[
https://issues.apache.org/jira/browse/MAPREDUCE-7101?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16507097#comment-16507097
]
Wangda Tan commented on MAPREDUCE-7101:
---------------------------------------
[~ste...@apache.org] / [~ehiggs] / [~rohithsharma]
Thanks for your suggestions. To me there're two major options.
1. As mentioned by [~rohithsharma], Make the behavior (skip dir timestamp
check) to be configurable to avoid surprising users of HDFS-backed clusters.
This config can be marked as private/unstable so we change this in the future.
2. Considering cloud storages are not identical as mentioned, another approach
is to add a option to make the whole {{scanIntermediateDirectory}} becomes
pluggable folder-scan policy, and {{UserLogDir}} is private to the policy. We
can implement list by recursive=true for some FS and =false for others, and can
poll special files, etc.
I would not prefer to {{turn off the directory timestamp check}}, and I prefer
#1 over #2.
Thoughts? Can we can a conclusion for this so we can start fixing the problem
sooner if possible?
> Revisit behavior of JHS scan file behavior
> ------------------------------------------
>
> Key: MAPREDUCE-7101
> URL: https://issues.apache.org/jira/browse/MAPREDUCE-7101
> Project: Hadoop Map/Reduce
> Issue Type: Bug
> Reporter: Wangda Tan
> Priority: Critical
>
> Currently, the JHS scan directory if the modification of *directory* changed:
> {code}
> public synchronized void scanIfNeeded(FileStatus fs) {
> long newModTime = fs.getModificationTime();
> if (modTime != newModTime) {
> <... omitted some logics ...>
> // reset scanTime before scanning happens
> scanTime = System.currentTimeMillis();
> Path p = fs.getPath();
> try {
> scanIntermediateDirectory(p);
> {code}
> This logic relies on an assumption that, the directory's modification time
> will be updated if a file got placed under the directory.
> However, the semantic of directory's modification time is not consistent in
> different FS implementations. For example, MAPREDUCE-6680 fixed some issues
> of truncated modification time. And HADOOP-12837 mentioned on S3, the
> directory's modification time is always 0.
> I think we need to revisit behavior of this logic to make it to more robustly
> work on different file systems.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: mapreduce-issues-unsubscr...@hadoop.apache.org
For additional commands, e-mail: mapreduce-issues-h...@hadoop.apache.org
