Thank you for the feedback so far. FYI, or as a reminder, this is now PEP 555, but the web version is still the same draft that I posted here.
The discussion of this was paused as there was a lot going on at that moment, but I'm now getting ready to make a next version of the draft. Below, I'll draft some changes I intend to make so they can already be discussed.
First of all, I'm considering calling the concept "context arguments" instead of "context variables", because that describes the concept better. But see below for some more.
On Tue, Sep 5, 2017 at 12:50 AM, Koos Zevenhoven email@example.com wrote:
as promised, here is a draft PEP for context variable semantics and implementation. Apologies for the slight delay; I had a not-so-minor autosave accident and had to retype the majority of this first draft.
During the past years, there has been growing interest in something like task-local storage or async-local storage. This PEP proposes an alternative approach to solving the problems that are typically stated as motivation for such concepts.
This proposal is based on sketches of solutions since spring 2015, with some minor influences from the recent discussion related to PEP 550. I can also see some potential implementation synergy between this PEP and PEP 550, even if the proposed semantics are quite different.
So, here it is. This is the first draft and some things are still missing, but the essential things should be there.
PEP: 999 Title: Context-local variables (contextvars) Version: $Revision$ Last-Modified: $Date$ Author: Koos Zevenhoven Status: Draft Type: Standards Track Content-Type: text/x-rst Created: DD-Mmm-YYYY Post-History: DD-Mmm-YYYY
Sometimes, in special cases, it is desired that code can pass information down the function call chain to the callees without having to explicitly pass the information as arguments to each function in the call chain. This proposal describes a construct which allows code to explicitly switch in and out of a context where a certain context variable has a given value assigned to it. This is a modern alternative to some uses of things like global variables in traditional single-threaded (or thread-unsafe) code and of thread-local storage in traditional concurrency-unsafe code (single- or multi-threaded). In particular, the proposed mechanism can also be used with more modern concurrent execution mechanisms such as asynchronously executed coroutines, without the concurrently executed call chains interfering with each other's contexts.
The "call chain" can consist of normal functions, awaited coroutines, or generators. The semantics of context variable scope are equivalent in all cases, allowing code to be refactored freely into subroutines (which here refers to functions, sub-generators or sub-coroutines) without affecting the semantics of context variables. Regarding implementation, this proposal aims at simplicity and minimum changes to the CPython interpreter and to other Python interpreters.
Consider a modern Python call chain (or call tree), which in this
proposal refers to any chained (nested) execution of subroutines, using
any possible combinations of normal function calls, or expressions using
yield from. In some cases, passing necessary
down the call chain as arguments can substantially complicate the required
function signatures, or it can even be impossible to achieve in practice.
In these cases, one may search for another place to store this information.
Let us look at some historical examples.
The most naive option is to assign the value to a global variable or similar, where the code down the call chain can access it. However, this immediately makes the code thread-unsafe, because with multiple threads, all threads assign to the same global variable, and another thread can interfere at any point in the call chain.
A somewhat less naive option is to store the information as per-thread information in thread-local storage, where each thread has its own "copy" of the variable which other threads cannot interfere with. Although non-ideal, this has been the best solution in many cases. However, thanks to generators and coroutines, the execution of the call chain can be suspended and resumed, allowing code in other contexts to run concurrently. Therefore, using thread-local storage is concurrency-unsafe, because other call chains in other contexts may interfere with the thread-local variable.
Note that in the above two historical approaches, the stored information
has the widest available scope without causing problems. For a third
solution along the same path, one would first define an equivalent of a
"thread" for asynchronous execution and concurrency. This could be seen as
the largest amount of code and nested calls that is guaranteed to be
executed sequentially without ambiguity in execution order. This might be
referred to as concurrency-local or task-local storage. In this meaning of
"task", there is no ambiguity in the order of execution of the code within
one task. (This concept of a task is close to equivalent to a
asyncio, but not exactly.) In such concurrency-locals, it is possible
to pass information down the call chain to callees without another code
path interfering with the value in the background.
