[
https://issues.apache.org/jira/browse/MAPREDUCE-4772?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Robert Joseph Evans updated MAPREDUCE-4772:
-------------------------------------------
Attachment: MR-4772-trunk.txt
MR-4772-0.23.txt
This patch changes the AM to restart a map task if 50% of the shuffling
reducers report errors for a given map task instead of 50% of the running
reducers.
It also changes how often a reduce reports fetch failures. If a
ConnectionException happens then it will report the error immediately and will
not wait. A ConnectionException indicates that there is no one listening on the
remote port. This is very different from a timeout where the port is overrun
and no one is able to get through.
It also adds in a maximum delay between fetch retries. In the original code
every time a fetch failure happened the reducer would add 30% to the delay. It
would also only report every 10th failure. This means that the first failure
would be reported after about 6 min, the second after 90 min and the third
after 20 hours. This is really bad when there is only one reducer because the
AM requires at least three reports for the map task to be restarted.
The default maximum delay is set to 1 min which would change the numbers to be
6 min, 15 min, and 25 min respectively. 25 min still seems very long to wait,
but is much better then 20 hours.
> Fetch failures can take way too long for a map to be restarted
> --------------------------------------------------------------
>
> Key: MAPREDUCE-4772
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-4772
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Components: mrv2
> Affects Versions: 0.23.4
> Reporter: Robert Joseph Evans
> Assignee: Robert Joseph Evans
> Priority: Critical
> Attachments: MR-4772-0.23.txt, MR-4772-trunk.txt
>
>
> In one particular case we saw a NM go down at just the right time, that most
> of the reducers got the output of the map tasks, but not all of them.
> The ones that failed to get the output reported to the AM rather quickly that
> they could not fetch from the NM, but because the other reducers were still
> running the AM would not relaunch the map task because there weren't more
> than 50% of the running reducers that had reported fetch failures. Then
> because of the exponential back-off for fetches on the reducers it took until
> 1 hour 45 min for the reduce tasks to hit another 10 fetch failures and
> report in again. At that point the other reducers had finished and the job
> relaunched the map task. If the reducers had still been running at 1:45 I
> have no idea how long it would have taken for each of the tasks to get to 30
> fetch failures.
> We need to trigger the map based off of percentage of reducers shuffling, not
> percentage of reducers running, we also need to have a maximum limit of the
> back off, so that we don't ever have the reducer waiting for days to try and
> fetch map output.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
