Chore!!: Migrate sqlglot to v30

[VaggelisD]

Summary of Changes

File	Change	Why
pyproject.toml	sqlglot[rs]~=28.10.1 → sqlglot~=30.0.1	Version bump; [rs] extra removed (replaced by sqlglotc)
60+ files across sqlmesh/, tests/, web/	exp.Expression → exp.Expr in all type hints, isinstance checks, and type casts. Class definitions remain as exp.Expression subclasses	sqlglot v30 introduced exp.Expr as the new base class; Expression is now a subclass. Functions like exp.alias_(), exp.func(), parse_one() return Expr, causing mypy errors when assigned to Expression-typed variables. Custom expression classes must still inherit Expression since Expr doesn't support constructor arguments
sqlmesh/core/schema_diff.py, tests/core/integration/utils.py	exp.DataType.Type → exp.DType in type annotations	DataType.Type enum is no longer recognized as a valid type by mypy; exp.DType is the v30 alias
sqlmesh/core/dialect.py	New import path for AthenaTrinoParser, updated self.expression() calls	sqlglot API: Athena parser renamed, expression() now takes pre-constructed objects
sqlmesh/utils/jinja.py	Import Expression from sqlglot.expressions instead of sqlglot	sqlglot.expressions is now a package; top-level re-export changed
sqlmesh/core/model/kind.py	Clear meta["sql"] in _time_data_type_validator	DataType.build now preserves the parsed expression directly. Our _parse_types extension sets meta["sql"], which the pydantic encoder prioritizes over dialect-aware rendering
sqlmesh/core/model/meta.py	Normalize column types through dialect roundtrip during deserialization	BigQuery INT no longer auto-maps to BIGINT during DataType.build; roundtripping through dialect SQL restores canonical form for stable data hashes
sqlmesh/utils/metaprogramming.py	Added _resolve_import_module() to walk module hierarchy for re-exports	to_table.__module__ is now sqlglot.expressions.builders but it's re-exported from sqlglot.expressions; generated import statements must use the public module
sqlmesh/core/context_diff.py	Added sqlglotc to IGNORED_PACKAGES	New sqlglot conditionally imports sqlglotc (C extension); dependency detection was picking it up as a user requirement
tests/core/test_macros.py	Simplified expected Snowflake ARRAY_GENERATE_RANGE arithmetic	sqlglot now simplifies (DATEDIFF(...) + 1 - 1) + 1 to DATEDIFF(...) + 1
tests/core/test_config.py	Relaxed regex for error message module path	Column class now reports __module__ as sqlglot.expressions.core instead of sqlglot.expressions

[VaggelisD]

The DuckDB test failure is preexisting

[georgesittas]

Is there a test for this? Why did this work before the bump? Do we risk messing up the serialized version of this field by popping sql?

[VaggelisD]

I think it was because of this commit tobymao/sqlglot#7092.

On v28, we'd:

1. Parse the string text into exp.DataType with parse_one (setting meta["sql"] in the process)
2. Reconstruct a fresh object again before returning, effectively discarding meta

On v30 we only do (1) now after the optimization PR.

We should have the same serialization and data hashes, otherwise e.g we'd notice BigQuery parsing "TIMESTAMP" into exp.DType.DATETIME but roundtripping it to "TIMESTAMP" again due to meta taking priority as raw text verbatim.

[georgesittas]

Ok, so the current behavior is coincidentally consistent with what we used to do before the migration, because of the buggy AST node reconstruction in v28?

[georgesittas]

Why do we need a cast here?

[VaggelisD]

Ah, on v30 that was widened to exp.Expr, I'll widen the type hints for all the call chains

[georgesittas]

Got it, thanks.

[georgesittas]

Doesn't this only work for 1 parent right now? For example, if the new module became sqlglot.expressions.foo.bar, would you stop at foo?

[VaggelisD]

This is under a for loop walking up to the root module, this should work I believe

[georgesittas]

I meant that if foo re-exports bar in the above example, don't we stop at foo? Which means that we'll still see a diff in the serialized import due to a different path.