Common to the above approaches is that they indeed use variables with a wide but just-narrow-enough scope. Thread-locals could also be called thread-wide globals---in single-threaded code, they are indeed truly global. And task-locals could be called task-wide globals, because tasks can be very big.
The issue here is that neither global variables, thread-locals nor task-locals are really meant to be used for this purpose of passing information of the execution context down the call chain. Instead of the widest possible variable scope, the scope of the variables should be controlled by the programmer, typically of a library, to have the desired scope---not wider. In other words, task-local variables (and globals and thread-locals) have nothing to do with the kind of context-bound information passing that this proposal intends to enable, even if task-locals can be used to emulate the desired semantics. Therefore, in the following, this proposal describes the semantics and the outlines of an implementation for context-local variables (or context variables, contextvars). In fact, as a side effect of this PEP, an async framework can use the proposed feature to implement task-local variables.
Because the proposed semantics are not a direct extension to anything
already available in Python, this proposal is first described in terms of
semantics and API at a fairly high level. In particular, Python
statements are heavily used in the description, as they are a good match
with the proposed semantics. However, the underlying
__exit__ methods correspond to functions in the lower-level
speed-optimized (C) API. For clarity of this document, the lower-level
functions are not explicitly named in the definition of the semantics.
After describing the semantics and high-level API, the implementation is
described, going to a lower level.
Core concept ''''''''''''
A context-local variable is represented by a single instance of
cvar. Any code that has access to the
object can ask for its value with respect to the current context. In the
high-level API, this value is given by the
cvar = contextvars.Var(default="the default value", description="example context variable")
Some points related to the arguments and naming:
Indeed, this might change to contextvars.Arg. After all, these are more like arguments than variables.
But just like with function arguments, you can use a mutable value, which allows more variable-like semantics. That is, however, not the primarily intended use. It may also cause more problems at inter-process or inter-interpreter boundaries etc., where direct mutation of objects may not be possible.
I might have to remove the
default argument, at least in this form. If
there is a default, it should be more explicit what the scope of the
default is. There could be thread-wide defaults or interpreter-wide
defaults and so on. It is not completely clear what a truly global default
One way to deal with this would be to always pass the context on to other threads and processes etc when they are created. But there are some ambiguities here too, so the safest way might be to let the user implement the desired semantics regarding defaults and thread boundaries etc.
assert cvar.value == "the default value" # default still applies # In code examples, all ``assert`` statements should # succeed according to the proposed semantics.
No assignments to
cvar have been applied for this context, so
cvar.value gives the default value. Assigning new values to contextvars
is done in a highly scope-aware manner::
with cvar.assign(new_value): assert cvar.value is new_value # Any code here, or down the call chain from here, sees: # cvar.value is new_value # unless another value has been assigned in a # nested context assert cvar.value is new_value # the assignment of ``cvar`` to ``new_value`` is no longer visible assert cvar.value == "the default value"
cvar.assign(value) returns another object, namely
contextvars.Assignment(cvar, new_value). The essential part here is
that applying a context variable assignment (
paired with a de-assignment (
Assignment.__exit__). These operations set
the bounds for the scope of the assigned value.
Assignments to the same context variable can be nested to override the outer assignment in a narrower context::
assert cvar.value == "the default value" with cvar.assign("outer"): assert cvar.value == "outer" with cvar.assign("inner"): assert cvar.value == "inner" assert cvar.value == "outer" assert cvar.value == "the default value"
Also multiple variables can be assigned to in a nested manner without affecting each other::
cvar1 = contextvars.Var() cvar2 = contextvars.Var() assert cvar1.value is None # default is None by default assert cvar2.value is None with cvar1.assign(value1): assert cvar1.value is value1 assert cvar2.value is None with cvar2.assign(value2): assert cvar1.value is value1 assert cvar2.value is value2 assert cvar1.value is value1 assert cvar2.value is None assert cvar1.value is None assert cvar2.value is None
Or with more convenient Python syntax::
with cvar1.assign(value1), cvar2.assign(value2): assert cvar1.value is value1 assert cvar2.value is value2
In another context, in another thread or otherwise concurrently executed task or code path, the context variables can have a completely different state. The programmer thus only needs to worry about the context at hand.
