[
https://issues.apache.org/jira/browse/HIVE-13232?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Owen O'Malley updated HIVE-13232:
---------------------------------
Description:
In Hive 0.11, when ORC's OutStream's were flushed they dropped all of the their
buffers. In the patch for HIVE-4243, we inadvertently changed that behavior so
that one of the buffers is held on to. For queries with a lot of writers and
thus under significant memory pressure this can have a significant impact on
the memory usage.
Note that "hive.optimize.sort.dynamic.partition" avoids this problem by sorting
on the dynamic partition key and thus only a single ORC writer is open at once.
This will use memory more effectively and avoid creating ORC files with very
small stripes, which will produce better downstream performance.
was:
In Hive 0.11, when ORC's OutStream's were flushed they dropped all of the their
buffers. In the patch for HIVE-4342, we inadvertently changed that behavior so
that one of the buffers is held on to. For queries with a lot of writers and
thus under significant memory pressure this can have a significant impact on
the memory usage.
Note that "hive.optimize.sort.dynamic.partition" avoids this problem by sorting
on the dynamic partition key and thus only a single ORC writer is open at once.
This will use memory more effectively and avoid creating ORC files with very
small stripes, which will produce better downstream performance.
> Aggressively drop compression buffers in ORC OutStreams
> -------------------------------------------------------
>
> Key: HIVE-13232
> URL: https://issues.apache.org/jira/browse/HIVE-13232
> Project: Hive
> Issue Type: Bug
> Components: ORC
> Reporter: Owen O'Malley
> Assignee: Owen O'Malley
>
> In Hive 0.11, when ORC's OutStream's were flushed they dropped all of the
> their buffers. In the patch for HIVE-4243, we inadvertently changed that
> behavior so that one of the buffers is held on to. For queries with a lot of
> writers and thus under significant memory pressure this can have a
> significant impact on the memory usage.
> Note that "hive.optimize.sort.dynamic.partition" avoids this problem by
> sorting on the dynamic partition key and thus only a single ORC writer is
> open at once. This will use memory more effectively and avoid creating ORC
> files with very small stripes, which will produce better downstream
> performance.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
