[
https://issues.apache.org/jira/browse/HIVE-22928?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17065109#comment-17065109
]
Denys Kuzmenko commented on HIVE-22928:
---------------------------------------
[~poeppt], just upload patch file with incremented version #, like
HIVE-22928.2.patch
> Allow hive.exec.stagingdir to be a fully qualified directory name
> -----------------------------------------------------------------
>
> Key: HIVE-22928
> URL: https://issues.apache.org/jira/browse/HIVE-22928
> Project: Hive
> Issue Type: Improvement
> Components: Configuration, Hive
> Affects Versions: 3.1.2
> Reporter: Thomas Poepping
> Assignee: Thomas Poepping
> Priority: Minor
> Attachments: HIVE-22928.patch
>
>
> Currently, {{hive.exec.stagingdir}} can only be set as a relative directory
> name that, for operations like {{insert}} or {{insert overwrite}}, will be
> placed either under the table directory or the partition directory.
> For cases where an HDFS cluster is small but the data being inserted is very
> large (greater than the capacity of the HDFS cluster, as mentioned in a
> comment by [~ashutoshc] on [HIVE-14270]), the client may want to set their
> staging directory to be an explicit blobstore path (or any filesystem path),
> rather than relying on Hive to intelligently build the blobstore path based
> on an interpretation of the job. We may lose locality guarantees, but because
> renames are just as expensive on blobstores no matter what the prefix is,
> this isn't considered a terribly large loss (assuming only blobstore
> customers use this functionality).
> Note that {{hive.blobstore.use.blobstore.as.scratchdir}} doesn't actually
> suffice in this case, as the stagingdir is not the same.
> This commit enables Hive customers to set an absolute location for all
> staging directories. For instances where the configured stagingdir scheme is
> not the same as the scheme for the table location, the default stagingdir
> configuration is used. This avoids a cross-filesystem rename, which is
> impossible anyway.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
