Non-goals

Just like in PEP 484 and PEP 526, it should be emphasized that Python will remain a dynamically typed language, and the authors have no desire to ever make type hints mandatory, even by convention.

This PEP is meant to solve the problem of forward references in type annotations. There are still cases outside of annotations where forward references will require usage of string literals. Those are listed in a later section of this document.

Annotations without forced evaluation enable opportunities to improve the syntax of type hints. This idea will require its own separate PEP and is not discussed further in this document.

Non-typing usage of annotations

While annotations are still available for arbitrary use besides type checking, it is worth mentioning that the design of this PEP, as well as its precursors (PEP 484 and PEP 526), is predominantly motivated by the type hinting use case.

In Python 3.8 PEP 484 will graduate from provisional status. Other enhancements to the Python programming language like PEP 544, PEP 557, or PEP 560, are already being built on this basis as they depend on type annotations and the typing module as defined by PEP 484. In fact, the reason PEP 484 is staying provisional in Python 3.7 is to enable rapid evolution for another release cycle that some of the aforementioned enhancements require.

With this in mind, uses for annotations incompatible with the aforementioned PEPs should be considered deprecated.

Enabling the future behavior in Python 3.7

The functionality described above can be enabled starting from Python 3.7 using the following special import:

from __future__ import annotations

A reference implementation of this functionality is available on GitHub.

Runtime annotation resolution and class decorators

Metaclasses and class decorators that need to resolve annotations for the current class will fail for annotations that use the name of the current class. Example:

def class_decorator(cls):
    annotations = get_type_hints(cls)  # raises NameError on 'C'
    print(f'Annotations for {cls}: {annotations}')
    return cls

@class_decorator
class C:
    singleton: 'C' = None

This was already true before this PEP. The class decorator acts on the class before it’s assigned a name in the current definition scope.

Runtime annotation resolution and TYPE_CHECKING

Sometimes there’s code that must be seen by a type checker but should not be executed. For such situations the typing module defines a constant, TYPE_CHECKING, that is considered True during type checking but False at runtime. Example:

import typing

if typing.TYPE_CHECKING:
    import expensive_mod

def a_func(arg: expensive_mod.SomeClass) -> None:
    a_var: expensive_mod.SomeClass = arg
    ...

This approach is also useful when handling import cycles.

Trying to resolve annotations of a_func at runtime using typing.get_type_hints() will fail since the name expensive_mod is not defined (TYPE_CHECKING variable being False at runtime). This was already true before this PEP.

Deprecation policy

Starting with Python 3.7, a __future__ import is required to use the described functionality. No warnings are raised.

NOTE: Whether this will eventually become the default behavior is currently unclear pending decision on PEP 649. In any case, use of annotations that depend upon their eager evaluation is incompatible with both proposals and is no longer supported.

Keeping the ability to use function local state when defining annotations

With postponed evaluation, this would require keeping a reference to the frame in which an annotation got created. This could be achieved for example by storing all annotations as lambdas instead of strings.

This would be prohibitively expensive for highly annotated code as the frames would keep all their objects alive. That includes predominantly objects that won’t ever be accessed again.

To be able to address class-level scope, the lambda approach would require a new kind of cell in the interpreter. This would proliferate the number of types that can appear in __annotations__, as well as wouldn’t be as introspectable as strings.

Note that in the case of nested classes, the functionality to get the effective “globals” and “locals” at definition time is provided by typing.get_type_hints().

If a function generates a class or a function with annotations that have to use local variables, it can populate the given generated object’s __annotations__ dictionary directly, without relying on the compiler.

Disallowing local state usage for classes, too

This PEP originally proposed limiting names within annotations to only allow names from the model-level scope, including for classes. The author argued this makes name resolution unambiguous, including in cases of conflicts between local names and module-level names.

This idea was ultimately rejected in case of classes. Instead, typing.get_type_hints() got modified to populate the local namespace correctly if class-level annotations are needed.

The reasons for rejecting the idea were that it goes against the intuition of how scoping works in Python, and would break enough existing type annotations to make the transition cumbersome. Finally, local scope access is required for class decorators to be able to evaluate type annotations. This is because class decorators are applied before the class receives its name in the outer scope.

