[
https://issues.apache.org/jira/browse/HDFS-16696?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17572023#comment-17572023
]
Erik Krogen commented on HDFS-16696:
------------------------------------
Generally +1 from me as long as we can get it done without introducing too much
additional complexity, and keep it off-by-default. If we assume there are many
clients using always-msync mode (including for Design B laid out in
HDFS-13522), then latency of msync ops becomes very critical. Even without
always-msync mode, latency of msync can have an impact on performance, so it's
still potentially useful.
> NameNode supports a new MsyncRPCServer to handle msync() RPC separately
> -----------------------------------------------------------------------
>
> Key: HDFS-16696
> URL: https://issues.apache.org/jira/browse/HDFS-16696
> Project: Hadoop HDFS
> Issue Type: Improvement
> Reporter: ZanderXu
> Assignee: ZanderXu
> Priority: Major
>
> HDFS-12943 introduced Consistent Reads from Standby Node. It use msync
> mechanism to guarantee the consistency. So the latency of msycn() rpc is
> very important, especially for some end users who need call msync() rpc every
> time.
> Unfortunately, NameNode handle msync() RPCs same with other RPCs, also need
> enqueue, wait, handled. So the msync() will be blocked by other RPCs, such as
> setQuota, rename, delete, etc.
> So we need a new mechanism to guarantee the latency of the msync() RPC.
> Such as:
> * We can supports a new MsyncRPCServer in NameNode to separately msync() RPC.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
