[
https://issues.apache.org/jira/browse/HBASE-11683?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14102310#comment-14102310
]
Jonathan Hsieh commented on HBASE-11683:
----------------------------------------
{quote}
I'm thinking how to implement the #2 mob reads, is it okay to record how many
times the scanner read from the mob files? I don't see HBase has metrics in the
normal scanner, is it necessary for the mob read? Please advise. Thanks.
{quote}
I'm thinking about this from the point of view of someone trying to decide if
they should use the mob or an operator verifying that the mobs are working.
Flushes should cover the write side metrics. Ideally i'd want to know how much
IO i'm saving or would save by using the mob feature and this helps me
understand that. We'd probably want some compaction related mob counts as well.
(# cells converted to mob, # converted from mob).
However, I really do care about the reads side as well. It would be great
actually if we got general sizes statistics for the cells when reading and
stats on the mob caches as well. There are two places I'm thinking the data
could be collected:
* Adding a counter every time the mob dereferences cell (specific to mob)
* Adding cell size count buckets that the server tracks when a Result is sent
from a get/scan.
> Metrics for MOB
> ---------------
>
> Key: HBASE-11683
> URL: https://issues.apache.org/jira/browse/HBASE-11683
> Project: HBase
> Issue Type: Sub-task
> Components: regionserver, Scanners
> Affects Versions: 2.0.0
> Reporter: Jonathan Hsieh
> Assignee: Jingcheng Du
> Attachments: HBASE-11683.diff
>
>
> We need to make sure to capture metrics about mobs.
> Some basic ones include:
> # of mob writes
> # of mob reads
> # avg size of mob (?)
> # mob files
> # of mob compactions / sweeps
--
This message was sent by Atlassian JIRA
(v6.2#6252)