Introducing a new dictionary for the string literal form instead

Yury Selivanov shared the following idea:

1. Add a new special attribute to functions: __annotations_text__.
2. Make __annotations__ a lazy dynamic mapping, evaluating expressions from the corresponding key in __annotations_text__ just-in-time.

This idea is supposed to solve the backwards compatibility issue, removing the need for a new __future__ import. Sadly, this is not enough. Postponed evaluation changes which state the annotation has access to. While postponed evaluation fixes the forward reference problem, it also makes it impossible to access function-level locals anymore. This alone is a source of backwards incompatibility which justifies a deprecation period.

A __future__ import is an obvious and explicit indicator of opting in for the new functionality. It also makes it trivial for external tools to recognize the difference between a Python files using the old or the new approach. In the former case, that tool would recognize that local state access is allowed, whereas in the latter case it would recognize that forward references are allowed.

Finally, just-in-time evaluation in __annotations__ is an unnecessary step if get_type_hints() is used later.

Dropping annotations with -O

There are two reasons this is not satisfying for the purpose of this PEP.

First, this only addresses runtime cost, not forward references, those still cannot be safely used in source code. A library maintainer would never be able to use forward references since that would force the library users to use this new hypothetical -O switch.

Second, this throws the baby out with the bath water. Now no runtime annotation use can be performed. PEP 557 is one example of a recent development where evaluating type annotations at runtime is useful.

All that being said, a granular -O option to drop annotations is a possibility in the future, as it’s conceptually compatible with existing -O behavior (dropping docstrings and assert statements). This PEP does not invalidate the idea.

Passing string literals in annotations verbatim to __annotations__

This PEP originally suggested directly storing the contents of a string literal under its respective key in __annotations__. This was meant to simplify support for runtime type checkers.

Mark Shannon pointed out this idea was flawed since it wasn’t handling situations where strings are only part of a type annotation.

The inconsistency of it was always apparent but given that it doesn’t fully prevent cases of double-wrapping strings anyway, it is not worth it.

Making the name of the future import more verbose

Instead of requiring the following import:

from __future__ import annotations

the PEP could call the feature more explicitly, for example string_annotations, stringify_annotations, annotation_strings, annotations_as_strings, lazy_annotations, static_annotations, etc.

The problem with those names is that they are very verbose. Each of them besides lazy_annotations would constitute the longest future feature name in Python. They are long to type and harder to remember than the single-word form.

There is precedence of a future import name that sounds overly generic but in practice was obvious to users as to what it does:

from __future__ import division

In PEP 484

The forward reference problem was discussed when PEP 484 was originally drafted, leading to the following statement in the document:

A compromise is possible where a __future__ import could enable turning all annotations in a given module into string literals, as follows:

from __future__ import annotations

class ImSet:
    def add(self, a: ImSet) -> List[ImSet]: ...

assert ImSet.add.__annotations__ == {
    'a': 'ImSet', 'return': 'List[ImSet]'
}

Such a __future__ import statement may be proposed in a separate PEP.

python/typing#400

The problem was discussed at length on the typing module’s GitHub project, under Issue 400. The problem statement there includes critique of generic types requiring imports from typing. This tends to be confusing to beginners:

Why this:

from typing import List, Set
def dir(o: object = ...) -> List[str]: ...
def add_friends(friends: Set[Friend]) -> None: ...

But not this:

def dir(o: object = ...) -> list[str]: ...
def add_friends(friends: set[Friend]) -> None ...

Why this:

up_to_ten = list(range(10))
friends = set()

But not this:

from typing import List, Set
up_to_ten = List[int](range(10))
friends = Set[Friend]()

While typing usability is an interesting problem, it is out of scope of this PEP. Specifically, any extensions of the typing syntax standardized in PEP 484 will require their own respective PEPs and approval.

Issue 400 ultimately suggests postponing evaluation of annotations and keeping them as strings in __annotations__, just like this PEP specifies. This idea was received well. Ivan Levkivskyi supported using the __future__ import and suggested unparsing the AST in compile.c. Jukka Lehtosalo pointed out that there are some cases of forward references where types are used outside of annotations and postponed evaluation will not help those. For those cases using the string literal notation would still be required. Those cases are discussed briefly in the “Forward References” section of this PEP.

