Regression on coercing Array of Structs

[kazuyukitanimura]

Describe the bug

https://github.com/apache/datafusion/pull/12753/files#diff-f1e354d4fe26237064d8194e10a6008efa4f88e2b68b8a8352086a5d011180b8R108 introduced to use type_union_resolution; however, this caused a regression on coerce_types against Array of Structs.

type_union_resolution returns None for such cases and let the plan fail.

To Reproduce

Found at apache/datafusion-comet#1001 (comment)

Expected behavior

The fix seems to be easy. We just need to addd

                .or_else(|| struct_coercion(lhs_type, rhs_type))

at the end of type_union_resolution_coercion

Then, type_union_resolution returns Some(...) for Array of Structs.

Additional context

No response

[alamb]

@jayzhan211 has a PR to fix this here; #12839

however, this caused a regression on coerce_types against Array of Structs.

Can someone help me understand what the end user query / behavior is expected when coercing arrays of structs? It is not clear what the semantics are supposed to be, and thus I am not sure if the PR completely implements the desired semantics

[jayzhan211]

I follow what DuckDB does, since they seem to have a mature support for struct type

[jayzhan211]

D create table t(s1 struct(a int, b varchar), s2 struct(a float, b varchar));
D insert into t values (row(1, 'r'), row(2.2, 'b'));
D select [s1, s2] from t;
┌──────────────────────────────────────────┐
│         main.list_value(s1, s2)          │
│       struct(a float, b varchar)[]       │
├──────────────────────────────────────────┤
│ [{'a': 1.0, 'b': r}, {'a': 2.2, 'b': b}] │
└──────────────────────────────────────────┘
D select coalesce(s1, s2) from t;
┌────────────────────────────┐
│      COALESCE(s1, s2)      │
│ struct(a float, b varchar) │
├────────────────────────────┤
│ {'a': 1.0, 'b': r}         │
└────────────────────────────┘
D select typeof(s1) from t;
┌──────────────────────────────┐
│          typeof(s1)          │
│           varchar            │
├──────────────────────────────┤
│ STRUCT(a INTEGER, b VARCHAR) │
└──────────────────────────────┘

Coercion rule is so inconsistent, array and coalesce converts int to float, but values doesn't