[
https://issues.apache.org/jira/browse/HDFS-13641?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16500535#comment-16500535
]
Erik Krogen commented on HDFS-13641:
------------------------------------
I really support this change, thanks for taking it up [~csun]! A few minor
comments:
* The changes to make {{assertQuantileGauges}} more general are good. I think
leaving the default as "Latency" is probably a good idea, but let's document
that in the Javadoc.
* For the {{startTime}} within {{doWork()}}, I think it should be fetched
_after_ locking the namespace, to avoid lock queue delays appearing as part of
the load time.
Also, a more general comment. I think setting this up for percentiles is
useful, but percentiles are disabled by default for good reason: they are
expensive. I would really want to have these metrics on our production
clusters, where we do not enable these percentiles. I am thinking if, in
addition to the percentiles, we can add an average for each one (using, say,
{{MutableRate}}) that is enabled by default. You can use this to get a decent
idea of the value, and enable percentiles for more fine-grained information.
Thoughts?
> Add metrics for edit log tailing
> ---------------------------------
>
> Key: HDFS-13641
> URL: https://issues.apache.org/jira/browse/HDFS-13641
> Project: Hadoop HDFS
> Issue Type: Improvement
> Components: metrics
> Reporter: Chao Sun
> Assignee: Chao Sun
> Priority: Major
> Attachments: HDFS-13641-HDFS-12943.000.patch, HDFS-13641.000.patch
>
>
> We should add metrics for each iteration of edit log tailing, including 1) #
> of edits loaded, 2) time spent in select input edit stream, 3) time spent in
> loading the edits, 4) interval between the iterations.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
