[
https://issues.apache.org/jira/browse/HBASE-21150?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16603782#comment-16603782
]
Ted Yu commented on HBASE-21150:
--------------------------------
bq. you use a thread pool to execute the registration
The pool is used to avoid adding latency to region open.
bq. will still do registration when getting
As I mentioned in the first comment of the JIRA, additional logging at the
above place didn't show up in the test output.
Meaning, by the time flushing needs to perform metrics operation, the
registration is already complete.
The call to helper above is a safe net - not to be executed under normal
situation.
Also, the helper doesn't rely on computeIfAbsent - making it suitable to be
backported (as Xu has done).
> Avoid delay in first flushes due to overheads in table metrics registration
> ---------------------------------------------------------------------------
>
> Key: HBASE-21150
> URL: https://issues.apache.org/jira/browse/HBASE-21150
> Project: HBase
> Issue Type: Bug
> Reporter: Ted Yu
> Assignee: Ted Yu
> Priority: Major
> Attachments: 21150.v1.txt, 21150.v2.txt, 21150.v3.txt
>
>
> After HBASE-15728 is integrated, the lazy table metrics registration results
> in penalty for the first flushes.
> Excerpt from log shows delay (note the same timestamp 08:18:23,234) :
> {code:java}
> 2018-09-02 08:18:23,232 DEBUG
> [rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-2]
> regionserver.MetricsTableSourceImpl(124): Creating new
> MetricsTableSourceImpl for table 'testtb-1535901500805'
> 2018-09-02 08:18:23,233 DEBUG
> [rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-2]
> regionserver.MetricsTableSourceImpl(137): registering metrics for testtb-
> 1535901500805
> 2018-09-02 08:18:23,234 INFO
> [rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-1]
> regionserver.HRegion(2822): Finished flush of dataSize ~2.29 KB/2343,
> heapSize ~5.16 KB/5280, currentSize=0 B/0 for
> fa403f6a4fb8dbc1a1c389744fce2d58 in 280ms, sequenceid=5, compaction
> requested=false
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52758,1535901497238)-snapshot-pool11-thread-1]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 6 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52758,1535901497238)-snapshot-pool11-thread-1,5,FailOnTimeoutGroup]
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-1]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 0 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-1,5,FailOnTimeoutGroup]
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52762,1535901497314)-snapshot-pool9-thread-1]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 6 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52762,1535901497314)-snapshot-pool9-thread-1,5,FailOnTimeoutGroup]
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52762,1535901497314)-snapshot-pool9-thread-2]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 6 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52762,1535901497314)-snapshot-pool9-thread-2,5,FailOnTimeoutGroup]
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52758,1535901497238)-snapshot-pool11-thread-2]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 5 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52758,1535901497238)-snapshot-pool11-thread-2,5,FailOnTimeoutGroup]
> 2018-09-02 08:18:23,234 DEBUG
> [rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-2]
> regionserver.MetricsTableAggregateSourceImpl(84): it took 6 ms to register
> testtb-1535901500805
> Thread[rs(hw13463.attlocal.net,52760,1535901497280)-snapshot-pool10-thread-2,5,FailOnTimeoutGroup]
> {code}
> This is a regression.
> When first region of the table is opened on region server, we can proactively
> register table metrics.
> This would avoid the penalty on first flushes for the table.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
