hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Nicolas Spiegelberg (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-3102) Enhance HRegionServerMetrics for Long-running Stats
Date Tue, 12 Oct 2010 19:03:47 GMT

    [ https://issues.apache.org/jira/browse/HBASE-3102?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12920308#action_12920308
] 

Nicolas Spiegelberg commented on HBASE-3102:
--------------------------------------------

'marathon' could be named something different. It's just a longer-running period and I didn't
want to use 'period2'. The intent of this metric is to be abstract, although RS compactions
& minMax are the only two uses currently. Good future candidates are HLogSplit and region
rebalancing times on the Master. I don't think you want custom reset times per-metric. Every
time you add a new metric, you'd have to make sure the time is properly synchronized. 

I was not going to make the modification in Hadoop to ensure we can support 0.20-append through
0.22.  My original intent was to subclass from MetricsTimeVaryingRate to avoid the pushMetric
hack, but the MetricContext implementation on 0.20-append used introspection and would require
a complete reimplementation of the class. Mostly cut-paste, but I thought that was more hacky.


Sent from my iPhone





> Enhance HRegionServerMetrics for Long-running Stats
> ---------------------------------------------------
>
>                 Key: HBASE-3102
>                 URL: https://issues.apache.org/jira/browse/HBASE-3102
>             Project: HBase
>          Issue Type: Improvement
>          Components: regionserver
>            Reporter: Nicolas Spiegelberg
>            Assignee: Nicolas Spiegelberg
>             Fix For: 0.90.0
>
>         Attachments: HBASE-3102_90.patch
>
>
> A useful metric to observe track is compaction count + duration. Since compactions are
long running and only happen infrequently, the avg/opcount stats are should be reset longer
than the default polling period (5 sec). In addition to 'hbase.period', we should allow a
different duration after which long-running metrics should expire. This would also fix our
existing metrics problem where min/max stats are never reset until the process is restarted/upgraded.

-- 
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.


Mime
View raw message