[raams95]

Description

When table_diff has more than one key column it builds a single __sqlmesh_join_key by casting each key to a string and concatenating them. On SQL Server that cast was CAST(x AS VARCHAR) with no length, and T-SQL silently treats that as VARCHAR(30) inside CAST. This causes two problems:

• UNIQUEIDENTIFIER keys fail outright with Insufficient result space to convert uniqueidentifier value to char (mssql: table_diff error with multiple keys #4478)
• longer string keys get cut at 30 characters, so different rows can end up with the same join key and the diff reports wrong results without any error (SQL Server: table_diff __sqlmesh_join_key parts truncated at 30 characters each, table_diff fails to use character_maximum_length to set (N)VARCHAR lengths #5146)

This overrides normalize_value in the MSSQL adapter to cast to VARCHAR(MAX) instead, along the lines of what was suggested in #5146. Other engines are unchanged. Fabric inherits the same behaviour since it uses the MSSQL adapter.

Fixes #4478
Fixes #5146

Test Plan

• Added test_concat_columns_uses_varchar_max in tests/core/engine_adapter/test_mssql.py, which checks the SQL generated for a composite key with a UNIQUEIDENTIFIER and an NVARCHAR(100) column. It fails on main (CAST(id AS VARCHAR)) and passes with this change.
• pytest tests/core/engine_adapter/test_mssql.py tests/core/engine_adapter/test_fabric.py tests/core/test_table_diff.py passes, apart from 4 table_diff console tests that fail the same way on main on my Windows machine (box-drawing characters in the rich output).
• ruff check, ruff format --check and mypy are clean on the changed files.
• I haven't run this against a live SQL Server instance.

Checklist

• I have run make style and fixed any issues (ruff and ruff-format pass; mypy and the migrations hook report the same environment-only errors on main on my Windows machine)
• I have added tests for my changes (if applicable)
• All existing tests pass (make fast-test) (no new failures compared to main on my Windows machine; the Windows-only failures on main are unchanged)
• My commits are signed off (git commit -s) per the DCO

[mday-io]

Thanks, this is the right fix. The T-SQL docs confirm the default length is 30 inside CAST, and VARCHAR(MAX) is supported on Fabric too. That's also why it shouldn't be NVARCHAR, since Fabric has no nvarchar; worth a word in the comment so nobody "fixes" it later. Two test asks before merging:

• Could you add a key longer than 30 characters (or a UUID) to test_table_diff_grain_check_multiple_keys in the integration suite? That way the MSSQL Docker job actually reproduces SQL Server: table_diff __sqlmesh_join_key parts truncated at 30 characters each, table_diff fails to use character_maximum_length to set (N)VARCHAR lengths #5146 instead of us only checking the SQL string.
• Since the PR says Fabric inherits this, could you add the same assertion to test_fabric.py?

I'll approve the CI run so the MSSQL integration tests run against it.