[
https://issues.apache.org/jira/browse/HBASE-16025?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15406713#comment-15406713
]
Gary Helmling commented on HBASE-16025:
---------------------------------------
If you mean caching for only the lifetime of the request, I think that would
still lead to a lot of request amplification against meta. It would still mean
a single additional request to meta anytime any request to any region in the
cluster needs to be retried.
I think we need to aggressively cache table state for the lifetime of the
client and only clear it when we encounter certain exception types, ie.
NotServingRegionException.
> Cache table state to reduce load on META
> ----------------------------------------
>
> Key: HBASE-16025
> URL: https://issues.apache.org/jira/browse/HBASE-16025
> Project: HBase
> Issue Type: Improvement
> Components: Client
> Reporter: Gary Helmling
> Assignee: Gary Helmling
> Priority: Critical
> Fix For: 2.0.0
>
>
> HBASE-12035 moved keeping table enabled/disabled state from ZooKeeper into
> hbase:meta. When we retry operations on the client, we check table state in
> order to return a specific message if the table is disabled. This means that
> in master we will be going back to meta for every retry, even if a region's
> location has not changed. This is going to cause performance issues when a
> cluster is already loaded, ie. in cases where regionservers may be returning
> CallQueueTooBigException.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
