Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
None
-
None
Description
In queries like TPC-DS Q67, the cardinality of the PARTITION BY expression of the analytic may be much lower than the parallelism of the input fragment. In this case the runtime of the sort can be skewed. We could mitigate the problem by doing the expensive sort before the exchange, so that the analytic fragment only needs to merge together its sorted input and evaluate the analytic over it.
The impact of this is greater with multithreading, so I am considering only change the default when mt_dop > 0
Attachments
Issue Links
- is related to
-
IMPALA-4969 With clustered hint, consider sort->exhchange->insert plan
- Open
- relates to
-
IMPALA-9853 Push rank() predicates into sort
- Resolved
-
IMPALA-9983 Push limit from a top level sort onto analytic sort when applicable
- Resolved