[jira] [Commented] (SPARK-20560) Review Spark's handling of filesystems returning "localhost" in getFileBlockLocations

["Steve Loughran (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/SPARK-20560?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15993008#comment-15993008 ] 

Steve Loughran commented on SPARK-20560:
----------------------------------------

{{FileSystem.getFileBlockLocations(path)}} is only invoked from from {{HdfsUtils.getFileSegmentLocations}}, and used as a source of data for {{RDD.preferredLocations}}

I don't see anything explicit through the code that detects & reacts to the FS call returning localhost; I'll do some test downstream to see what surfaces against S3. Unless the scheduler has some explicit "localhost -> anywhere" map, it might make sense for HdfsUtils.getFileSegmentLocation to downgrade "localhost" to None, on the basis that in a cluster FS, the data clearly doesn't know where it is.



> Review Spark's handling of filesystems returning "localhost" in getFileBlockLocations
> -------------------------------------------------------------------------------------
>
>                 Key: SPARK-20560
>                 URL: https://issues.apache.org/jira/browse/SPARK-20560
>             Project: Spark
>          Issue Type: Bug
>          Components: Scheduler
>    Affects Versions: 2.1.0
>            Reporter: Steve Loughran
>            Priority: Minor
>
> Some filesystems (S3a, Azure WASB) return "localhost" as the response to {{FileSystem.getFileBlockLocations(path)}}. If this is then used as the preferred host when scheduling work, there's a risk that work will be queued on one host, rather than spread across the cluster.
> HIVE-14060 and TEZ-3291 have both seen it in their schedulers.
> I don't know if Spark does it, someone needs to look at the code, maybe write some tests



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscribe@spark.apache.org
For additional commands, e-mail: issues-help@spark.apache.org