[jira] [Commented] (HBASE-27531) AsyncRequestFutureImpl unnecessarily clears meta cache for full server

["Bryan Beaudreault (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HBASE-27531?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17646667#comment-17646667 ] 

Bryan Beaudreault commented on HBASE-27531:
-------------------------------------------

Digging further into the git history, actually the tableName == null check was included in the original commit. But then it was removed in HBASE-21775. I'm going to do a bit more digging here when I have time. Ideally we can keep BufferedMutator working without having to always clear the cache for the entire server for every small per-request failure.

> AsyncRequestFutureImpl unnecessarily clears meta cache for full server
> ----------------------------------------------------------------------
>
>                 Key: HBASE-27531
>                 URL: https://issues.apache.org/jira/browse/HBASE-27531
>             Project: HBase
>          Issue Type: Improvement
>            Reporter: Bryan Beaudreault
>            Priority: Major
>
> In AsyncRequestFutureImpl.receiveGlobalFailure, the meta cache is cleared twice. First the entire cache is cleared for the serverName, then each individual region's cache is further cleared. See [here.|https://github.com/apache/hbase/blob/branch-2/hbase-client/src/main/java/org/apache/hadoop/hbase/client/AsyncRequestFutureImpl.java#L766-L780]
> This behavior was introduced in HBASE-19900, which was a Jira more related to return value corruption rather than meta cache. I believe the change was an accident in the refactoring. Take a look at the commit [here.|https://github.com/apache/hbase/commit/e2636c81f7700a34368d73079375dbce6e180547#diff-7c58ffd83c150488599591ed5a3a068599646ebdbbbfdcd2233386e5472cca35L709]
> In the commit, the cleanServerCache call should have been wrapped in tableName == null. Then the updateCachedLocations call should have been wrapped in tableName != null. So we should only go down one path or another based on the existence of a tableName.
> Full server cache clears are extremely expensive in a high concurrency client. This is especially true in AsyncProcess, where a single request might hit many servers. We should reinstate the tableName checks so that we only do one or the other.



--
This message was sent by Atlassian Jira
(v8.20.10#820010)