[
https://issues.apache.org/jira/browse/HADOOP-490?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#action_12475496
]
Andrzej Bialecki commented on HADOOP-490:
------------------------------------------
I'd like to call for re-evaluation of this issue. With the introduction of
TaskTrackerAction it seems to me that signals could be accommodated easier than
before, simply by sending yet another type of TaskTrackerAction. The original
reasons for this issue are still valid - the need to pass bits of information
to all tasks in a job.
The message queue approach mentioned before has been tested in practice, and
found useful for small-scale clusters and infrequent (control-type) messages.
However, it's not scalable due to the heavy load it puts on the namenode.
> Add ability to send "signals" to jobs and tasks
> -----------------------------------------------
>
> Key: HADOOP-490
> URL: https://issues.apache.org/jira/browse/HADOOP-490
> Project: Hadoop
> Issue Type: New Feature
> Components: mapred
> Affects Versions: 0.6.0
> Reporter: Andrzej Bialecki
> Assigned To: Owen O'Malley
>
> In some cases it would be useful to be able to "signal" a job and its tasks
> about some external condition, or to broadcast a specific message to all
> tasks in a job. Currently we can only send a single pseudo-signal, that is
> to kill a job.
> Example 1: some jobs may be gracefully terminated even if they didn't
> complete all their work, e.g. Fetcher in Nutch may be running for a very long
> time if it blocks on relatively few sites left over from the fetchlist. In
> such case it would be very useful to send it a message requesting that it
> discards the rest of its input and gracefully completes its map tasks.
> Example 2: available bandwidth for fetching may be different at different
> times of day, e.g. daytime vs. nighttime, or total external link usage by
> other applications. Fetcher jobs often run for several hours. It would be
> good to be able to send a "signal" to the Fetcher to throttle or un-throttle
> its bandwidth usage depending on external conditions.
> Job implementations could react to these messages either by implementing a
> method, or by registering a listener, whichever seems more natural.
> I'm not quite sure how to go about implementing it, I guess this would have
> to be a part of TaskUmbilicalProtocol but my knowledge here is a bit fuzzy
> ... ;) Comments are welcome.
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
