[HIVE-16869] Hive returns wrong result when predicates on non-existing columns are pushed down to Parquet reader

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Fixed
• Fix Version/s: 3.0.0
• Assignee: Yibing Shi Assign to me
• Reporter: Yibing Shi
• Created: 09/Jun/17 14:59
• Updated: 22/May/18 23:58
• Resolved: 27/Jun/17 13:23

Description

When hive.optimize.ppd and hive.optimize.index.filter are turned, and a select query has a condition on a column that doesn't exist in Parquet file (such as a partition column), Hive often returns wrong result.

Please see below example for details:

hive> create table test_parq (a int, b int) partitioned by (p int) stored as parquet;
OK
Time taken: 0.292 seconds
hive> insert overwrite table test_parq partition (p=1) values (1, 2);
OK
Time taken: 5.08 seconds
hive> select * from test_parq where a=1 and p=1;
OK
1	2	1
Time taken: 0.441 seconds, Fetched: 1 row(s)
hive> select * from test_parq where (a=1 and p=1) or (a=999 and p=999);
OK
1	2	1
Time taken: 0.197 seconds, Fetched: 1 row(s)
hive> set hive.optimize.index.filter=true;
hive> select * from test_parq where (a=1 and p=1) or (a=999 and p=999);
OK
Time taken: 0.167 seconds
hive> select * from test_parq where (a=1 or a=999) and (a=999 or p=1);
OK
Time taken: 0.563 seconds

Issue Links

• breaks: HIVE-17052 Remove logging of predicate filters
• is related to: HIVE-16661 Parquet storage does not handle 'or' statement properly