Refactoring into subroutines ''''''''''''''''''''''''''''
Code using contextvars can be refactored into subroutines without affecting the semantics. For instance::
assi = cvar.assign(new_value) def apply(): assi.__enter__() assert cvar.value == "the default value" apply() assert cvar.value is new_value assi.__exit__() assert cvar.value == "the default value"
Or similarly in an asynchronous context where
await expressions are
used. The subroutine can now be a coroutine::
assi = cvar.assign(new_value) async def apply(): assi.__enter__() assert cvar.value == "the default value" await apply() assert cvar.value is new_value assi.__exit__() assert cvar.value == "the default value"
Or when the subroutine is a generator::
def apply(): yield assi.__enter__()
which is called using
yield from apply() or with calls to
.send. This is discussed further in later sections.
Semantics for generators and generator-based coroutines '''''''''''''''''''''''''''''''''''''''''''''''''''''''
Generators, coroutines and async generators act as subroutines in much the
same way that normal functions do. However, they have the additional
possibility of being suspended by
yield expressions. Assignment
contexts entered inside a generator are normally preserved across yields::
def genfunc(): with cvar.assign(new_value): assert cvar.value is new_value yield assert cvar.value is new_value g = genfunc() next(g) assert cvar.value == "the default value" with cvar.assign(another_value): next(g)
However, the outer context visible to the generator may change state across yields::
def genfunc(): assert cvar.value is value2 yield assert cvar.value is value1 yield with cvar.assign(value3): assert cvar.value is value3 with cvar.assign(value1): g = genfunc() with cvar.assign(value2): next(g) next(g) next(g) assert cvar.value is value1
Similar semantics apply to async generators defined by
async def ...
yield ... ).
By default, values assigned inside a generator do not leak through yields
to the code that drives the generator. However, the assignment contexts
entered and left open inside the generator do become visible outside the
generator after the generator has finished with a
assi = cvar.assign(new_value) def genfunc(): yield assi.__enter__(): yield g = genfunc() assert cvar.value == "the default value" next(g) assert cvar.value == "the default value" next(g) # assi.__enter__() is called here assert cvar.value == "the default value" next(g) assert cvar.value is new_value assi.__exit__()
Frameworks, such as
asyncio or third-party libraries, can use
additional functionality in
contextvars to achieve the desired
semantics in cases which are not determined by the Python interpreter. Some
of the semantics described in this section are also afterwards used to
describe the internal implementation.
Leaking yields ''''''''''''''
contextvars.leaking_yields decorator, one can choose to leak
the context through
yield expressions into the outer context that
drives the generator::
@contextvars.leaking_yields def genfunc(): assert cvar.value == "outer" with cvar.assign("inner"): yield assert cvar.value == "inner" assert cvar.value == "outer" g = genfunc(): with cvar.assign("outer"): assert cvar.value == "outer" next(g) assert cvar.value == "inner" next(g) assert cvar.value == "outer"
Unfortunately, we actually need a third kind of generator semantics, something like this:
@contextvars.caller_context def genfunc(): assert cvar.value is the_value yield assert cvar.value is the_value
with cvar.assign(the_value): gen = genfunc()
with cvar.assign(1234567890): try: next(gen) except StopIteration: pass
Nick, Yury and I (and Nathaniel, Guido, Jim, ...?) somehow just narrowly missed the reasons for this in discussions related to PEP 550. Perhaps because we had mostly been looking at it from an async angle.
