Following system colour scheme Selected dark colour scheme Selected light colour scheme

Python Enhancement Proposals

PEP 568 – Generator-sensitivity for Context Variables

Nathaniel J. Smith <njs at>
Standards Track

Table of Contents


Context variables provide a generic mechanism for tracking dynamic, context-local state, similar to thread-local storage but generalized to cope work with other kinds of thread-like contexts, such as asyncio Tasks. PEP 550 proposed a mechanism for context-local state that was also sensitive to generator context, but this was pretty complicated, so the BDFL requested it be simplified. The result was PEP 567, which is targeted for inclusion in 3.7. This PEP then extends PEP 567’s machinery to add generator context sensitivity.

This PEP is starting out in the “deferred” status, because there isn’t enough time to give it proper consideration before the 3.7 feature freeze. The only goal right now is to understand what would be required to add generator context sensitivity in 3.8, so that we can avoid shipping something in 3.7 that would rule it out by accident. (Ruling it out on purpose can wait until 3.8 ;-).)


[Currently the point of this PEP is just to understand how this would work, with discussion of whether it’s a good idea deferred until after the 3.7 feature freeze. So rationale is TBD.]

High-level summary

Instead of holding a single Context, the threadstate now holds a ChainMap of Contexts. ContextVar.get and ContextVar.set are backed by the ChainMap. Generators and async generators each have an associated Context that they push onto the ChainMap while they’re running to isolate their context-local changes from their callers, though this can be overridden in cases like @contextlib.contextmanager where “leaking” context changes from the generator into its caller is desirable.


Review of PEP 567

Let’s start by reviewing how PEP 567 works, and then in the next section we’ll describe the differences.

In PEP 567, a Context is a Mapping from ContextVar objects to arbitrary values. In our pseudo-code here we’ll pretend that it uses a dict for backing storage. (The real implementation uses a HAMT, which is semantically equivalent to a dict but with different performance trade-offs.):

class Context(
    def __init__(self):
        self._data = {}
        self._in_use = False

    def __getitem__(self, key):
        return self._data[key]

    def __iter__(self):
        return iter(self._data)

    def __len__(self):
        return len(self._data)

At any given moment, the threadstate holds a current Context (initialized to an empty Context when the threadstate is created); we can use to temporarily switch the current Context:

def run(self, fn, *args, **kwargs):
    if self._in_use:
        raise RuntimeError("Context already in use")
    tstate = get_thread_state()
    old_context = tstate.current_context
    tstate.current_context = self
    self._in_use = True
        return fn(*args, **kwargs)
        state.current_context = old_context
        self._in_use = False

We can fetch a shallow copy of the current Context by calling copy_context; this is commonly used when spawning a new task, so that the child task can inherit context from its parent:

def copy_context():
    tstate = get_thread_state()
    new_context = Context()
    new_context._data = dict(tstate.current_context)
    return new_context

In practice, what end users generally work with is ContextVar objects, which also provide the only way to mutate a Context. They work with a utility class Token, which can be used to restore a ContextVar to its previous value:

class Token:
    MISSING = sentinel_value()

    # Note: constructor is private
    def __init__(self, context, var, old_value):
        self._context = context
        self.var = var
        self.old_value = old_value

    # XX: PEP 567 currently makes this a method on ContextVar, but
    # I'm going to propose it switch to this API because it's simpler.
    def reset(self):
        # XX: should we allow token reuse?
        # XX: should we allow tokens to be used if the saved
        # context is no longer active?
        if self.old_value is self.MISSING:
            del self._context._data[self.context_var]
            self._context._data[self.context_var] = self.old_value

# XX: the handling of defaults here uses the simplified proposal from
# This can be updated to whatever we settle on, it was just less
# typing this way :-)
class ContextVar:
    def __init__(self, name, *, default=None): = name
        self.default = default

    def get(self):
        context = get_thread_state().current_context
        return context.get(self, self.default)

    def set(self, new_value):
        context = get_thread_state().current_context
        token = Token(context, self, context.get(self, Token.MISSING))
        context._data[self] = new_value
        return token

Changes from PEP 567 to this PEP

In general, Context remains the same. However, now instead of holding a single Context object, the threadstate stores a stack of them. This stack acts just like a collections.ChainMap, so we’ll use that in our pseudocode. then becomes:

def run(self, fn, *args, **kwargs):
    if self._in_use:
        raise RuntimeError("Context already in use")
    tstate = get_thread_state()
    old_context_stack = tstate.current_context_stack
    tstate.current_context_stack = ChainMap([self])     # changed
    self._in_use = True
        return fn(*args, **kwargs)
        state.current_context_stack = old_context_stack
        self._in_use = False

Aside from some updated variables names (e.g., tstate.current_contexttstate.current_context_stack), the only change here is on the marked line, which now wraps the context in a ChainMap before stashing it in the threadstate.

We also add a Context.push method, which is almost exactly like, except that it temporarily pushes the Context onto the existing stack, instead of temporarily replacing the whole stack:

# Context.push
def push(self, fn, *args, **kwargs):
    if self._in_use:
        raise RuntimeError("Context already in use")
    tstate = get_thread_state()
    tstate.current_context_stack.maps.insert(0, self)  # different from run
    self._in_use = True
        return fn(*args, **kwargs)
        tstate.current_context_stack.maps.pop(0)       # different from run
        self._in_use = False

In most cases, we don’t expect push to be used directly; instead, it will be used implicitly by generators. Specifically, every generator object and async generator object gains a new attribute .context. When an (async) generator object is created, this attribute is initialized to an empty Context (self.context = Context()). This is a mutable attribute; it can be changed by user code. But trying to set it to anything that isn’t a Context object or None will raise an error.

