[
https://issues.apache.org/jira/browse/PIG-116?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12571880#action_12571880
]
Pi Song commented on PIG-116:
-----------------------------
In short term that will be fine but I would like to give a bit more thought
here that this approach seems relying on the trash feature of HDFS which is not
common among other file systems.
By looking at commercial DBMSs, what they do is:-
- Rely on a temp folder and temp files *on top of* file systems.
- User can customize location and size limit.
- The folder is clean-up every time the system restarts (comparable to when pig
restarts, not hadoop restarts so this can be more frequent)
- The old temp files are removed when the system running low on temp space
- Temp files where lifecycle is explicitly known may be marked for collection
at appropriate time (like the concept of GC)
This way looks good to me that
- temp file management will not rely on file system implementation
- good when running on a file system that has disk quota enabled (I saw a
discussion about disk quota on HDFS a few months ago).
- If JVM crashes during operation, it's still ok because when running pig next
time the temp folder will be clean up.
However this doesn't consider checkpointing/resumed run after crash.
My 2 cents.
> pig leaves temp files behind
> ----------------------------
>
> Key: PIG-116
> URL: https://issues.apache.org/jira/browse/PIG-116
> Project: Pig
> Issue Type: Bug
> Reporter: Olga Natkovich
> Assignee: Olga Natkovich
>
> Currently, pig creates temp dirs via call to FileLocalizer.getTemporaryPath.
> They are created on the client and are mainly used to store data between 2
> M-R jobs. Pig then attempts to clean them up in the client's shutdown hook.
> The problem with this approach is that, because there is now way to order the
> shutdown hooks, in some cases, the DFS is already closed when we try to
> delete the files in which case a substention amount of data can be left in
> DFS. I see this issue more frequently with hadoop 0.16 perhaps because I had
> to add an extra shutdown hook to handle hod disconnects.
> The short term, I would like to propose the approach below:
> (1) If trash is configured on the cluster, use trash location to create temp
> directory that will expire in 7 days. The hope is that most jobs don't run
> longer that 7 days. The user can specify a longer interval via a command line
> switch
> (2) If trash is not enabled on the cluster, the location that we use now will
> be used
> (3) In the shutdown hook, we will attempt to cleanup. If the attempt fails
> and trash is enabled, we let trash handle it; otherwise we provide the list
> of locations to the user to clean. (I realize that this is not ideal but
> could not figure out a better way.)
> Longer term, I am talking with hadoop team to have better temp file support:
> https://issues.apache.org/jira/browse/HADOOP-2815
> Comments? Suggestions?
--
This message is automatically generated by JIRA.
-
You can reply to this email to add a comment to the issue online.