[In addition to this, all context changes (Assignment __enter__ or __exit__) would be leaked out when the generator finishes iff there are no outer context changes. If there are outer context changes, an attempt to leak changes will fail. (I will probably need to explain this better).]
Capturing contextvar assignments ''''''''''''''''''''''''''''''''
contextvars.capture(), one can capture the assignment contexts
that are entered by a block of code. The changes applied by the block of
code can then be reverted and subsequently reapplied, even in another
assert cvar1.value is None # default assert cvar2.value is None # default assi1 = cvar1.assign(value1) assi2 = cvar1.assign(value2) with contextvars.capture() as delta: assi1.__enter__() with cvar2.assign("not captured"): assert cvar2.value is "not captured" assi2.__enter__() assert cvar1.value is value2 delta.revert() assert cvar1.value is None assert cvar2.value is None ... with cvar1.assign(1), cvar2.assign(2): delta.reapply() assert cvar1.value is value2 assert cvar2.value == 2
However, reapplying the "delta" if its net contents include deassignments may not be possible (see also Implementation and Open Issues).
Getting a snapshot of context state '''''''''''''''''''''''''''''''''''
contextvars.get_local_state() returns an object
representing the applied assignments to all context-local variables in the
context where the function is called. This can be seen as equivalent to
contextvars.capture() to capture all context changes from the
beginning of execution. The returned object supports methods
reapply() as above.
We will probably need also a
use() method (or another name) here. That
would return a context manager that applies the full context on __enter__
and reapplies the previous one on __exit__.
Running code in a clean state '''''''''''''''''''''''''''''
Although it is possible to revert all applied context changes using the above primitives, a more convenient way to run a block of code in a clean context is provided::
with context_vars.clean_context(): # here, all context vars start off with their default values # here, the state is back to what it was before the with block.
As an additional tool, there could be contextvars.callback:
@contextvars.callback def some_callback():
# do stuff
This would provide some of the functionality of this PEP if callbacks are used, so that the callback would be run with the same context as the code that creates the callback.
The implementation of this would be essentially:
def callback(func): context = contextvars.get_local_context(): def wrapped(args, **kwargs): with context.use(): func(args, **kwargs return wrapped
With some trickery this might allow an async framework based on callbacks instead of coroutines to use context arguments. But using this might be a bit awkward sometimes. A contextlib.ExitStack might help here.
This section describes to a variable level of detail how the described semantics can be implemented. At present, an implementation aimed at simplicity but sufficient features is described. More details will be added later.
Alternatively, a somewhat more complicated implementation offers minor additional features while adding some performance overhead and requiring more code in the implementation.
Data structures and implementation of the core concept ''''''''''''''''''''''''''''''''''''''''''''''''''''''
Each thread of the Python interpreter keeps its on stack of
contextvars.Assignment objects, each having a pointer to the previous
(outer) assignment like in a linked list. The local state (also returned by
contextvars.get_local_state()) then consists of a reference to the
top of the stack and a pointer/weak reference to the bottom of the stack.
This allows efficient stack manipulations. An object produced by
contextvars.capture() is similar, but refers to only a part of the
stack with the bottom reference pointing to the top of the stack as it was
in the beginning of the capture block.
Now, the stack evolves according to the assignment
__exit__ methods. For example::
cvar1 = contextvars.Var() cvar2 = contextvars.Var() # stack:  assert cvar1.value is None assert cvar2.value is None with cvar1.assign("outer"): # stack: [Assignment(cvar1, "outer")] assert cvar1.value == "outer" with cvar1.assign("inner"): # stack: [Assignment(cvar1, "outer"), # Assignment(cvar1, "inner")] assert cvar1.value == "inner" with cvar2.assign("hello"): # stack: [Assignment(cvar1, "outer"), # Assignment(cvar1, "inner"), # Assignment(cvar2, "hello")] assert cvar2.value == "hello" # stack: [Assignment(cvar1, "outer"), # Assignment(cvar1, "inner")] assert cvar1.value == "inner" assert cvar2.value is None # stack: [Assignment(cvar1, "outer")] assert cvar1.value == "outer" # stack:  assert cvar1.value is None assert cvar2.value is None
Getting a value from the context using
cvar1.value can be implemented
as finding the topmost occurrence of a
cvar1 assignment on the stack
and returning the value there, or the default value if no assignment is
found on the stack. However, this can be optimized to instead be an O(1)
operation in most cases. Still, even searching through the stack may be
reasonably fast since these stacks are not intended to grow very large.
