Make propagate_unknown respect any explicit value

propagate_unknown will still traverse any series of nested documents,
meaning that once you set propagate_unknown=True, it is true for the
whole schema structure.

However, this introduces tracking for whether or not `unknown` was set
explicitly. If `unknown=RAISE` is set because no value was specified,
we will set a new flag on the schema, `auto_unknown=True`.

propagate_unknown now has the following behavior:
- if the nested schema has auto_unknown=False, use the current value
  for `unknown` in the nested `load` call
- if a nested field has its `unknown` attribute set, use that in place
  of any value sent via `propagate_unknown`

Effectively, this means that if you set `unknown` explicitly anywhere
in a nested schema structure, it will propagate downwards from that
point.

Combined with the fact that propagate_unknown=True propagates
downwards across all schema barriers, including if
`propagate_unknown=False` is set explicitly somewhere, this could be
confusing. However, because the idea is for `propagate_unknown=True`
to eventually be the only supported behavior for marshmallow,
this is acceptable as a limitation.

auto_unknown is an attribute of schema opts and of schema instances,
with the same kind of inheritance behavior as other fields.

Author: sirosen

src/marshmallow/fields.py

@@ -608,6 +608,15 @@ def _deserialize(
             Add ``partial`` parameter.
         """
         self._test_collection(value)
+        # check if self.unknown or self.schema.unknown is set
+        # however, we should only respect `self.schema.unknown` if
+        # `auto_unknown` is False, meaning that it was set explicitly on the
+        # schema class or instance
+        explicit_unknown = self.unknown or (
+            self.schema.unknown if not self.schema.auto_unknown else None
+        )
+        if explicit_unknown:
+            unknown = explicit_unknown
         return self._load(
             value,
             data,

src/marshmallow/schema.py

@@ -227,7 +227,14 @@ def __init__(self, meta, ordered: bool = False):
         self.include = getattr(meta, "include", {})
         self.load_only = getattr(meta, "load_only", ())
         self.dump_only = getattr(meta, "dump_only", ())
-        self.unknown = getattr(meta, "unknown", RAISE)
+        # self.unknown defaults to "RAISE", but note whether it was explicit or
+        # not, so that when we're handling propagate_unknown we can decide
+        # whether or not to propagate based on whether or not it was set
+        # explicitly
+        self.unknown = getattr(meta, "unknown", None)
+        self.auto_unknown = self.unknown is None
+        if self.auto_unknown:
+            self.unknown = RAISE
         self.propagate_unknown = getattr(meta, "propagate_unknown", False)
         self.register = getattr(meta, "register", True)
 
@@ -391,6 +398,9 @@ def __init__(
         self.dump_only = set(dump_only) or set(self.opts.dump_only)
         self.partial = partial
         self.unknown = unknown or self.opts.unknown
+        # if unknown was not set explicitly AND self.opts.auto_unknown is true,
+        # then the value should be considered "automatic"
+        self.auto_unknown = (not unknown) and self.opts.auto_unknown
         self.propagate_unknown = propagate_unknown or self.opts.propagate_unknown
         self.context = context or {}
         self._normalize_nested_options()

tests/test_schema.py

@@ -2901,3 +2901,66 @@ class DefinitelyUniqueSchema(Schema):
 
     SchemaClass = class_registry.get_class(DefinitelyUniqueSchema.__name__)
     assert SchemaClass is DefinitelyUniqueSchema
+
+
+def test_propagate_unknown_stops_at_explicit_value_for_nested():
+    # propagate_unknown=True should traverse any "auto_unknown" values and
+    # replace them with the "unknown" value from the parent context (schema or
+    # load arguments)
+    # this test makes sure that it stops when a nested field or schema has
+    # "unknown" set explicitly (so auto_unknown=False)
+
+    class Bottom(Schema):
+        x = fields.Str()
+
+    class Middle(Schema):
+        x = fields.Str()
+        # set unknown explicitly on a nested field, so auto_unknown will be
+        # false going into Bottom
+        child = fields.Nested(Bottom, unknown=EXCLUDE)
+
+    class Top(Schema):
+        x = fields.Str()
+        child = fields.Nested(Middle)
+
+    data = {
+        "x": "hi",
+        "y": "bye",
+        "child": {"x": "hi", "y": "bye", "child": {"x": "hi", "y": "bye"}},
+    }
+    result = Top(unknown=INCLUDE, propagate_unknown=True).load(data)
+    assert result == {
+        "x": "hi",
+        "y": "bye",
+        "child": {"x": "hi", "y": "bye", "child": {"x": "hi"}},
+    }
+
+
+def test_propagate_unknown_stops_at_explicit_value_for_meta():
+    # this is the same as the above test of propagate_unknown stopping where
+    # auto_unknown=False, but it checks that this applies when `unknown` is set
+    # by means of `Meta`
+
+    class Bottom(Schema):
+        x = fields.Str()
+
+    class Middle(Schema):
+        x = fields.Str()
+        child = fields.Nested(Bottom)
+
+        # set unknown explicitly on a nested field, so auto_unknown will be
+        # false going into Bottom
+        class Meta:
+            unknown = EXCLUDE
+
+    class Top(Schema):
+        x = fields.Str()
+        child = fields.Nested(Middle)
+
+    data = {
+        "x": "hi",
+        "y": "bye",
+        "child": {"x": "hi", "y": "bye", "child": {"x": "hi", "y": "bye"}},
+    }
+    result = Top(unknown=INCLUDE, propagate_unknown=True).load(data)
+    assert result == {"x": "hi", "y": "bye", "child": {"x": "hi", "child": {"x": "hi"}}}