Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
Impala 2.9.0
-
None
Description
Decimal multiplication indicates overflow before you've actually overflowed the result type. For example:
[dhecht-desktop.ca.cloudera.com:21000] > select typeof(cast(.1 as decimal(20,20)) * cast(1 as decimal(20,19))); Query: select typeof(cast(.1 as decimal(20,20)) * cast(1 as decimal(20,19))) Query submitted at: 2017-02-15 14:08:11 (Coordinator: http://dhecht-desktop.ca.cloudera.com:25000) Query progress can be monitored at: http://dhecht-desktop.ca.cloudera.com:25000/query_plan?query_id=ac4e45202a0850c4:c757c44600000000 +-----------------------------------------------------------------+ | typeof(cast(0.1 as decimal(20,20)) * cast(1 as decimal(20,19))) | +-----------------------------------------------------------------+ | DECIMAL(38,38) | +-----------------------------------------------------------------+ [dhecht-desktop.ca.cloudera.com:21000] > select cast(.1 as decimal(20,20)) * cast(1 as decimal(20,19)); Query: select cast(.1 as decimal(20,20)) * cast(1 as decimal(20,19)) Query submitted at: 2017-02-15 14:07:58 (Coordinator: http://dhecht-desktop.ca.cloudera.com:25000) Query progress can be monitored at: http://dhecht-desktop.ca.cloudera.com:25000/query_plan?query_id=c94daa3dc258be9c:559353e800000000 +---------------------------------------------------------+ | cast(0.1 as decimal(20,20)) * cast(1 as decimal(20,19)) | +---------------------------------------------------------+ | NULL | +---------------------------------------------------------+ WARNINGS: UDF WARNING: Expression overflowed, returning NULL [dhecht-desktop.ca.cloudera.com:21000] > select cast(.1 as decimal(20,20)) * cast(1 as decimal(19,18)); Query: select cast(.1 as decimal(20,20)) * cast(1 as decimal(19,18)) Query submitted at: 2017-02-15 14:12:48 (Coordinator: http://dhecht-desktop.ca.cloudera.com:25000) Query progress can be monitored at: http://dhecht-desktop.ca.cloudera.com:25000/query_plan?query_id=8d4a0189a544e194:12148c3100000000 +---------------------------------------------------------+ | cast(0.1 as decimal(20,20)) * cast(1 as decimal(19,18)) | +---------------------------------------------------------+ | 0.10000000000000000000000000000000000000 | +---------------------------------------------------------+ Fetched 1 row(s) in 0.01s
I.e. 0.1 * 1 fits in a DECIMAL(38,38), but our implementation of decimal multiply can overflow when S1+S2 > 38 even if the result would fit the result type.
The problem is that when starting with Decimal16Values, we do a 128 * 128 => 128-bit multiply, rather than 128 * 128 => 256 bit multiply. We check the intermediate result which has full (S1+S2) scale against MAX_UNSCALED_DECIMAL16, even though we later have code to scale down the result as needed to fit the result type.
Attachments
Issue Links
- blocks
-
IMPALA-4924 Make DECIMAL_V2 the default mode
- Resolved
-
IMPALA-4940 DECIMAL V2 multiply result type
- Resolved
- is related to
-
IMPALA-4964 Decimal modulo operator is overflowing
- Resolved