I will still need to explain the O(1) algorithm, but one nice thing is that an implementation like micropython does not necessarily need to include that optimization.
The above description is already sufficient for implementing the core concept. Suspendable frames require some additional attention, as explained in the following.
Implementation of generator and coroutine semantics '''''''''''''''''''''''''''''''''''''''''''''''''''
Within generators, coroutines and async generators, assignments and
deassignments are handled in exactly the same way as anywhere else.
However, some changes are needed in the builtin generator methods
close. Here is the Python
equivalent of the
changes needed in
send for a generator (here
the behavior in Python 3.6)::
def send(self, value): # if decorated with contextvars.leaking_yields if self.gi_contextvars is LEAK: # nothing needs to be done to leak context through yields :) return self._old_send(value) try: with contextvars.capture() as delta: if self.gi_contextvars: # non-zero captured content from previous iteration self.gi_contextvars.reapply() ret = self._old_send(value) except Exception: raise else: # suspending, revert context changes but delta.revert() self.gi_contextvars = delta return ret
The corresponding modifications to the other methods is essentially identical. The same applies to coroutines and async generators.
For code that does not use
contextvars, the additions are O(1) and
essentially reduce to a couple of pointer comparisons. For code that does
contextvars, the additions are still O(1) in most cases.
More on implementation ''''''''''''''''''''''
The rest of the functionality, including
are in fact quite straightforward to
implement, but their implementation will be discussed further in later
versions of this proposal. Caching of assigned values is somewhat more
complicated, and will be discussed later, but it seems that most cases
should achieve O(1) complexity.
There are no direct backwards-compatibility concerns, since a completely new feature is proposed.
However, various traditional uses of thread-local storage may need a
smooth transition to
contextvars so they can be concurrency-safe. There
are several approaches to this, including emulating task-local storage with
a little bit of help from async frameworks. A fully general implementation
cannot be provided, because the desired semantics may depend on the design
of the framework.
I have a preliminary design for this, but probably doesn't need to be in this PEP.
Another way to deal with the transition is for code to first look for a
context created using
contextvars. If that fails because a new-style
context has not been set or because the code runs on an older Python
version, a fallback to thread-local storage is used.
If context variables are renamed context arguments, then there could be a settable variant called a context variable (could also be a third-party thing on top of context arguments, depending on what is done with decimal contexts).
In this proposal, all variable deassignments are made in the opposite
order compared to the preceding assignments. This has two useful
properties: it encourages using
with statements to define assignment
scope and has a tendency to catch errors early (forgetting a
.__exit__() call often results in a meaningful error. To have this as a
requirement requirement is beneficial also in terms of implementation
simplicity and performance. Nevertheless, allowing out-of-order context
exits is not completely out of the question, and reasonable implementation
strategies for that do exist.
The scope of value visibility should not be determined by the way the code is refactored into subroutines. It is necessary to have per-variable control of the assignment scope.
In fact, in early sketches, my approach was closer to this. The context
variables (or async variables) were stored in frame locals in a namespace
__async__ and they were propagated through subroutine calls to
callees. But this introduces problems when new scope layers are added,
and ended up being more complicated (and slightly similar to PEP 550).
Anyway, for starters, this was a glimpse of the changes I have planned, and open for discussion.