[jira] [Commented] (HADOOP-10300) Allowed deferred sending of call responses

["Daryn Sharp (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HADOOP-10300?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14549124#comment-14549124 ] 

Daryn Sharp commented on HADOOP-10300:
--------------------------------------

Modifying the tests isn't very feasible.  From what I recall, there are tests which use {{ExitUtil}} to convert fatal shutdowns into a {{RuntimeException}}.  These exceptions rip up through the stack and don't set {{rpcResponse}}.  The connection was left in limbo but I needed the connection closed, as it was prior to deferred responses, for the tests to pass.

Also in this "not supposed to happen" case, it's far better to abort the connection instead of leaving the client dangling to wait for a call response that will never be sent.

> Allowed deferred sending of call responses
> ------------------------------------------
>
>                 Key: HADOOP-10300
>                 URL: https://issues.apache.org/jira/browse/HADOOP-10300
>             Project: Hadoop Common
>          Issue Type: Sub-task
>          Components: ipc
>    Affects Versions: 2.0.0-alpha, 3.0.0
>            Reporter: Daryn Sharp
>            Assignee: Daryn Sharp
>              Labels: BB2015-05-TBR
>         Attachments: HADOOP-10300.patch, HADOOP-10300.patch
>
>
> RPC handlers currently do not return until the RPC call completes and response is sent, or a partially sent response has been queued for the responder.  It would be useful for a proxy method to notify the handler to not yet the send the call's response.
> An potential use case is a namespace handler in the NN might want to return before the edit log is synced so it can service more requests and allow increased batching of edits per sync.  Background syncing could later trigger the sending of the call response to the client.



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