[
https://issues.apache.org/jira/browse/HBASE-15160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15254361#comment-15254361
]
Enis Soztutar commented on HBASE-15160:
---------------------------------------
bq. 1. From the latest patch, we're adding keys for read/write count, could you
clarify the reason for this when we already have the num_ops couting in
histogram?
What I have noticed is that, the num_ops coming from the histograms are reset
everytime the histograms are reset. We are relying on these counts at the
regionserver level as well (like get_numOps, etc), but I think it is wrong and
very hard to interpret because
bq. 2. Since the read op happens inside a lock in HFileReaderImpl#getMetaBlock,
cost of update histogram hurts, and confirmed to be the root cause of the ~3%
performance regression in my test
Thanks [~carp84] for the perf test. I did not put up this patch against YCSB
yet. Did you try with block cache disabled? The metrics will only get updated
when an actual read happens obviously, so I was thinking of doing the test with
block cache turned off. Let me try your suggestion.
I originally changed the location for the histogram update to be inside the
HFileBlock.readAtOffset() rather than at the HFileReader level since even if
argument {{pread=false}}, we might end up doing a {{pread}} if we cannot get
the lock. Otherwise reporting for pread vs read will be slightly wrong if there
is contention for the input stream lock.
> Put back HFile's HDFS op latency sampling code and add metrics for monitoring
> -----------------------------------------------------------------------------
>
> Key: HBASE-15160
> URL: https://issues.apache.org/jira/browse/HBASE-15160
> Project: HBase
> Issue Type: Sub-task
> Affects Versions: 2.0.0, 1.1.2
> Reporter: Yu Li
> Assignee: Yu Li
> Attachments: HBASE-15160.patch, HBASE-15160_v2.patch,
> HBASE-15160_v3.patch, hbase-15160_v4.patch, hbase-15160_v5.patch
>
>
> In HBASE-11586 all HDFS op latency sampling code, including fsReadLatency,
> fsPreadLatency and fsWriteLatency, have been removed. There was some
> discussion about putting them back in a new JIRA but never happened.
> According to our experience, these metrics are useful to judge whether issue
> lies on HDFS when slow request occurs, so we propose to put them back in this
> JIRA, and add the metrics for monitoring as well.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
