[SPARK-56054][SQL] Undo handling aliased assignments in MERGE INTO schema evolution and add tests

[johanl-db]

What changes were proposed in this pull request?

Follow up from #54891 in particular this comment.

Adds more tests covering schema evolution in MERGE INTO when the assignment contains an alias.
This change also undoes the fix from #54891. On closer inspection, this isn't needed in Spark as Spark already removes trivial aliases on nested field accessors in MERGE in resolveExprInAssignment, which is what the change aimed to allow. See this test added here covering trivial aliases implictily added on nested field access.

Delta doesn't strip aliases during resolution in that way and will require special handling, but that's up to Delta to implement in its custom resolution logic to replicate what Spark already does

How was this patch tested?

Adds tests covering MERGE INTO schema evolution with aliases in assignments

[szehon-ho]

these are all DataFrame API, wondering is there any equivalent in SQL? Be good to abstract it out if possible

If not, maybe we can we make a new file MergeIntoSchemaEvolutionExtraDataFrameTests to make them only run in DataFrame mode?

It is a bit of a mess now due to the inheritance patterns

[johanl-db]

I moved the tests to a dedicated trait for dataframe tests.

There's no SQL equivalent, it's not possible to specify an alias in an assignment expression using SQL afaict: SET col = source.col AS other_col

[aokolnychyi]

Okay, this seems like a behavior change compared to the initial PR then? The code in master would evolve the schema in this case because we had that alias support?

[johanl-db]

Correct. The main motivation for the initial PR adding support for alias was to support nested assignments SET col.x = s.col.x, which translates to Assignment("col.x", Alias(GetStructField("col", "x"), "x")
I've noticed in the meantime that Spark actually strips the alias on struct field access after resolution (Delta doesn't).

We don't have a strong argument to support the remaining cases where the user provides an explicit alias using the dataframe API, so I'd rather not allow these

[aokolnychyi]

How does Delta behave in this case?

[johanl-db]

Delta allows schema evolution here, but only because it allows arbitrary expressions on the right hand side of the assignment. col("source.info").as("something_else") can qualifies for schema evolution the same way col("source.other_column").plus(lit(1)) does.
Spark rejects both (with this change)

[aokolnychyi]

@johanl-db, what about cases when the source query has an alias?

MERGE INTO t
USING (SELECT 'blah' AS col1, 1 AS col2 FROM source) AS s
ON false
WHEN NOT MATCHED THEN INSERT *

[johanl-db]

@johanl-db, what about cases when the source query has an alias?

MERGE INTO t
USING (SELECT 'blah' AS col1, 1 AS col2 FROM source) AS s
ON false
WHEN NOT MATCHED THEN INSERT *

Discussed directly with @aokolnychyi and @szehon-ho :
The shape of source doesn't matter at all, the only parts that matter when performing schema evolution are:

• The schema of the source
• The assignment expressions in the match clauses

MERGE doesn't look inside the source query to see what it contains, so the aliases there are irrelevant.
This case will trigger schema evolution for all new columns in the source schema, since there's an INSERT * assignment.

[johanl-db]

CI is failing due to formatting on files that are not touched in this PR. Likely from #55281

[szehon-ho]

I think this one is ready to go @cloud-fan

i also added the test requested by @aokolnychyi in #55304 (i didnt see the comment), we can add it separately , or optionally as part this pr, i guess it doesnt matter. As we looked into it together, its a bit orthogonal to this pr (the aliases within source doens't matter, to MERGE it just comes as a source table)

[szehon-ho]

also , apparently this ci error is fixed by #55334

[johanl-db]

I think this one is ready to go @cloud-fan

i also added the test requested by @aokolnychyi in #55304 (i didnt see the comment), we can add it separately , or optionally as part this pr, i guess it doesnt matter. As we looked into it together, its a bit orthogonal to this pr (the aliases within source doens't matter, to MERGE it just comes as a source table)

@szehon-ho I picked up your test from #55304 in this PR