmark the bulk insert version of this api:
> https://github.com/apache/aurora/blob/b436ec52101cfd5d6f65075a5f7a857ed5a41b5f/src/main/java/org/apache/aurora/scheduler/storage/JobUpdateStore.java#L152
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
