fix(mssql): cast table_diff values to VARCHAR(MAX)

sqlmesh/core/engine_adapter/mssql.py

@@ -81,6 +81,18 @@ class MSSQLEngineAdapter(
     VARIABLE_LENGTH_DATA_TYPES = {"binary", "varbinary", "char", "varchar", "nchar", "nvarchar"}
     INSERT_OVERWRITE_STRATEGY = InsertOverwriteStrategy.MERGE
 
+    def normalize_value(
+        self,
+        expr: exp.Expr,
+        type: exp.DataType,
+        decimal_precision: int = 3,
+        timestamp_precision: int = RowDiffMixin.MAX_TIMESTAMP_PRECISION,
+    ) -> exp.Expr:
+        # T-SQL truncates CAST(x AS VARCHAR) to 30 characters when no length is given
+        normalized = super().normalize_value(expr, type, decimal_precision, timestamp_precision)
+        normalized.set("to", exp.DataType.build("VARCHAR(MAX)", dialect=self.dialect))
+        return normalized
+
     @property
     def catalog_support(self) -> CatalogSupport:
         # MSSQL and AzureSQL both use this engine adapter, but they differ in catalog support.

tests/core/engine_adapter/test_mssql.py

@@ -1055,3 +1055,17 @@ def test_comments(make_mocked_engine_adapter: t.Callable, mocker: MockerFixture)
         adapter._build_create_comment_table_exp(table, comment),
         adapter._build_create_comment_column_exp(table, "a", comment),
     ]
+
+
+def test_concat_columns_uses_varchar_max(adapter: MSSQLEngineAdapter):
+    # CAST(x AS VARCHAR) without a length truncates to 30 characters in T-SQL, which
+    # breaks the table_diff join key for long or UNIQUEIDENTIFIER key columns
+    columns_to_types = {
+        "id": exp.DataType.build("UNIQUEIDENTIFIER", dialect=adapter.dialect),
+        "name": exp.DataType.build("NVARCHAR(100)", dialect=adapter.dialect),
+    }
+
+    assert (
+        adapter.concat_columns(columns_to_types).sql(dialect=adapter.dialect)
+        == "COALESCE(CAST(id AS VARCHAR(MAX)), '') + ',' + COALESCE(CAST(name AS VARCHAR(MAX)), '')"
+    )