Details
-
Bug
-
Status: Open
-
Major
-
Resolution: Unresolved
-
0.8.0
-
None
Description
The suite of unit tests in org.apache.drill.TestTpchDistributed are supposed to have parallel joins, parallel aggregations and sort. In order to force parallelism, the slice_target value is set to 10 (default is 100K). However, it looks like the plans of at least a few queries that I tried (query 03, 05, 14) did not have parallelism for the hash joins. The aggregations and sort do have Exchanges, so they are getting parallelized.
There is a likelihood that many plans have BroadcastExchange on the right input of the join and since the left input of the join is typically a single file we may be dropping the exchange on the right input. This should be fixed.