[
https://issues.apache.org/jira/browse/HIVE-16113?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16002677#comment-16002677
]
Remus Rusanu commented on HIVE-16113:
-------------------------------------
I think the issue is that the partition pruning logic expects {{null}} values
to percolate through the expression tree and trigger the {{isUnknown}} case in
{{PartitionPrunner.prunePartitionNames}}. But expressions like CASE, COALESCE
or NVL (maybe other?) stop null bubbling and can evaluate the tree to a
resolute {{false}} instead, causing overaggressive partition elimination. If
the expression has these functions (NVL, COALESCE) to handle {{null}} values at
a *row* level we're evaluating them and taking a decision at *partition* level
(prune/not prune), and the 'prune' case is not safe. Perhaps we should consider
functions like NVL/COALESCE 'special' and inject 'true' instead in the pruning
expression?
> PartitionPruner::removeNonPartCols needs to handle AND/OR cases
> ---------------------------------------------------------------
>
> Key: HIVE-16113
> URL: https://issues.apache.org/jira/browse/HIVE-16113
> Project: Hive
> Issue Type: Bug
> Components: Logical Optimizer
> Affects Versions: 1.2.1, 2.1.1, 2.2.0
> Reporter: Gopal V
> Assignee: Remus Rusanu
> Attachments: HIVE-16113.1.patch
>
>
> {code}
> create table daysales (customer int) partitioned by (dt string);
> insert into daysales partition(dt='2001-01-01') values(1);
> select * from daysales where nvl(dt='2001-01-01' and customer=1, false);
> 0 ROWS
> {code}
> https://github.com/apache/hive/blob/master/ql/src/java/org/apache/hadoop/hive/ql/optimizer/ppr/PartitionPruner.java#L384
> {code}
> 2017-03-05T12:37:47,153 WARN [6f053d71-6ad6-4ad0-833d-337f2d499c82 main]
> ppr.PartitionPruner: The expr = NVL(((dt = '2001-01-01') and null),false)
> {code}
> Because {{true and null => null}}, this turns into {{NVL(null, false)}}
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
