hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Heng Chen (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-14674) Rpc handler / task monitoring seems to be broken after 0.98
Date Fri, 23 Oct 2015 00:59:27 GMT

    [ https://issues.apache.org/jira/browse/HBASE-14674?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14970219#comment-14970219
] 

Heng Chen commented on HBASE-14674:
-----------------------------------

{quote}
This can be done at the constructor of the handler threads I think, but we do need to extract
out the inlined Thread into a class probably.
+          MonitoredRPCHandler status = getStatus();
{quote}

handler thread was created in {{RpcExecutor.start}} in MAIN thread.  
If  we move it into constructor, it will not list idle handler thread....

How do you think about it? [~enis]


> Rpc handler / task monitoring seems to be broken after 0.98
> -----------------------------------------------------------
>
>                 Key: HBASE-14674
>                 URL: https://issues.apache.org/jira/browse/HBASE-14674
>             Project: HBase
>          Issue Type: Bug
>            Reporter: Enis Soztutar
>             Fix For: 1.3.0, 1.2.1, 1.0.3, 1.1.4, 0.98.17
>
>         Attachments: HBASE-14674.patch
>
>
> In 0.96, we have the RPC handlers listed as tasks and show them in the web UI as well:

> {code}
> Tasks:
> ===========================================================
> Task: RpcServer.handler=0,port=64231
> Status: WAITING:Waiting for a call
> Running for 932s
> Task: RpcServer.handler=1,port=64231
> Status: WAITING:Waiting for a call
> Running for 932s
> Task: RpcServer.handler=2,port=64231
> Status: WAITING:Waiting for a call
> Running for 932s
> {code}
> After pluggable RPC scheduler, the way the tasks work for the handlers got changed. We
no longer list idle RPC handlers in the tasks, but we register them dynamically to {{TaskMonitor}}
through {{CallRunner}}. However, the IPC readers are still registered the old way (meaning
that idle readers are listed as tasks, but not idle handlers). 
> From the javadoc of {{MonitoredRPCHandlerImpl}}, it seems that we are NOT optimizing
the allocation for the MonitoredTask anymore, but instead allocate one for every RPC call
breaking the pattern (See CallRunner.getStatus()). 
> {code}
> /**
>  * A MonitoredTask implementation designed for use with RPC Handlers 
>  * handling frequent, short duration tasks. String concatenations and object 
>  * allocations are avoided in methods that will be hit by every RPC call.
>  */
> @InterfaceAudience.Private
> public class MonitoredRPCHandlerImpl extends MonitoredTaskImpl
> {code}
> There is also one more side affect that, since the CallRunner is a per-RPC object and
created in the RPC listener thread, the created task ends up having a name "listener" although
the actual processing happens in a handler thread. This is obviously very confusing during
debugging. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message