[
https://issues.apache.org/jira/browse/HIVE-19097?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16511246#comment-16511246
]
Zoltan Haindrich commented on HIVE-19097:
-----------------------------------------
I've just checked it out; and now when the IN operators are opened; and if I
use a 1.17 calcite snapshot; further optimizations are happening (as they
should)...however, opening these INs has a negative effect on other things;
like on [HIVE-19460-in_estimator] - which after opening the ORs could not do
anything...
I see the following options right now:
# do a min/max predicate pruning inside a calcite rule; that will easily do
what [HIVE-19460-in_estimator] have done ... but it needs the INs to be
open(this patch)
# set HIVEPOINTLOOKUPOPTIMIZERMIN to a smaller value; like 3 - I think this
would also enable [HIVE-19460-in_estimator] to do better work
# add a similar logic as [HIVE-19460-in_estimator] to the stats rules for
binary operators (<,>,=,..)
There might be other options which I've not found yet....I think 1. would be
probably harder to achieve - or at least I'm not entirely sure how to get
started with it :)
I think 2+3 would be good for now...
[~jcamachorodriguez], [~ashutoshc]: what do you think about it?
> related equals and in operators may cause inaccurate stats estimations
> ----------------------------------------------------------------------
>
> Key: HIVE-19097
> URL: https://issues.apache.org/jira/browse/HIVE-19097
> Project: Hive
> Issue Type: Bug
> Reporter: Zoltan Haindrich
> Assignee: Zoltan Haindrich
> Priority: Major
> Attachments: HIVE-19097.01.patch, HIVE-19097.partial.patch
>
>
> tpcds#74 is optimized in a way that for date_dim the condition contains IN
> and = for the same column
> {code:java}
> | Map Operator Tree: |
> | TableScan |
> | alias: date_dim |
> | filterExpr: (((d_year) IN (2001, 2002) and (d_year =
> 2002) and d_date_sk is not null) or ((d_year) IN (2001, 2002) and (d_year =
> 2001) and d_date_sk is not null)) (type: boolean) |
> | Statistics: Num rows: 73049 Data size: 876588 Basic
> stats: COMPLETE Column stats: COMPLETE |
> | Filter Operator |
> | predicate: ((d_year) IN (2001, 2002) and (d_year =
> 2002) and d_date_sk is not null) (type: boolean) |
> | Statistics: Num rows: 4 Data size: 48 Basic stats:
> COMPLETE Column stats: COMPLETE |
> {code}
> the "real" row count will be 365
> for separate {{IN}} and {{=}} the estimation is very good; but if both are
> present it becomes (very) underestimated.
> {code:java}
> set hive.query.results.cache.enabled=false;
> drop table if exists t1;
> drop table if exists t8;
> create table t1 (a integer,b integer);
> create table t8 like t1;
> insert into t1 values (1,1),(2,2),(3,3),(4,4),(5,5);
> insert into t8
> select * from t1 union all select * from t1 union all select * from t1 union
> all select * from t1 union all
> select * from t1 union all select * from t1 union all select * from t1 union
> all select * from t1
> ;
> analyze table t1 compute statistics for columns;
> analyze table t8 compute statistics for columns;
> explain analyze select sum(a) from t8 where b in (2,3) group by b;
> explain analyze select sum(a) from t8 where b=2 group by b;
> explain analyze select sum(a) from t1 where b in (2,3) and b=2 group by b;
> explain analyze select sum(a) from t8 where b in (2,3) and b=2 group by b;
> {code}
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
