[
https://issues.apache.org/jira/browse/FLINK-1419?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14282393#comment-14282393
]
Fabian Hueske commented on FLINK-1419:
--------------------------------------
I think it is fine to make files in the DC immutable (read only). An operator
that wants to modify files, can create a local writable copy.
Files in the DC should only be copied once per TM and stay until the job is
finished, IMO.
The question is, who initiates the copy process. The first task that requires
the file? In that case, all other tasks need to recognize that the file is
copied by another task and wait until the copying is completed.
> DistributedCache doesn't preserver files for subsequent operations
> ------------------------------------------------------------------
>
> Key: FLINK-1419
> URL: https://issues.apache.org/jira/browse/FLINK-1419
> Project: Flink
> Issue Type: Bug
> Affects Versions: 0.8, 0.9
> Reporter: Chesnay Schepler
> Assignee: Chesnay Schepler
>
> When subsequent operations want to access the same files in the DC it
> frequently happens that the files are not created for the following operation.
> This is fairly odd, since the DC is supposed to either a) preserve files when
> another operation kicks in within a certain time window, or b) just recreate
> the deleted files. Both things don't happen.
> Increasing the time window had no effect.
> I'd like to use this issue as a starting point for a more general discussion
> about the DistributedCache.
> Currently:
> 1. all files reside in a common job-specific directory
> 2. are deleted during the job.
>
> One thing that was brought up about Trait 1 is that it basically forbids
> modification of the files, concurrent access and all. Personally I'm not sure
> if this a problem. Changing it to a task-specific place solved the issue
> though.
> I'm more concerned about Trait #2. Besides the mentioned issue, the deletion
> is realized with the scheduler, which adds a lot of complexity to the current
> code. (It really is a pain to work on...)
> If we moved the deletion to the end of the job it could be done as a clean-up
> step in the TaskManager, With this we could reduce the DC to a
> cacheFile(String source) method, the delete method in the TM, and throw out
> everything else.
> Also, the current implementation implies that big files may be copied
> multiple times. This may be undesired, depending on how big the files are.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
