[
https://issues.apache.org/jira/browse/HIVE-14246?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Gopal V updated HIVE-14246:
---------------------------
Description:
The CUSTOM_SIMPLE_EDGE impl has differences between the size constraints of
either edge which cannot be represented by the ShuffleVertexManager presently.
Reducing the width based on the hashtable build side vs the streaming probe
side have different consequences since there is no order of runtime between
them.
Until the two parent vertices of the shuffle hash-join are related, this
feature causes massive inconsistency of performance across runs.
For inner & semi joins, the hashtable side should have a higher priority than
the streaming side and for left outer joins, the streaming side can over-take
the hashtable side, being the more dominant factor in the final row-counts.
Until such priorities can be bubbled up into ShuffleVertexManager, this feature
can be disabled.
was:
The CUSTOM_SIMPLE_EDGE impl has differences between the size constraints of
either edge which cannot be represented by the ShuffleVertexManager presently.
Reducing the width based on the hashtable build side vs the streaming probe
side have different consequences since there is no order of runtime between
them.
Until the two parent vertices of the shuffle hash-join are related, this
feature causes massive inconsistency of performance across runs.
> Tez: disable auto-reducer parallelism when CUSTOM_EDGE is in place
> ------------------------------------------------------------------
>
> Key: HIVE-14246
> URL: https://issues.apache.org/jira/browse/HIVE-14246
> Project: Hive
> Issue Type: Bug
> Components: Tez
> Affects Versions: 2.2.0
> Reporter: Gopal V
> Assignee: Gopal V
> Fix For: 2.2.0
>
>
> The CUSTOM_SIMPLE_EDGE impl has differences between the size constraints of
> either edge which cannot be represented by the ShuffleVertexManager presently.
> Reducing the width based on the hashtable build side vs the streaming probe
> side have different consequences since there is no order of runtime between
> them.
> Until the two parent vertices of the shuffle hash-join are related, this
> feature causes massive inconsistency of performance across runs.
> For inner & semi joins, the hashtable side should have a higher priority than
> the streaming side and for left outer joins, the streaming side can over-take
> the hashtable side, being the more dominant factor in the final row-counts.
> Until such priorities can be bubbled up into ShuffleVertexManager, this
> feature can be disabled.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)