[
https://issues.apache.org/jira/browse/SPARK-17159?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15428677#comment-15428677
]
Steve Loughran commented on SPARK-17159:
----------------------------------------
# the most minimal change is to get rid of that directoryFilter and just do it
as a filter on the returned list of generated FileStatus entries. Against s3a,
That will eliminate 1-4 HTTP calls per path which matches the pattern. It will
mean a larger array of FileStatus entries are returned, but otherwise has
# it *may* be possible to go further and have the glob also pickup all files
underneath. That may or may not provide a speedup.
# The listStatus() operatin can be speeded up by having its file filter
executed after the run (and so there being an existing filestatus entry: no
need to go near the FS to get the modification time
> Improve FileInputDStream.findNewFiles list performance
> ------------------------------------------------------
>
> Key: SPARK-17159
> URL: https://issues.apache.org/jira/browse/SPARK-17159
> Project: Spark
> Issue Type: Improvement
> Components: Streaming
> Affects Versions: 2.0.0
> Environment: spark against object stores
> Reporter: Steve Loughran
> Priority: Minor
>
> {{FileInputDStream.findNewFiles()}} is doing a globStatus with a fitler that
> calls getFileStatus() on every file, takes the output and does listStatus()
> on the output.
> This going to suffer on object stores, as dir listing and getFileStatus calls
> are so expensive. It's clear this is a problem, as the method has code to
> detect timeouts in the window and warn of problems.
> It should be possible to make this faster
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
