[IMPALA-1387] On-clause conjuncts of anti joins must be evaluated by the anti join.

• Type: Bug
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.0
• Fix Version/s: Impala 2.1
• Labels: correctness
• Assignee: Dimitris Tsirogiannis
• Reporter: Alexander Behm
• Created: 15/Oct/14 20:10
• Updated: 14/May/15 21:48
• Resolved: 21/Nov/14 17:37

Description

Query: explain select a.* from functional.alltypes a cross join functional.alltypes b where not exists (select * from functional.alltypes c where a.id = c.id and a.int_col < b.int_col)
+-----------------------------------------------------------+
| Explain String                                            |
+-----------------------------------------------------------+
| Estimated Per-Host Requirements: Memory=320.06MB VCores=2 |
|                                                           |
| 07:EXCHANGE [UNPARTITIONED]                               |
| |                                                         |
| 04:HASH JOIN [LEFT ANTI JOIN, BROADCAST]                  |
| |  hash predicates: a.id = c.id                           |
| |                                                         |
| |--06:EXCHANGE [BROADCAST]                                |
| |  |                                                      |
| |  02:SCAN HDFS [functional.alltypes c]                   |
| |     partitions=24/24 size=478.45KB                      |
| |                                                         |
| 03:CROSS JOIN [BROADCAST]                                 |
| |  predicates: a.int_col < b.int_col                      |
| |                                                         |
| |--05:EXCHANGE [BROADCAST]                                |
| |  |                                                      |
| |  01:SCAN HDFS [functional.alltypes b]                   |
| |     partitions=24/24 size=478.45KB                      |
| |                                                         |
| 00:SCAN HDFS [functional.alltypes a]                      |
|    partitions=24/24 size=478.45KB                         |
+-----------------------------------------------------------+

It's incorrect to assign "a.int_col < b.int_col" in the cross join because of NULLs. We also need to preserve those roes from the cross join where "(a.int_col < b.int_col) IS NULL" but the simpler fix is to just assign the predicate at the anti join.

Issue Links

• relates to: IMPALA-1371 Predicate applied incorrectly when FULL OUTER JOIN is present