[
https://issues.apache.org/jira/browse/HBASE-17016?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15638102#comment-15638102
]
Mikhail Antonov commented on HBASE-17016:
-----------------------------------------
[~enis] not necessary close as won't fix; I meant to say that I think the unit
of request rate outliers is often a single hot region; the unit of latency
outlier is mostly (almost always?) RS - GC stall, WAL append failed due to
dfsclient hitting error, that kind of thing, that makes latency per region not
super useful imo. If we remove them and see any improvement in terms of "less
latency outliers since less Counters etc" - great, less remove them. If we can
bring them back, and they are cheap - sure, why not. If we find out that any
latency histograms are relatively expensive (in some visible form) I'd be in
favor or removing them, unless someone has the usecase when they are actually
useful.
> Reimplement per-region latency histogram metrics
> ------------------------------------------------
>
> Key: HBASE-17016
> URL: https://issues.apache.org/jira/browse/HBASE-17016
> Project: HBase
> Issue Type: Task
> Components: metrics
> Affects Versions: 2.0.0, 1.4.0
> Reporter: Andrew Purtell
> Fix For: 2.0.0, 1.4.0
>
>
> Follow up from HBASE-10656, where [~enis] says:
> {quote}
> the main problem is that we have A LOT of per-region metrics that are latency
> histograms. These latency histograms create many many Counter / LongAdder
> objects. We should get rid of per-region latencies and maybe look at reducing
> the per-region metric overhead.
> {quote}
> And [~ghelmling] gives us a good candidate to implement pre-region latency
> histograms [HdrHistogram|https://github.com/HdrHistogram/HdrHistogram].
> Let's consider removing the per-region latency histograms and reimplement
> using HdrHistogram.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
