[
https://issues.apache.org/jira/browse/MAPREDUCE-6478?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Junping Du updated MAPREDUCE-6478:
----------------------------------
Description:
In some of our test cases for MR on public cloud scenario, a very big MR job
with hundreds or thousands of reducers cannot finish successfully because of
Job Cleanup failures which is caused by different scale/performance impact for
File System on the cloud (like AzureFS) which replacing HDFS's deletion for
whole directory with REST API calls on deleting each sub-directories
recursively. Even it get successfully, that could take much longer time (hours)
which is not necessary and waste time/resources especially in public cloud
scenario.
In these scenarios, some failures of cleanupJob can be ignored or user choose
to skip cleanupJob() completely make more sense. This is because making whole
job finish successfully with side effect of wasting some user spaces is much
better as user's jobs are usually comes and goes in public cloud, so have
choices to tolerant some temporary files exists with get rid of big job re-run
(or saving job's running time) is quite effective in time/resource cost.
We should allow user to have this option (ignore failure or skip job cleanup
stage completely) especially when user know the cleanup failure is not due to
HDFS abnormal status but other FS' different performance trade-off.
was:
In some our test cases for MR on public cloud scenario, a very big MR job with
hundreds or thousands of reducers cannot finish successfully because of Job
Cleanup failures which is caused by different scale/performance impact for File
System on the cloud (like AzureFS) which replacing HDFS's deletion for whole
directory with REST API calls on deleting each sub-directories recursively.
That could take much longer time (hours) which is not necessary in public cloud
scenario.
Also, it also more easily to get failed to cleanup in these cases, so some
failures of cleanupJob can be ignored in this case. Making whole job finish
successfully with side effect of wasting some user spaces make more sense in
these cases as user's job is usually comes and goes in public cloud, so have a
trade off to tolerant some temporary files exists with get rid of big job
re-run is quite cost effective.
We should allow user to have this option (ignore failure or skip job cleanup
stage completely) especially when user know the cleanup failure is not due to
HDFS abnormal status but other FS' different performance trade-off.
> Add an option to skip cleanupJob stage or ignore cleanup failure during
> commitJob().
> ------------------------------------------------------------------------------------
>
> Key: MAPREDUCE-6478
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-6478
> Project: Hadoop Map/Reduce
> Issue Type: Improvement
> Reporter: Junping Du
> Assignee: Junping Du
>
> In some of our test cases for MR on public cloud scenario, a very big MR job
> with hundreds or thousands of reducers cannot finish successfully because of
> Job Cleanup failures which is caused by different scale/performance impact
> for File System on the cloud (like AzureFS) which replacing HDFS's deletion
> for whole directory with REST API calls on deleting each sub-directories
> recursively. Even it get successfully, that could take much longer time
> (hours) which is not necessary and waste time/resources especially in public
> cloud scenario.
> In these scenarios, some failures of cleanupJob can be ignored or user choose
> to skip cleanupJob() completely make more sense. This is because making whole
> job finish successfully with side effect of wasting some user spaces is much
> better as user's jobs are usually comes and goes in public cloud, so have
> choices to tolerant some temporary files exists with get rid of big job
> re-run (or saving job's running time) is quite effective in time/resource
> cost.
> We should allow user to have this option (ignore failure or skip job cleanup
> stage completely) especially when user know the cleanup failure is not due to
> HDFS abnormal status but other FS' different performance trade-off.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
