[moraistejerina]

Which issue does this PR close?

• Closes optimize_projections fails with "No field named ..." after propagate_empty_relation removes the first branch of a UNION #26516.

Rationale for this change

A UNION ALL fails to plan with Schema error: No field named ... when the optimizer removes its first input and a later input remains. For example:

SELECT 1 WHERE false UNION ALL (SELECT 2 LIMIT 1);

propagate_empty_relation puts a projection over the remaining input with Projection::new_from_schema. That function makes column references from the union's field names, which come from the first input. When the remaining input has other names, the projection refers to columns that do not exist, and optimize_projections fails.

What changes are included in this PR?

• When one input remains and its schema is different from the union's schema, map its columns to the union's output fields by position, alias each one to the union's name, and build the projection with Projection::try_new (which also checks the result).
• Three existing snapshots change their text from Projection: a, b, c to Projection: test.a AS a, test.b AS b, test.c AS c. In those tests the remaining input is the first input, so the old by-name references worked. The output schema is the same.
• This also changes the EXPLAIN output of the struct.slt test added in fix: resolve pass-through columns against the input when merging an extraction projection #25412. That PR made push_down_leaf_projections handle the bare-name projection that propagate_empty_relation left after it removed an empty union input. With this change, the projection aliases the qualified input columns (leaf_merge_source.s AS s). The test still plans and returns the same rows. fix: resolve pass-through columns against the input when merging an extraction projection #25412 also handles this same-name alias form, and its unit tests still build the bare-name shape directly.

What is the testing strategy for this PR?

• New unit test propagate_union_remaining_input_with_other_names: the first input is empty and the second input has other column names. Without the fix, the rule produces Projection: a, b, c over an input that only has x, y, z.
• New union.slt cases with the two queries from the issue.

Are there any user-facing changes?

No. The queries that failed now return results. There are no public API changes.

[moraistejerina]

@alamb could you take a look when you have a chance? This fixes #26516: when propagate_empty_relation removes the first input of a union, it builds a projection with column names that the remaining input does not have.

[gabotechs]

👍 no comments, all good here!

I've verified that this is an actual problem present today that other SQL engines do not have:

• Query failing in the DataFusion fiddle
• Query succeeding in the Postgres fiddle

[gabotechs]

Thanks @moraistejerina!