Improve UX for UNION vs UNION ALL (introduce a LogicalPlan::Distinct)

[andygrove]

Is your feature request related to a problem or challenge? Please describe what you are trying to do.
The current support for UNION vs UNION ALL is confusing to me, in the context of trying to map this to another query engine's support for these operators.

• The logical plan simply has the Union operator and this is currently assumed to represent UNION ALL.
• The SQL planner has logic to wrap a union with an aggregate query and that seems like something we would want in the physical plan but not in the logical plan
• The DataFrame API does not have a function for performing a regular UNION. It is possible to manually wrap a union in a distinct to achieve this but that might not be obvious to users. Also, the documentation for distinct is incorrect and says that it performs a union

DataFusion Logical Plan

❯ explain select * from foo union select * from foo;
+---------------+----------------------------------------------------------------------------------------------------------------+
| plan_type     | plan                                                                                                           |
+---------------+----------------------------------------------------------------------------------------------------------------+
| logical_plan  | Projection: #a, #b                                                                                             |
|               |   Aggregate: groupBy=[[#a, #b]], aggr=[[]]                                                                     |
|               |     Union                                                                                                      |
|               |       Projection: #foo.a, #foo.b                                                                               |
|               |         TableScan: foo projection=Some([0, 1])                                                                 |
|               |       Projection: #foo.a, #foo.b                                                                               |
|               |         TableScan: foo projection=Some([0, 1])                                                                 |
+---------------+----------------------------------------------------------------------------------------------------------------+

Spark Logical Plan

'Distinct
+- 'Union false, false
   :- 'Project ['a]
   :  +- 'UnresolvedRelation [t1], [], false
   +- 'Project [unresolvedalias(1, None)]
      +- 'UnresolvedRelation [t1], [], false

Spark Physical Plan

*(2) HashAggregate(keys=[a#4], functions=[], output=[a#4])
+- Exchange hashpartitioning(a#4, 200), ENSURE_REQUIREMENTS, [id=#26]
   +- *(1) HashAggregate(keys=[a#4], functions=[], output=[a#4])
      +- Union
         :- LocalTableScan [a#4]
         +- LocalTableScan [1#6]

Describe the solution you'd like
I think what we want is:

• Introduce a Distinct operator in the logical plan
• Have the DataFrame distinct function create the Distinct operator instead of the aggregate query
• The DataFrame API should have functions for both union (existing method representing UNION ALL) and union_distinct
• The physical planner should translate Distinct to an aggregate query

Describe alternatives you've considered
Leave things as they are and improve the documentation.

Additional context
For users of DataFusion for SQL query planning, it would be easier to map union/union all to other engines rather than trying to reverse engineer the aggregate query wrapping the union.

[andygrove]

@alamb I would appreciate your insights here when you have time

[alamb]

Given the usecase is trying to map Union plans to what Spark does, using an explicit Distinct in the logical plan (that is converted into a GroupBy when doing physical planning makes sense to me 👍

Doing the UNION --> GroupBy transformation during SQL planning was probably a shortcut.

[mrob95]

Will take a look at this