[
https://issues.apache.org/jira/browse/HBASE-11876?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14118908#comment-14118908
]
Andrew Purtell commented on HBASE-11876:
----------------------------------------
Attached patch removes metrics update from HRegion.RegionScannerImpl#nextRaw,
keeps track of the equivalent metric in RSRpcServices#scan, and updates the
metric once per batch. When limiting the result set by size we use
KeyValue#heapSize, which also accounts for object heap overheads, but when
calculating scan metrics we use KeyValue#getLength, which is the on wire size
of the key value(s). I've maintained this distinction.
> RegionScanner.nextRaw(...) should not update metrics
> ----------------------------------------------------
>
> Key: HBASE-11876
> URL: https://issues.apache.org/jira/browse/HBASE-11876
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.98.6
> Reporter: Lars Hofhansl
> Assignee: Andrew Purtell
> Fix For: 0.99.0, 2.0.0, 0.98.6
>
> Attachments: HBASE-11876.patch
>
>
> I added the RegionScanner.nextRaw(...) to allow "smart" client to avoid some
> of the default work that HBase is doing, such as {start|stop}RegionOperation
> and synchronized(scanner) for each row.
> Metrics should follow the same approach. Collecting them per row is expensive
> and a caller should have the option to collect those later or to avoid
> collecting them completely.
> We can also save some cycles in RSRcpServices.scan(...) if we updated the
> metric only once/batch instead of each row.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
