[
https://issues.apache.org/jira/browse/HADOOP-16430?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16886461#comment-16886461
]
Steve Loughran commented on HADOOP-16430:
-----------------------------------------
Attached: screenshot of on-demand DDB load during a test run. IOPs goes up to
300 for a read, write to ~250 but for slightly longer. DDB shows 3ms per
operation.
Delete doesn't do batches; it just goes through file-by-file. If we were
confident that you did see a bulk write speedup (probably uses up fewer HTTPS
connections at the very least), it could be considered/justlfied
> S3AFilesystem.delete to incrementally update s3guard with deletions
> -------------------------------------------------------------------
>
> Key: HADOOP-16430
> URL: https://issues.apache.org/jira/browse/HADOOP-16430
> Project: Hadoop Common
> Issue Type: Sub-task
> Components: fs/s3
> Reporter: Steve Loughran
> Assignee: Steve Loughran
> Priority: Major
> Attachments: Screenshot 2019-07-16 at 22.08.31.png
>
>
> Currently S3AFilesystem.delete() only updates the delete at the end of a
> paged delete operation. This makes it slow when there are many thousands of
> files to delete ,and increases the window of vulnerability to failures
> Preferred
> * after every bulk DELETE call is issued to S3, queue the (async) delete of
> all entries in that post.
> * at the end of the delete, await the completion of these operations.
> * inside S3AFS, also do the delete across threads, so that different HTTPS
> connections can be used.
> This should maximise DDB throughput against tables which aren't IO limited.
> When executed against small IOP limited tables, the parallel DDB DELETE
> batches will trigger a lot of throttling events; we should make sure these
> aren't going to trigger failures
--
This message was sent by Atlassian JIRA
(v7.6.14#76016)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
