Migrating from AnyIO 3 to AnyIO 4
The non-standard exception group class was removed
AnyIO 3 had its own ExceptionGroup
class which predated the PEP 654 exception
group classes. This class has now been removed in favor of the built-in
BaseExceptionGroup
and ExceptionGroup
classes. If your code was either
raising the old ExceptionGroup
exception or catching it, you need to make the switch
to these standard classes. Otherwise you can ignore this part.
If you’re targeting Python releases older than 3.11, you need to use the exceptiongroup
backport and import one of those classes from exceptiongroup
. The only difference
between BaseExceptionGroup
and ExceptionGroup
is that the latter can
only contain exceptions derived from Exception
, and likewise can be caught with
except Exception:
.
Task groups now wrap single exceptions in groups
The most prominent backwards incompatible change in AnyIO 4 was that task groups now
always raise exception groups when either the host task or any child tasks raise an
exception (other than a cancellation exception). Previously, an exception group was only
raised when more than one exception needed to be raised from the task group. The
practical consequence is that if your code previously expected to catch a specific kind
of exception falling out of a task group, you now need to either switch to the
except*
syntax (if you’re fortunate enough to work solely with Python 3.11 or
later), or use the catch()
context manager from the exceptiongroup backport.
So, if you had code like this:
try:
await function_using_a_taskgroup()
except ValueError as exc:
...
The Python 3.11+ equivalent would look almost the same:
try:
await function_using_a_taskgroup()
except* ValueError as excgrp:
# Note: excgrp is an ExceptionGroup now!
...
If you need to stay compatible with older Python releases, you need to use the backport:
from exceptiongroup import ExceptionGroup, catch
def handle_value_errors(excgrp: ExceptionGroup) -> None:
...
with catch({ValueError: handle_value_errors}):
await function_using_a_taskgroup()
This difference often comes up in test suites too. For example, if you had this before in a pytest-based test suite:
with pytest.raises(ValueError):
await function_using_a_taskgroup()
You now need to change it to:
from exceptiongroup import ExceptionGroup
with pytest.raises(ExceptionGroup) as exc:
await function_using_a_taskgroup()
assert len(exc.value.exceptions) == 1
assert isinstance(exc.value.exceptions[0], ValueError)
If you need to stay compatible with both AnyIO 3 and 4, you can use the following compatibility code to “collapse” single-exception groups by unwrapping them:
import sys
from contextlib import contextmanager
from typing import Generator
has_exceptiongroups = True
if sys.version_info < (3, 11):
try:
from exceptiongroup import BaseExceptionGroup
except ImportError:
has_exceptiongroups = False
@contextmanager
def collapse_excgroups() -> Generator[None, None, None]:
try:
yield
except BaseException as exc:
if has_exceptiongroups:
while isinstance(exc, BaseExceptionGroup) and len(exc.exceptions) == 1:
exc = exc.exceptions[0]
raise exc
Syntax for type annotated memory object streams has changed
Where previously, creating type annotated memory object streams worked by passing the desired type as the second argument:
send, receive = create_memory_object_stream(100, int)
In 4.0, create_memory_object_stream()
is a class
masquerading as a function, so you need to parametrize it:
send, receive = create_memory_object_stream[int](100)
If you didn’t parametrize your memory object streams before, then you don’t need to make any changes in this regard.
Event loop factories instead of event loop policies
If you’re using a custom asyncio event loop policy with run()
, you need to switch
to passing an event loop factory, that is, a callable that returns a new event loop.
Using uvloop as an example, code like the following:
anyio.run(main, backend_options={"event_loop_policy": uvloop.EventLoopPolicy()})
should be converted into:
anyio.run(main, backend_options={"loop_factory": uvloop.new_event_loop})
Make sure not to actually call the factory function!
Migrating from AnyIO 2 to AnyIO 3
AnyIO 3 changed some functions and methods in a way that needs some adaptation in your code. All deprecated functions and methods will be removed in AnyIO 4.
Asynchronous functions converted to synchronous
AnyIO 3 changed several previously asynchronous functions and methods into regular ones for two reasons:
to better serve use cases where synchronous callbacks are used by third party libraries
to better match the API of Trio
The following functions and methods were changed:
When migrating to AnyIO 3, simply remove the await
from each call to these.
Note
For backwards compatibility reasons, current_time()
,
current_effective_deadline()
and get_running_tasks()
return objects which
are awaitable versions of their original types (float
and list
,
respectively). These awaitable versions are subclasses of the original types so they
should behave as their originals, but if you absolutely need the pristine original
types, you can either use maybe_async
or float()
/ list()
on the returned
value as appropriate.
The following async context managers changed to regular context managers:
open_cancel_scope()
(now justCancelScope()
)
When migrating, just change async with
into a plain with
.
With the exception of
MemoryObjectReceiveStream.receive_nowait()
,
all of them can still be used like before – they will raise DeprecationWarning
when used this way on AnyIO 3, however.
If you’re writing a library that needs to be compatible with both major releases, you
will need to use the compatibility functions added in AnyIO 2.2: maybe_async()
and
maybe_async_cm()
. These will let you safely use functions/methods and context
managers (respectively) regardless of which major release is currently installed.
Example 1 – setting an event:
from anyio.abc import Event
from anyio import maybe_async
async def foo(event: Event):
await maybe_async(event.set())
...
Example 2 – opening a cancel scope:
from anyio import CancelScope, maybe_async_cm
async def foo():
async with maybe_async_cm(CancelScope()) as scope:
...
Starting tasks
The TaskGroup.spawn()
coroutine method has been deprecated in favor of the
synchronous method TaskGroup.start_soon()
(which mirrors start_soon()
in
Trio’s nurseries). If you’re fully migrating to AnyIO 3, simply switch to calling the
new method (and remove the await
).
If your code needs to work with both AnyIO 2 and 3, you can keep using
TaskGroup.spawn()
(until AnyIO 4) and suppress the deprecation warning:
import warnings
async def foo():
async with create_task_group() as tg:
with warnings.catch_warnings():
await tg.spawn(otherfunc)
Blocking portal changes
AnyIO now requires from_thread.start_blocking_portal()
to be used as a
context manager:
from anyio import sleep
from anyio.from_thread import start_blocking_portal
with start_blocking_portal() as portal:
portal.call(sleep, 1)
As with TaskGroup.spawn()
, the BlockingPortal.spawn_task()
method has also been
renamed to start_task_soon()
, so as to be consistent
with task groups.
The create_blocking_portal()
factory function was also deprecated in favor of
instantiating BlockingPortal
directly.
For code requiring cross compatibility, catching the deprecation warning (as above) should work.
Synchronization primitives
Synchronization primitive factories (create_event()
etc.) were deprecated in favor
of instantiating the classes directly. So convert code like this:
from anyio import create_event
async def main():
event = create_event()
into this:
from anyio import Event
async def main():
event = Event()
or, if you need to work with both AnyIO 2 and 3:
try:
from anyio import Event
create_event = Event
except ImportError:
from anyio import create_event
from anyio.abc import Event
async def foo() -> Event:
return create_event()
Threading functions moved
Threading functions were restructured to submodules, following the example of Trio:
current_default_worker_thread_limiter
→to_thread.current_default_thread_limiter()
(NOTE: the function was renamed too!)run_sync_in_worker_thread()
→to_thread.run_sync()
run_async_from_thread()
→from_thread.run()
run_sync_from_thread()
→from_thread.run_sync()
The old versions are still in place but emit deprecation warnings when called.