[
https://issues.apache.org/jira/browse/HBASE-27532?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Bryan Beaudreault updated HBASE-27532:
--------------------------------------
Description:
Initially in HBASE-14978, but more recently improved in HBASE-27558 and
HBASE-27570, we now track the amount of block bytes scanned by read requests.
Random read use-cases (which might read 1 small row from a 64kb block) and
heavily filtered scans (which might scan many blocks in order to return a
needle in a haystack row) can unknowingly cause RegionServers to do more work
than a user expects. Now that we have block IO accounting, we should expose
this to the user in the form of jmx metrics and ScanMetrics.
was:Way back in HBASE-14978 we started tracking the block size retained for
responses in addition to the actual size of the cells returned. It's currently
used for evaluating quotas, but this is extremely useful information for
monitoring the overall cost of queries being executed. We should report this
to JMX, like we do with call.getResponseCellSize(). If possible I'd love to see
it added to region/table metrics as well.
> Add block bytes scanned metrics
> -------------------------------
>
> Key: HBASE-27532
> URL: https://issues.apache.org/jira/browse/HBASE-27532
> Project: HBase
> Issue Type: Improvement
> Reporter: Bryan Beaudreault
> Priority: Major
>
> Initially in HBASE-14978, but more recently improved in HBASE-27558 and
> HBASE-27570, we now track the amount of block bytes scanned by read requests.
> Random read use-cases (which might read 1 small row from a 64kb block) and
> heavily filtered scans (which might scan many blocks in order to return a
> needle in a haystack row) can unknowingly cause RegionServers to do more work
> than a user expects. Now that we have block IO accounting, we should expose
> this to the user in the form of jmx metrics and ScanMetrics.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
