SIMILAR TO: SQL planner rejects non-literal LargeUtf8 / Utf8View patterns

[adriangb]

Is your feature request related to a problem or challenge?

Follow-up to #23704 (which fixed the SIMILAR TO "failed to downcast array" panic from #22886 by coercing operands in the analyzer).

The analyzer now coerces SIMILAR TO operands to a common string type, but the SQL planner still rejects LargeUtf8 and Utf8View patterns before the analyzer ever runs. In datafusion/sql/src/expr/mod.rs:

let pattern_type = pattern.get_type(schema)?;
if pattern_type != DataType::Utf8 && pattern_type != DataType::Null {
    return plan_err!("Invalid pattern in SIMILAR TO expression");
}

So a non-literal pattern of type LargeUtf8 or Utf8View fails at plan time:

-- errors: "Invalid pattern in SIMILAR TO expression"
SELECT s FROM test, patterns WHERE s SIMILAR TO arrow_cast(pat, 'Utf8View');

#23704's examples all keep the Utf8View/LargeUtf8 on the input side with a Utf8 pattern, so this path isn't exercised there. Now that the analyzer coerces both operands, this planner restriction is unnecessary.

Describe the solution you'd like

Relax the pattern-type check to also accept LargeUtf8 and Utf8View:

if !matches!(
    pattern_type,
    DataType::Utf8 | DataType::LargeUtf8 | DataType::Utf8View | DataType::Null
) {
    return plan_err!("Invalid pattern in SIMILAR TO expression");
}

plus a sqllogictest case with a non-literal Utf8View (and LargeUtf8) pattern.

Describe alternatives you've considered

An earlier alternative implementation of the same fix (#22887) already included this planner relaxation and the corresponding test, so the change is small and self-contained.

Additional context

cc @u70b3 (author of #23704) — flagging in case you want to fold this into the same area. Happy to send the PR otherwise.

[u70b3]

I'll take this — PR opened: #23735. It relaxes the planner pattern-type check to accept Utf8/LargeUtf8/Utf8View/Null and adds sqllogictest cases with non-literal Utf8View/LargeUtf8 patterns in type_coercion.slt (next to the #22886 regression tests), borrowing from #22887 as suggested.