@@ -102,27 +102,25 @@ impl OptimizerRule for ReplaceDistinctWithAggregate {
102102 let schema = input. schema ( ) ;
103103 let field_count = schema. fields ( ) . len ( ) ;
104104 for dep in schema. functional_dependencies ( ) . iter ( ) {
105- // If the input is already unique on all of its columns (e.g.
106- // it is a GROUP BY over exactly these columns), the DISTINCT
107- // is a no-op and we can simply remove it.
108- //
109- // The dependency mode must be `Single`: a `Multi`
110- // dependence (e.g. a former key downgraded by a join) means
111- // equal rows may occur multiple times, so the DISTINCT
112- // still has work to do.
113- //
114- // The grouping columns must also not contain NULLs because
115- // a nullable UNIQUE constraint permits multiple NULL keys,
116- // but DISTINCT treats NULLs as equal and must still
117- // collapse them.
105+ // We can remove the DISTINCT entirely if functional
106+ // dependencies prove it is already unique and therefore a
107+ // DISTINCT would be a no op.
118108 let source_indices = & dep. source_indices ;
119109 let any_source_field_nullable = source_indices
120110 . iter ( )
121111 . any ( |& idx| schema. field ( idx) . is_nullable ( ) ) ;
122112 let nullable = dep. nullable && any_source_field_nullable;
123113 if !nullable
114+ // The dependency mode must be `Single` because a `Multi`
115+ // dependence means equal rows may occur multiple times
124116 && dep. mode == Dependency :: Single
125117 && source_indices. len ( ) >= field_count
118+ // The grouping columns must not contain NULLs because a
119+ // nullable dependency allows multiple rows with NULL in
120+ // the same column (because NULL != NULL). However,
121+ // DISTINCT treats NULLs as equal and collapses them, so
122+ // we can only remove the DISTINCT if we know that no
123+ // source column can actually be NULL.
126124 && source_indices[ ..field_count]
127125 . iter ( )
128126 . enumerate ( )
0 commit comments