Repository navigation
infer_expression_type takes IfThen and SwitchExpression nullability from the first branch only #277
Description
infer_expression_type derives an IfThen's type from ifs[0].then alone (src/substrait/type_inference.py:426) and a SwitchExpression's from switch_expression.ifs[0].then (type_inference.py:433). The remaining then branches and the else branch are never consulted, so when the branches disagree on nullability the answer depends on branch order — and in one order it is unsound, declaring REQUIRED for an expression that can evaluate to null.
A COALESCE-style rewrite over a left join is exactly this shape: one branch is a required column, the other a null-padded one.
Reproducer
import substrait.algebra_pb2 as stalg, substrait.type_pb2 as stt
from substrait.type_inference import infer_expression_type
R, N = stt.Type.NULLABILITY_REQUIRED, stt.Type.NULLABILITY_NULLABLE
NAMES = {R: "REQUIRED", N: "NULLABLE"}
# An input row where field 0 is required and field 1 is nullable.
scope = stt.Type.Struct(
types=[stt.Type(i64=stt.Type.I64(nullability=R)), stt.Type(i64=stt.Type.I64(nullability=N))],
nullability=R,
)
def ref(i):
return stalg.Expression(selection=stalg.Expression.FieldReference(
root_reference=stalg.Expression.FieldReference.RootReference(),
direct_reference=stalg.Expression.ReferenceSegment(
struct_field=stalg.Expression.ReferenceSegment.StructField(field=i))))
def if_then(then, else_):
true = stalg.Expression(literal=stalg.Expression.Literal(boolean=True))
return stalg.Expression(if_then=stalg.Expression.IfThen(
ifs=[stalg.Expression.IfThen.IfClause(**{"if": true, "then": then})],
**{"else": else_}))
for label, expr in (
("then=f0 REQUIRED, else=f1 NULLABLE", if_then(ref(0), ref(1))),
("then=f1 NULLABLE, else=f0 REQUIRED", if_then(ref(1), ref(0))),
):
t = infer_expression_type(expr, scope)
print(f"{label}: {NAMES[getattr(t, t.WhichOneof('kind')).nullability]}")then=f0 REQUIRED, else=f1 NULLABLE: REQUIRED
then=f1 NULLABLE, else=f0 REQUIRED: NULLABLE
Both spellings can return null, so the first is wrong. A SwitchExpression with the same branches behaves identically.
Measured on main at afa3c52. Pre-existing.
The output nullability should be nullable if any branch — including else — is nullable. This is the same shape as the Expand switching-field gap in #269 and the grouping-key gap in #268.
🤖 Filed with AI assistance as a follow-up to review of #272.
Activity