[
https://issues.apache.org/jira/browse/HBASE-25727?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17319044#comment-17319044
]
Xiaolin Ha commented on HBASE-25727:
------------------------------------
Yes. I think both read response size and write request size are needed. Using
MemStoreSizeCostFunction and StoreFileCostFunction is a little indirectly for
reflecting IO pressures, though it can reflect compaction and flush IO pressure
to some extent. But there are small regions with high read pressure, and also
small regions with too fast flush speed, which can not be noticed by balancer.
> Support Read/Write request size cost functions in StochasticLoadBalancer
> ------------------------------------------------------------------------
>
> Key: HBASE-25727
> URL: https://issues.apache.org/jira/browse/HBASE-25727
> Project: HBase
> Issue Type: Improvement
> Components: Balancer
> Affects Versions: 3.0.0-alpha-1, 2.0.0, 1.4.13
> Reporter: Xiaolin Ha
> Priority: Major
>
> Currently, StochasticLoadBalancer has several cost functions to reflect
> pressure of regions, they are Read/Write/CP RequestCostFunction,
> MemStoreSizeCostFunction, StoreFileCostFunction. Memstore and storefile cost
> functions are used to reflect the write io pressure. But when some regions
> has high read io pressures, balancer can not aware of it and then make some
> actions.
> Like request count cost functions in StochasticLoadBalancer, we could also
> support request size cost functions. Because in some circumstances, request
> size is the most important factor reflecting the pressure of RSes.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
