Add presence tracking to structures (#52)

designs/shapes.md

@@ -253,12 +253,16 @@ whose constructors only allow keyword arguments. For example:
 
 ```python
 class ExampleStructure:
+    required_param: str
+    struct_param: OtherStructure
+    optional_param: str | None
+
     def __init__(
         self,
         *, # This prevents positional arguments
         required_param: str,
         struct_param: OtherStructure,
-        optional_param: Optional[str] = None,
+        optional_param: str | None = None
     ):
         self.required_param = required_param
         self.struct_param = struct_param
@@ -270,14 +274,14 @@ class ExampleStructure:
             "StructParam": self.struct_param.as_dict(),
         }
 
-        if self.optional_param:
+        if self.optional_param is not None:
             d["OptionalParam"] = self.optional_param
 
     @staticmethod
     def from_dict(d: Dict) -> ExampleStructure:
         return ExampleStructure(
             required_param=d["RequiredParam"],
-            struct_param=OtherStructure.from_dict(d["StructParam"])
+            struct_param=OtherStructure.from_dict(d["StructParam"]),
             optional_param=d.get("OptionalParam"),
         )
 ```
@@ -369,6 +373,204 @@ print(example_dict.get("OptionalParam")) # None
 This is a small example of a minor annoyance, but one that you must always be
 aware of when using dicts.
 
