hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Elliott Clark (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-15518) Add Per-Table metrics back
Date Tue, 22 Mar 2016 18:55:26 GMT

    [ https://issues.apache.org/jira/browse/HBASE-15518?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15207035#comment-15207035
] 

Elliott Clark commented on HBASE-15518:
---------------------------------------

They were removed explicitly because they give very skewed numbers based on how many regions
are on each region server. If one regionserver has 10 regions of a table and another has 5
then the numbers become almost nonsensical. You then need to normalize them per number of
regions and you're back at a per region metrics. If a user is interested in per table across
the whole cluster then aggregating the region metrics is the best way. If they are looking
for a certain region server then not aggregating is the correct thing.

On top of all that the other reason they were removed is because it caused a whole lot of
contention for the debatable gains. Per region metrics are much less contented so adding them
is lower cost. Per server metrics are contended so we should only add the minimum set of metrics
needed there.

> Add Per-Table metrics back
> --------------------------
>
>                 Key: HBASE-15518
>                 URL: https://issues.apache.org/jira/browse/HBASE-15518
>             Project: HBase
>          Issue Type: Sub-task
>            Reporter: Enis Soztutar
>            Assignee: Alicia Ying Shu
>             Fix For: 2.0.0, 1.4.0
>
>
> We used to have per-table metrics, but it was removed in some restructuring. We have
per-region metrics, and per-regionserver metrics, but nothing in between. 
> For majority of users, per-region is too granular, they are mostly interested in table
level aggregates. This is especially useful in multi-tenant cases where a table's disk usage,
number of requests, etc can be made much more visible. 
> In this jira, we'll add the basic infrastructure to add a single (or a few) per-table
metrics. Than we can improve on that by adding remaining metrics from the region server level.

> The plan is to NOT aggregate per-table metrics at master for now. Just aggregation of
per-region metrics at the per-table level for every regionserver. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message