[
https://issues.apache.org/jira/browse/HIVE-15093?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15637570#comment-15637570
]
Sergio Peña commented on HIVE-15093:
------------------------------------
I agree with [~yalovyyi] that we should leave file system operations to HDFS
and just use the rename() API, but currently Hadoop is leaving Hive behind on
performance because the rename() on S3 is slow due to serial copies.
[~yalovyyi] Do you happen to know if Hadoop 2.8.0 will make renames() perform
better on blobstore systems? When will Hadoop 2.8.0 be released to the public?
[~stakiar] I don't know if Hadoop will perform different when doing renames on
other blobstores. I think we would need to create a configuration flag specific
for each blobstore in case S3 is slow but Azure blobstore is fast. We would
want users to use Hive parallel copies with S3, but leave Azure use their own
rename() operation.
> S3-to-S3 Renames: Files should be moved individually rather than at a
> directory level
> -------------------------------------------------------------------------------------
>
> Key: HIVE-15093
> URL: https://issues.apache.org/jira/browse/HIVE-15093
> Project: Hive
> Issue Type: Sub-task
> Components: Hive
> Affects Versions: 2.1.0
> Reporter: Sahil Takiar
> Assignee: Sahil Takiar
> Attachments: HIVE-15093.1.patch, HIVE-15093.2.patch,
> HIVE-15093.3.patch, HIVE-15093.4.patch, HIVE-15093.5.patch,
> HIVE-15093.6.patch, HIVE-15093.7.patch
>
>
> Hive's MoveTask uses the Hive.moveFile method to move data within a
> distributed filesystem as well as blobstore filesystems.
> If the move is done within the same filesystem:
> 1: If the source path is a subdirectory of the destination path, files will
> be moved one by one using a threapool of workers
> 2: If the source path is not a subdirectory of the destination path, a single
> rename operation is used to move the entire directory
> The second option may not work well on blobstores such as S3. Renames are not
> metadata operations and require copying all the data. Client connectors to
> blobstores may not efficiently rename directories. Worst case, the connector
> will copy each file one by one, sequentially rather than using a threadpool
> of workers to copy the data (e.g. HADOOP-13600).
> Hive already has code to rename files using a threadpool of workers, but this
> only occurs in case number 1.
> This JIRA aims to modify the code so that case 1 is triggered when copying
> within a blobstore. The focus is on copies within a blobstore because
> needToCopy will return true if the src and target filesystems are different,
> in which case a different code path is triggered.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
