Skip to content
Docs Pydantic Docs

functools module

Monty implements a small subset of functools. The implemented callables match CPython 3.14 for arguments, values, repr() and error messages, apart from the notes below.

Implemented

reduce(function, iterable, /[, initial]) and partial(func, /, *args, **keywords).

Not implemented

Everything else: cache, lru_cache, cached_property, wraps, update_wrapper, partialmethod, total_ordering, singledispatch, singledispatchmethod, cmp_to_key, get_cache_token, WRAPPER_ASSIGNMENTS, WRAPPER_UPDATES.

functools.Placeholder, added in 3.14 to skip a positional slot when binding (partial(f, Placeholder, 2)), is also absent, so every bound positional fills a leading slot.

These names are absent from the module namespace rather than stubbed, so they are rejected at type-check time (Module 'functools' has no member 'lru_cache') and raise AttributeError at runtime.

Behavioural divergences

  • reduce() cannot call a host function. The reduction function runs through the same synchronous path as map()’s, which cannot suspend the VM. An external function raises NotImplementedError: reduce(): external function 'f' is not yet supported in this context; one that touches the filesystem raises the same error naming the OS function it maps to, e.g. reduce(): OS function 'Path.iterdir' is not yet supported in this context for os.listdir(). This covers a partial that wraps one. Calling a partial anywhere else is unaffected, since that goes through the ordinary call path.
  • Calling a partial charges the native re-entry budget. A partial stored as a class attribute binds as a bound method whose __func__ is another partial, so a chain of them nests on the interpreter’s own call stack without pushing a Python frame. Monty bounds that chain at the fixed native re-entry depth (see resource_limits.md), raising RecursionError: maximum recursion depth exceeded beyond roughly a dozen levels; CPython runs such a chain into the thousands before its own C stack gives out. Ordinary use is unaffected — nested partial(partial(f, 1), 2) is flattened at construction, and a partial passed to map() or sorted(key=) costs one level.
  • partial objects have no __dict__. CPython allows arbitrary attributes on one (p.x = 1), which Monty rejects with AttributeError: 'functools.partial' object has no attribute 'x' and no __dict__ for setting new attributes. Assigning to func, args or keywords fails in both, but CPython words it AttributeError: readonly attribute.
  • args and keywords are rebuilt on each access. p.args is p.args is False, where CPython returns the same objects every time. Mutating the dict from p.keywords therefore has no effect on what the partial passes on; CPython returns the live dict, so mutating it changes later calls.
  • type(...).__name__ is the dotted name. type(functools.partial(f)).__name__ is 'functools.partial', where CPython reports the bare 'partial'. This is Monty’s general treatment of types whose CPython tp_name is dotted (re.Pattern and itertools.count behave the same way); str(type(p)) matches CPython’s "<class 'functools.partial'>", as do error messages naming the type.
  • partial objects carry no dunder attributes. p.__call__ and p.__doc__ both raise AttributeError, where CPython has a method-wrapper and the type’s docstring respectively.
  • A partial crossing the host boundary marshals as its repr. Python and JavaScript hosts receive MontyObject::repr("functools.partial(...)") rather than a callable, since neither side can call back into a value that only exists inside the sandbox.

Notes

partial is a descriptor, as it is in CPython 3.14: one stored as a class attribute binds the instance as the next argument after those it already carries.