+### Default Values
+
+Default values on structures are indicated by wrapping them in a simple class.
+
+```python
+class _DEFAULT:
+    def __init__(self, wrapped: Any):
+        """Wraps a value to signal it was provided by default.
+
+        These values will be immediately unwrapped in the associated
+        initializers so the values can be used as normal, the defaultedness
+        will then be tracked separately.
+        """
+        self._wrapped = wrapped
+
+    @property
+    def value(self) -> Any:
+        # Prevent mutations from leaking by simply returning copies of mutable
+        # defaults. We could also just make immutable subclasses.
+        if isinstance(self._wrapped, list):
+            return list(self._wrapped)
+        if isinstance(self._wrapped, dict):
+            return dict(self._wrapped)
+        return self._wrapped
+
+    def __repr__(self) -> str:
+        return f"_DEFAULT({repr(self._wrapped)})"
+
+    def __str__(self) -> str:
+        return str(self._wrapped)
+
+
+D = TypeVar("D")
+
+
+def _default(value: D) -> D:
+    """Wraps a value to signal it was provided by default.
+
+    These values will be immediately unwrapped in the associated
+    initializers so the values can be used as normal, the defaultedness
+    will then be tracked separately.
+
+    We use this wrapper function for brevity, but also because many completion
+    tools will show the code of the default rather than the result, and
+    `_default(7)` is a bit more clear than `cast(int, _DEFAULT(7))`.
+    """
+    return cast(D, _DEFAULT(value))
+
+
+class StructWithDefaults:
+    default_int: int
+    default_list: list
+
+    def __init__(
+        self,
+        *,
+        default_int: int = _default(7),
+        default_list: list = _default([]),
+    ):
+        self._has: dict[str, bool] = {}
+        self._set_default_attr("default_int", default_int)
+        self._set_default_attr("default_list", default_list)
+
+    def _set_default_attr(self, name: str, value: Any) -> None:
+        # Setting the attributes this way saves a ton of lines of repeated
+        # code.
+        if isinstance(value, _DEFAULT):
+            object.__setattr__(self, name, value.value())
+            self._has[name] = False
+        else:
+            setattr(self, name, value)
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        object.__setattr__(self, name, value)
+        self._has[name] = True
+
+    def _hasattr(self, name: str) -> bool:
+        if self._has[name]:
+            return True
+        # Lists and dicts are mutable. We could make immutable variants, but
+        # that's kind of a bad experience. Instead we can just check to see if
+        # the value is empty.
+        if isinstance((v := getattr(self, name)), (dict, list)) and len(v) == 0:
+            self._has[name] = True
+            return True
+        return False
+```
+
+One of the goals of customizable default values is to reduce the amount of
+nullable members that are exposed. With that in mind, the typical strategy of
+assigning the default value to `None` can't be used since that implicitly adds
+`None` to the type signature. That would also make IntelliSense marginally
+worse since you can't easily see the actual default value.
+
+Instead, a default wrapper is used. The presence of the wrapper signifies to
+the initializer function that a default was used. The value is then immediately
+unwrapped so it can be used where needed. The defaultedness is stored in an
+internal dict that is updated whenever the property is re-assigned. A private
+function exists to give the serializer this information.
+
+To make this wrapper class pass the type checker, it is simply "cast" to the
+needed type. This isn't a problem since the true value is immediately unwrapped
+in the initializer. A wrapper function performs the actual wrapping. This has
+the advantage of not requiring the type signature to be repeated since it can
+be inferred from the type of the function's input. It also has the advantage of
+looking a bit nicer in many IntelliSense tools, who show the code assigned to
+as the default value rather than the resolved value.
+
+#### Alternative: Subclassing
+
+One potential alternative is to create "default" subclasses of the various
+defaultable types.
+
+```python
+class _DEFAULT_INT(int):
+    pass
+
+
+class _DEFAULT_STR(str):
+    pass
+
+
+class WithWrappers:
+    def __init__(
+        self,
+        *,
+        default_int: int = _DEFAULT_INT(7),
+        default_str: str = _DEFAULT_STR("foo"),
+    ):
+        self.default_int = default_int
+        self.default_str = default_str
+```
+
+The advantage of this is that it requires no upkeep and no lying to the type
+system. These values are real, normal value that can be used everywhere their
+base classes can. During serialization we can check if it's the default
+type.
+
+Unfortunately, this isn't wholly possible because not all of the defaultable
+values can be subclassed. Neither `bool` nor `NoneType` can have subclasses,
+so we'd need to create our own sentinel values. This risks unexpected behavior
+if a customer were to use an `is` check.
+
+#### Alternative: kwargs
+
+Another possible alternative is to use the keyword arguments dictionary
+feature.
+
+```python
+class _WithKwargsType(TypedDict):
+    default_int: NotRequired[int]
+    default_str: NotRequired[str]
+    default_list: NotRequired[list[str]]
+
+
+class WithKwargs:
+    default_int: int
+    default_str: str
+    default_list: list[str]
+
+    # This syntax for typing kwargs requires PEP 692
+    def __init__(self, **kwargs: **_WithKwargsType):
+        self._has = {}
+        self.default_int = kwargs.get("default_int", 7)
+        self._has["default_int"] = "default_int" in kwargs
+        self.default_str = kwargs.get("default_str", "foo")
+        self._has["default_str"] = "default_str" in kwargs
+        self.default_list = kwargs.get("default_list", [])
+        self._has["default_list"] = "default_list" in kwargs
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        object.__setattr__(self, name, value)
+        self._has[name] = True
+
+    def _hasattr(self, name: str) -> bool:
+        if self._has[name]:
+            return True
+        if isinstance((v := getattr(self, name)), (dict, list)) and len(v) == 0:
+            self._has[name] = True
+            return True
+        return False
+```
+
+This leverages another feature of python that natively allows for presence
+checks. The kwargs dictionary implicitly contains that metadata because keys
+not set simply aren't present. This otherwise uses the same internal dict
+mechanism to continue to keep track of defaultedness.
+
+The major disadvantage to this is that it essentially forgoes IntelliSense
+and type checking until [PEP 692](https://peps.python.org/pep-0692/) lands.
+This isn't expected to happen until 3.12 at the earliest, which is expected
+in late 2023 / early 2024. Then the tools need to be updated for support,
+which isn't straight-forward.
+
+Another disadvantage is that it excludes the ability to include the default
+value in the IntelliSense since the typing of kwargs relies on TypedDicts
+which don't support default values.
+
 ## Errors
 
 Modeled errors are specialized structures that have a `code` and `message`.