[HIVE-22513] Constant propagation of casted column in filter ops can cause incorrect results

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Fix Version/s: 4.0.0-alpha-1
• Component/s: Query Planning
• Assignee: Ádám Szita
• Reporter: Ádám Szita
• Created: 19/Nov/19 12:45
• Updated: 17/Nov/22 08:49
• Resolved: 20/Nov/19 12:47

Description

This issue happens if CBO is disabled.

We should not be propagating constants if the corresponding ExprNodeColumnDesc instance is wrapped inside a CAST operator as casting might truncate information from the original column.

This can happen if we're using CAST in a WHERE clause, which will cause the projected columns to be replaced in a SELECT operator. Their new value will be the result of casting which could be a different value compared to that in the original column:

set hive.cbo.enable=false;
set hive.fetch.task.conversion=more; --just for testing convenience

create table testtb (id string);
insert into testtb values('2019-11-05 01:01:11');

select id, CAST(id AS VARCHAR(10)) from testtb where CAST(id AS VARCHAR(9)) = '2019-11-0';

+------------+------------+
|     id     |    _c1     |
+------------+------------+
| 2019-11-0  | 2019-11-0  |
+------------+------------+
1 row selected (0.168 seconds)

-- VS expected: 2019-11-05 01:01:11 | 2019-11-05 

As to what types of casting (from and where types) cause information loss it's hard to properly keep track of, and I don't think it should be taken into consideration when deciding whether or not to propagate a constant. Rather than adding a big and potentially convoluted and fragile check for this, I propose to prevent constant mappings to be spawned out of CASTed columns.