[
https://issues.apache.org/jira/browse/HBASE-8481?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15089209#comment-15089209
]
Matthew Byng-Maddick commented on HBASE-8481:
---------------------------------------------
In discussion with Lars, the metrics we should expose are:
* for a block cache miss, the size of the block read (this is already there, I
suspect, though maybe not quite as explicitly)
* for a block cache miss in reading a KV, the size of the KV read
Between the two, we can expose a trend of how much excess data we're reading
off the disk to help with making decisions on how to size blocks in the HFile.
> Add cost tracking to API calls for resource management
> ------------------------------------------------------
>
> Key: HBASE-8481
> URL: https://issues.apache.org/jira/browse/HBASE-8481
> Project: HBase
> Issue Type: New Feature
> Reporter: Lars George
>
> All calls that require resource should be tracked internally using a
> lightweight mechanism, such as a simple counter record. The idea is to be
> able to tell from calls how many resources they did require, for example a
> Get or Scan, how much true I/O was used. This should then be exposed to the
> Coprocessor context so that a plugin can make cost based decisions going
> forward with other calls. The record should include custom tags that in the
> preXYZ() calls can be used to for example tag a record with a user group and
> therefore be able to aggregate the statistics and accounting for groups of
> users.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
