[
https://issues.apache.org/jira/browse/HBASE-15742?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15265365#comment-15265365
]
Phil Yang commented on HBASE-15742:
-----------------------------------
This issue doesn't need unit test, I think. I test it in our staging cluster,
we have a metric service which will get metrics from each server by JMX
periodically (several seconds), and these servers have a long minor GC STW
time, after this patch the minor GC time decrease much.
> Reduce allocation of objects in metrics
> ---------------------------------------
>
> Key: HBASE-15742
> URL: https://issues.apache.org/jira/browse/HBASE-15742
> Project: HBase
> Issue Type: Bug
> Affects Versions: 1.3.0, 1.2.1, 1.0.3, 1.1.4, 0.98.19
> Reporter: Phil Yang
> Attachments: HBASE-15742-v1.patch
>
>
> We use JMX and o.a.h.metrics2 to do some metrics on regions, tables, region
> servers and cluster. We use MetricsInfo to show the information of metrics,
> and we use Interns to cache MetricsInfo objects because it won't be changed.
> However, in Interns there are some static values to limit the max cached
> objects. We can only cache 2010 metrics, but we have dozens of metrics for
> one region and we have some RS-level metrics in each RS and all metrics for
> all regions will be saved in master. So each server will have thousands of
> metrics, and we can not cache most of them. When we collect metrics by JMX,
> we will create many objects which can be avoid. It increases the pressure of
> GC and JMX has some caching logic so the objects can not be removed
> immediately which increases the pressure more.
> Interns is in Hadoop project, and I think the implementation is not suitable
> for HBase. Because we can not know how many MetricsInfo we have, it depends
> on the number of regions. And we can not set it unlimited because we should
> remove the objects whose region is split, moved, or dropped. I think we can
> use Guava's cache with expireAfterAccess which is very simple and convenient.
> So we can add a new Interns class in HBase project first, and put it to
> upstream later.
> Moreover, in MutableHistogram#snapshot we create same Strings each time, we
> can create them only in the first time.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
