Fix: cast NULLs in VALUES clause (df->values) to avoid coercion issues (#2856)

Author: georgesittas

sqlmesh/core/dialect.py

@@ -897,6 +897,14 @@ def select_from_values_for_batch_range(
         ]
 
     values_exp = exp.values(expressions, alias=alias, columns=columns_to_types)
+    if values:
+        # BigQuery crashes on `SELECT CAST(x AS TIMESTAMP) FROM UNNEST([NULL]) AS x`, but not
+        # on `SELECT CAST(x AS TIMESTAMP) FROM UNNEST([CAST(NULL AS TIMESTAMP)]) AS x`. This
+        # ensures nulls under the `Values` expression are cast to avoid similar issues.
+        for value, kind in zip(values_exp.expressions[0].expressions, columns_to_types.values()):
+            if isinstance(value, exp.Null):
+                value.replace(exp.cast(value, to=kind))
+
     return exp.select(*casted_columns).from_(values_exp, copy=False).where(where, copy=False)
 
 

tests/core/test_dialect.py

@@ -375,6 +375,20 @@ def test_select_from_values_for_batch_range_json():
     )
 
 
+def test_select_from_values_that_include_null():
+    values = [(1, exp.null())]
+    columns_to_types = {
+        "id": exp.DataType.build("int", dialect="bigquery"),
+        "ts": exp.DataType.build("timestamp", dialect="bigquery"),
+    }
+
+    values_expr = select_from_values_for_batch_range(values, columns_to_types, 0, len(values))
+    assert values_expr.sql(dialect="bigquery") == (
+        "SELECT CAST(id AS INT64) AS id, CAST(ts AS TIMESTAMP) AS ts FROM "
+        "UNNEST([STRUCT(1 AS id, CAST(NULL AS TIMESTAMP) AS ts)]) AS t"
+    )
+
+
 @pytest.fixture(params=["mysql", "duckdb", "postgres", "snowflake"])
 def normalization_dialect(request):
     if request.param == "duckdb":