Whenever we enter an generator via __next__, send, throw, or close, or enter an async generator by calling one of those methods on its __anext__, asend, athrow, or aclose coroutines, then its .context attribute is checked, and if non-None, is automatically pushed:

# GeneratorType.__next__
def __next__(self):
    if self.context is not None:
        return self.context.push(self.__real_next__)
        return self.__real_next__()

While we don’t expect people to use Context.push often, making it a public API preserves the principle that a generator can always be rewritten as an explicit iterator class with equivalent semantics.

Also, we modify contextlib.(async)contextmanager to always set its (async) generator objects’ .context attribute to None:

# contextlib._GeneratorContextManagerBase.__init__
def __init__(self, func, args, kwds):
    self.gen = func(*args, **kwds)
    self.gen.context = None                  # added

This makes sure that code like this continues to work as expected:

def decimal_precision(prec):
    with decimal.localcontext() as ctx:
        ctx.prec = prec

with decimal_precision(2):

The general idea here is that by default, every generator object gets its own local context, but if users want to explicitly get some other behavior then they can do that.

Otherwise, things mostly work as before, except that we go through and swap everything to use the threadstate ChainMap instead of the threadstate Context. In full detail:

The copy_context function now returns a flattened copy of the “effective” context. (As an optimization, the implementation might choose to do this flattening lazily, but if so this will be made invisible to the user.) Compared to our previous implementation above, the only change here is that tstate.current_context has been replaced with tstate.current_context_stack:

def copy_context() -> Context:
    tstate = get_thread_state()
    new_context = Context()
    new_context._data = dict(tstate.current_context_stack)
    return new_context

Token is unchanged, and the changes to ContextVar.get are trivial:

# ContextVar.get
def get(self):
    context_stack = get_thread_state().current_context_stack
    return context_stack.get(self, self.default)

ContextVar.set is a little more interesting: instead of going through the ChainMap machinery like everything else, it always mutates the top Context in the stack, and – crucially! – sets up the returned Token to restore its state later. This allows us to avoid accidentally “promoting” values between different levels in the stack, as would happen if we did old = var.get(); ...; var.set(old):

# ContextVar.set
def set(self, new_value):
    top_context = get_thread_state().current_context_stack.maps[0]
    token = Token(top_context, self, top_context.get(self, Token.MISSING))
    top_context._data[self] = new_value
    return token

And finally, to allow for introspection of the full context stack, we provide a new function contextvars.get_context_stack:

def get_context_stack() -> List[Context]:
    return list(get_thread_state().current_context_stack.maps)

That’s all.

Comparison to PEP 550

The main difference from PEP 550 is that it reified what we’re calling “contexts” and “context stacks” as two different concrete types (LocalContext and ExecutionContext respectively). This led to lots of confusion about what the differences were, and which object should be used in which places. This proposal simplifies things by only reifying the Context, which is “just a dict”, and makes the “context stack” an unnamed feature of the interpreter’s runtime state – though it is still possible to introspect it using get_context_stack, for debugging and other purposes.

Implementation notes

Context will continue to use a HAMT-based mapping structure under the hood instead of dict, since we expect that calls to copy_context are much more common than ContextVar.set. In almost all cases, copy_context will find that there’s only one Context in the stack (because it’s rare for generators to spawn new tasks), and can simply re-use it directly; in other cases HAMTs are cheap to merge and this can be done lazily.

Rather than using an actual ChainMap object, we’ll represent the context stack using some appropriate structure – the most appropriate options are probably either a bare list with the “top” of the stack being the end of the list so we can use push/pop, or else an intrusive linked list (PyThreadStateContextContext → …), with the “top” of the stack at the beginning of the list to allow efficient push/pop.

A critical optimization in PEP 567 is the caching of values inside ContextVar. Switching from a single context to a context stack makes this a little bit more complicated, but not too much. Currently, we invalidate the cache whenever the threadstate’s current Context changes (on thread switch, and when entering/exiting The simplest approach here would be to invalidate the cache whenever stack changes (on thread switch, when entering/exiting, and when entering/leaving Context.push). The main effect of this is that iterating a generator will invalidate the cache. It seems unlikely that this will cause serious problems, but if it does, then I think it can be avoided with a cleverer cache key that recognizes that pushing and then popping a Context returns the threadstate to its previous state. (Idea: store the cache key for a particular stack configuration in the topmost Context.)

It seems unavoidable in this design that uncached get will be O(n), where n is the size of the context stack. However, n will generally be very small – it’s roughly the number of nested generators, so usually n=1, and it will be extremely rare to see n greater than, say, 5. At worst, n is bounded by the recursion limit. In addition, we can expect that in most cases of deep generator recursion, most of the Contexts in the stack will be empty, and thus can be skipped extremely quickly during lookup. And for repeated lookups the caching mechanism will kick in. So it’s probably possible to construct some extreme case where this causes performance problems, but ordinary code should be essentially unaffected.


Last modified: 2023-09-09 17:39:29 GMT