[
https://issues.apache.org/jira/browse/FLINK-6380?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15996474#comment-15996474
]
Stephan Ewen commented on FLINK-6380:
-------------------------------------
I think we should refrain from using the "delete parent if empty" pattern in
any place. Always risky.
Rather, we should have a proper life cycle when these things are initialized
and cleaned up (for example when the HaServices are created / destroyed).
> BlobService concurrency issues between delete and put/get methods
> -----------------------------------------------------------------
>
> Key: FLINK-6380
> URL: https://issues.apache.org/jira/browse/FLINK-6380
> Project: Flink
> Issue Type: Bug
> Components: Network
> Affects Versions: 1.3.0
> Reporter: Nico Kruber
>
> {{BlobCache#deleteAll(JobID)}} deletes the job directory which is only
> created at the start of {{BlobCache#getURL(BlobKey)}} which then relies on
> the directory being present.
> This is not restricted to the {{BlobCache}}, though, but also affects the
> {{BlobServer}} in two ways:
> 1) its own local storage and
> 2) its backing {{BlobStore}}
> For the latter, i.e. in {{FileSystemBlobStore}}, there is no guarantee that a
> directory will not be deleted concurrently (from a {{delete}} method) between
> its creation and writing a file (in a {{get}} method):
> * the {{delete}} method for name-addressable blobs always deletes the
> job-specific storage directory if there is no further blob for this job
> * the content-addressable blobs do that similarly but are shared among jobs
> and thus only delete directories if there is no other blob.
> Since name-addressable blobs have not been used so far and the latter case
> typically does not occur concurrently with get/put requests, this has not
> been a problem so far but is more relevant after applying FLINK-6046.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
