[
https://issues.apache.org/jira/browse/HIVE-15546?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15829006#comment-15829006
]
Sahil Takiar commented on HIVE-15546:
-------------------------------------
According to the code, dummy files are added to empty partitions so that the
operator pipeline does not need to be special cased to handle such scenarios.
As far as I can tell, this has been the behavior for a while:
{quote}
If any input path points to an empty table or partition a dummy file in the
scratch dir is instead created and added to the list. This is needed to avoid
special casing the operator pipeline for these cases.
{quote}
I originally tried to remove the dummy file logic, but that caused a bunch of
test failures (ref the first Hive QA run). So I decided to make the logic
multi-threaded instead.
Sergey kindly pointed out some code in NullScanFileSystem and
ZeroRowsInputFormat that could potentially get around this, but I was going to
leave that investigation for a separate patch. Even when using the
NullScanFileSystem and ZeroRowsInputFormat, a listStatus still needs to be done
to see if the partition is empty or not, which can be costly on blobstores like
S3.
> Optimize Utilities.getInputPaths() so each listStatus of a partition is done
> in parallel
> ----------------------------------------------------------------------------------------
>
> Key: HIVE-15546
> URL: https://issues.apache.org/jira/browse/HIVE-15546
> Project: Hive
> Issue Type: Sub-task
> Components: Hive
> Reporter: Sahil Takiar
> Assignee: Sahil Takiar
> Attachments: HIVE-15546.1.patch, HIVE-15546.2.patch,
> HIVE-15546.3.patch, HIVE-15546.4.patch, HIVE-15546.5.patch
>
>
> When running on blobstores (like S3) where metadata operations (like
> listStatus) are costly, Utilities.getInputPaths() can add significant
> overhead when setting up the input paths for an MR / Spark / Tez job.
> The method performs a listStatus on all input paths in order to check if the
> path is empty. If the path is empty, a dummy file is created for the given
> partition. This is all done sequentially. This can be really slow when there
> are a lot of empty partitions. Even when all partitions have input data, this
> can take a long time.
> We should either:
> (1) Just remove the logic to check if each input path is empty, and handle
> any edge cases accordingly.
> (2) Multi-thread the listStatus calls
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