The biggest controversy on the issue was Guido van Rossum’s concern that untokenizing annotation expressions back to their string form has no precedent in the Python programming language and feels like a hacky workaround. He said:

One thing that comes to mind is that it’s a very random change to the language. It might be useful to have a more compact way to indicate deferred execution of expressions (using less syntax than lambda:). But why would the use case of type annotations be so all-important to change the language to do it there first (rather than proposing a more general solution), given that there’s already a solution for this particular use case that requires very minimal syntax?

Eventually, Ethan Smith and schollii voiced that feedback gathered during PyCon US suggests that the state of forward references needs fixing. Guido van Rossum suggested coming back to the __future__ idea, pointing out that to prevent abuse, it’s important for the annotations to be kept both syntactically valid and evaluating correctly at runtime.

First draft discussion on python-ideas

Discussion happened largely in two threads, the original announcement and a follow-up called PEP 563 and expensive backwards compatibility.

The PEP received rather warm feedback (4 strongly in favor, 2 in favor with concerns, 2 against). The biggest voice of concern on the former thread being Steven D’Aprano’s review stating that the problem definition of the PEP doesn’t justify breaking backwards compatibility. In this response Steven seemed mostly concerned about Python no longer supporting evaluation of annotations that depended on local function/class state.

A few people voiced concerns that there are libraries using annotations for non-typing purposes. However, none of the named libraries would be invalidated by this PEP. They do require adapting to the new requirement to call eval() on the annotation with the correct globals and locals set.

This detail about globals and locals having to be correct was picked up by a number of commenters. Nick Coghlan benchmarked turning annotations into lambdas instead of strings, sadly this proved to be much slower at runtime than the current situation.

The latter thread was started by Jim J. Jewett who stressed that the ability to properly evaluate annotations is an important requirement and backwards compatibility in that regard is valuable. After some discussion he admitted that side effects in annotations are a code smell and modal support to either perform or not perform evaluation is a messy solution. His biggest concern remained loss of functionality stemming from the evaluation restrictions on global and local scope.

Nick Coghlan pointed out that some of those evaluation restrictions from the PEP could be lifted by a clever implementation of an evaluation helper, which could solve self-referencing classes even in the form of a class decorator. He suggested the PEP should provide this helper function in the standard library.

Second draft discussion on python-dev

Discussion happened mainly in the announcement thread, followed by a brief discussion under Mark Shannon’s post.

Steven D’Aprano was concerned whether it’s acceptable for typos to be allowed in annotations after the change proposed by the PEP. Brett Cannon responded that type checkers and other static analyzers (like linters or programming text editors) will catch this type of error. Jukka Lehtosalo added that this situation is analogous to how names in function bodies are not resolved until the function is called.

A major topic of discussion was Nick Coghlan’s suggestion to store annotations in “thunk form”, in other words as a specialized lambda which would be able to access class-level scope (and allow for scope customization at call time). He presented a possible design for it (indirect attribute cells). This was later seen as equivalent to “special forms” in Lisp. Guido van Rossum expressed worry that this sort of feature cannot be safely implemented in twelve weeks (i.e. in time before the Python 3.7 beta freeze).

After a while it became clear that the point of division between supporters of the string form vs. supporters of the thunk form is actually about whether annotations should be perceived as a general syntactic element vs. something tied to the type checking use case.

Finally, Guido van Rossum declared he’s rejecting the thunk idea based on the fact that it would require a new building block in the interpreter. This block would be exposed in annotations, multiplying possible types of values stored in __annotations__ (arbitrary objects, strings, and now thunks). Moreover, thunks aren’t as introspectable as strings. Most importantly, Guido van Rossum explicitly stated interest in gradually restricting the use of annotations to static typing (with an optional runtime component).

Nick Coghlan got convinced to PEP 563, too, promptly beginning the mandatory bike shedding session on the name of the __future__ import. Many debaters agreed that annotations seems like an overly broad name for the feature name. Guido van Rossum briefly decided to call it string_annotations but then changed his mind, arguing that division is a precedent of a broad name with a clear meaning.

The final improvement to the PEP suggested in the discussion by Mark Shannon was the rejection of the temptation to pass string literals through to __annotations__ verbatim.

A side-thread of discussion started around the runtime penalty of static typing, with topic like the import time of the typing module (which is comparable to re without dependencies, and three times as heavy as re when counting dependencies).