[
https://issues.apache.org/jira/browse/HBASE-29265?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hernan Gelaf-Romer updated HBASE-29265:
---------------------------------------
Description:
Somewhat related to https://issues.apache.org/jira/browse/HBASE-27487
In HBase 2.X, the AsyncRequestFutureImpl will clear the meta cache if it
receives a batch call that fails to execute. When AsyncRequestFutureImpl
receives an exception from a multi call, it will call receiveGlobalFailure and
force a meta cache clear, even if the exception received does not usually
trigger meta cache clears.
This can cause issues on HBase clusters that are currently overloaded. If a
server is overloaded, it can force clients to repeatedly clear meta cache,
causing a meta hotspot, which further slows down queries. I'm not entirely sure
why we force meta cache clears, but I think we should exempt exceptions that
indicate that a server is overloaded from clearing the meta cache.
I'm also open to only clearing the meta cache if the exception received is a
cache clearing exception, but I'd need to make sure this doesn't invalidate
some assumptions made when the code was first written.
was:
Somewhat related to https://issues.apache.org/jira/browse/HBASE-27487
In HBase 2.X, the AsyncRequestFutureImpl will clear the meta cache if it
receives
> Batch calls to overloaded cluster can cause meta hotspotting
> ------------------------------------------------------------
>
> Key: HBASE-29265
> URL: https://issues.apache.org/jira/browse/HBASE-29265
> Project: HBase
> Issue Type: Improvement
> Reporter: Hernan Gelaf-Romer
> Assignee: Hernan Gelaf-Romer
> Priority: Major
> Labels: pull-request-available
>
> Somewhat related to https://issues.apache.org/jira/browse/HBASE-27487
>
> In HBase 2.X, the AsyncRequestFutureImpl will clear the meta cache if it
> receives a batch call that fails to execute. When AsyncRequestFutureImpl
> receives an exception from a multi call, it will call receiveGlobalFailure
> and force a meta cache clear, even if the exception received does not usually
> trigger meta cache clears.
>
> This can cause issues on HBase clusters that are currently overloaded. If a
> server is overloaded, it can force clients to repeatedly clear meta cache,
> causing a meta hotspot, which further slows down queries. I'm not entirely
> sure why we force meta cache clears, but I think we should exempt exceptions
> that indicate that a server is overloaded from clearing the meta cache.
>
> I'm also open to only clearing the meta cache if the exception received is a
> cache clearing exception, but I'd need to make sure this doesn't invalidate
> some assumptions made when the code was first written.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
