Javadoc
We need to override this method since Calcite and Drill calculate
joined row count in different ways. It helps avoid a case when
at the first time was used Drill join row count but at the second time
Calcite row count was used. It may happen when
RelMdDistinctRowCount#getDistinctRowCount(Join,RelMetadataQuery,ImmutableBitSet,RexNode) method is used and after that used
another getDistinctRowCount method for parent rel, which just uses
row count of input rel node (our join rel).
It causes cost increase of best rel node when
RelSubset#propagateCostImprovements is called.
This is a part of the fix for CALCITE-2018.