You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@hive.apache.org by "Gopal V (JIRA)" <ji...@apache.org> on 2018/06/15 23:19:00 UTC
[jira] [Comment Edited] (HIVE-11573) PointLookupOptimizer can be
pessimistic at a low nDV
[ https://issues.apache.org/jira/browse/HIVE-11573?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16514489#comment-16514489 ]
Gopal V edited comment on HIVE-11573 at 6/15/18 11:18 PM:
----------------------------------------------------------
[~kgyrtkirk]: this was written to prevent algorithmic complexity of the expression walkers because (a = 1 || a = 2) is actually UDFOr(UDFOpEq(a, 1), UDFOpEq(a, 2)) while the other one is GenericUDFIn(a, [1,2]) which is faster to walk through.
The original bad case had 680 x (a = 1 and b = 2) or (a=1 and b=3) ... etc, which got generated as (a=1) and (b=(2,3,...)) and ((a,b) IN ((1,2),(2,3)...) so that the disjunction could be pushed down to the ORC PPD.
was (Author: gopalv):
[~kgyrtkirk]: this was written to prevent algorithmic complexity of the expression walkers because (a = 1 || a = 2) is actually UDFOr(UDFOpEq(a, 1), UDFOpEq(a, 2)) while the other one is GenericUDFIn(a, [1,2]) which is faster to walk through.
The original bad case had 680 (a = 1 and b = 2) or (a=1 and b=3) ... etc, which got generated as (a=1) and (b=(2,3,...)) and ((a,b) IN ((1,2),(2,3)...) so that the disjunction could be pushed down to the ORC PPD.
> PointLookupOptimizer can be pessimistic at a low nDV
> ----------------------------------------------------
>
> Key: HIVE-11573
> URL: https://issues.apache.org/jira/browse/HIVE-11573
> Project: Hive
> Issue Type: Bug
> Affects Versions: 2.0.0
> Reporter: Gopal V
> Assignee: Gopal V
> Priority: Major
> Fix For: 2.0.0
>
> Attachments: HIVE-11573.1.patch, HIVE-11573.2.patch, HIVE-11573.3.patch, HIVE-11573.4.patch, HIVE-11573.5.patch, HIVE-11573.6.patch
>
>
> The PointLookupOptimizer can turn off some of the optimizations due to its use of tuple IN() clauses.
> Limit the application of the optimizer for very low nDV cases and extract the sub-clause as a pre-condition during runtime, to trigger the simple column predicate index lookups.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)