[ https://issues.apache.org/jira/browse/HIVE-15121?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16180458#comment-16180458 ]
wangzhihao commented on HIVE-15121: ----------------------------------- Hi [~stakiar] Should not [this statement | https://github.com/apache/hive/blob/aa7edfefe20cede2c37d8c8b6c864c3b6039923f/ql/src/java/org/apache/hadoop/hive/ql/Context.java#L497 ] be {code:java} if (!isFinalJob && BlobStorageUtils.areOptimizationsEnabled(conf)) { {code} We should use HDFS staging dir when optimization is enabled and the MR is not final job. But current logic will call getMRTmpPath() code when areOptimizationsEnabled() return false. > Last MR job in Hive should be able to write to a different scratch directory > ---------------------------------------------------------------------------- > > Key: HIVE-15121 > URL: https://issues.apache.org/jira/browse/HIVE-15121 > Project: Hive > Issue Type: Sub-task > Components: Hive > Reporter: Sahil Takiar > Assignee: Sahil Takiar > Fix For: 2.3.0 > > Attachments: HIVE-15121.1.patch, HIVE-15121.2.patch, > HIVE-15121.3.patch, HIVE-15121.patch, HIVE-15121.WIP.1.patch, > HIVE-15121.WIP.2.patch, HIVE-15121.WIP.patch > > > Hive should be able to configure all intermediate MR jobs to write to HDFS, > but the final MR job to write to S3. > This will be useful for implementing parallel renames on S3. The idea is that > for a multi-job query, all intermediate MR jobs write to HDFS, and then the > final job writes to S3. Writing to HDFS should be faster than writing to S3, > so it makes more sense to write intermediate data to HDFS. > The advantage is that any copying of data that needs to be done from the > scratch directory to the final table directory can be done server-side, > within the blobstore. The MoveTask simply renames data from the scratch > directory to the final table location, which should translate to a > server-side COPY request. This way HiveServer2 doesn't have to actually copy > any data, it just tells the blobstore to do all the work. -- This message was sent by Atlassian JIRA (v6.4.14#64029)