[
https://issues.apache.org/jira/browse/HBASE-15134?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16100763#comment-16100763
]
Andrew Purtell commented on HBASE-15134:
----------------------------------------
I think per region flush and compaction metrics are useful. They indicate IO
load on the region that might be hard to detect through other means. Op
counters don't tell a story about the store level consequences of client
activity like flushes and compactions do. FWIW
If we're going to commit this, I'd like to see it brought back at least as far
as branch-1.4. Thanks.
> Add visibility into Flush and Compaction queues
> -----------------------------------------------
>
> Key: HBASE-15134
> URL: https://issues.apache.org/jira/browse/HBASE-15134
> Project: HBase
> Issue Type: New Feature
> Components: Compaction, metrics, regionserver
> Reporter: Elliott Clark
> Assignee: Abhishek Singh Chouhan
> Fix For: 2.0.0
>
> Attachments: HBASE-15134.master.001.patch,
> HBASE-15134.master.002.patch, HBASE-15134.patch, HBASE-15134.patch
>
>
> On busy spurts we can see regionservers start to see large queues for
> compaction. It's really hard to tell if the server is queueing a lot of
> compactions for the same region, lots of compactions for lots of regions, or
> just falling behind.
> For flushes much the same. There can be flushes in queue that aren't being
> run because of delayed flushes. There's no way to know from the metrics how
> many flushes are for each region, how many are delayed. Etc.
> We should add either more metrics around this ( num per region, max per
> region, min per region ) or add on a UI page that has the list of compactions
> and flushes.
> Or both.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
