Anyio Versions Save

• Added support for the Python 3.12 walk_up keyword argument in anyio.Path.relative_to() (PR by Colin Taylor)

• Fixed passing total_tokens to anyio.CapacityLimiter() as a keyword argument not working on the trio backend (#515)

• Fixed Process.aclose() not performing the minimum level of necessary cleanup when cancelled. Previously:

  • Cancellation of Process.aclose() could leak an orphan process
  • Cancellation of run_process() could very briefly leak an orphan process.
  • Cancellation of Process.aclose() or run_process() on Trio could leave standard streams unclosed

  (PR by Ganden Schaffner)

• Fixed Process.stdin.aclose(), Process.stdout.aclose(), and Process.stderr.aclose() not including a checkpoint on asyncio (PR by Ganden Schaffner)

• Fixed documentation on how to provide your own typed attributes

• Add support for byte-based paths in connect_unix, create_unix_listeners, create_unix_datagram_socket, and create_connected_unix_datagram_socket. (PR by Lura Skye)

• Enabled the Event and CapacityLimiter classes to be instantiated outside an event loop thread

• Broadly improved/fixed the type annotations. Among other things, many functions and methods that take variadic positional arguments now make use of PEP 646 TypeVarTuple to allow the positional arguments to be validated by static type checkers. These changes affected numerous methods and functions, including:

  • anyio.run()
  • TaskGroup.start_soon()
  • anyio.from_thread.run()
  • anyio.from_thread.run_sync()
  • anyio.to_thread.run_sync()
  • anyio.to_process.run_sync()
  • BlockingPortal.call()
  • BlockingPortal.start_task_soon()
  • BlockingPortal.start_task()

  (also resolves #560)

• Fixed various type annotations of anyio.Path to match Typeshed:

  • anyio.Path.__lt__()
  • anyio.Path.__le__()
  • anyio.Path.__gt__()
  • anyio.Path.__ge__()
  • anyio.Path.__truediv__()
  • anyio.Path.__rtruediv__()
  • anyio.Path.hardlink_to()
  • anyio.Path.samefile()
  • anyio.Path.symlink_to()
  • anyio.Path.with_segments()

  (PR by Ganden Schaffner)

• Fixed adjusting the total number of tokens in a CapacityLimiter on asyncio failing to wake up tasks waiting to acquire the limiter in certain edge cases (fixed with help from Egor Blagov)

• Fixed loop_factory and use_uvloop options not being used on the asyncio backend (#643)

• Fixed cancellation propagating on asyncio from a task group to child tasks if the task hosting the task group is in a shielded cancel scope (#642)

• Adapted to API changes made in Trio v0.23:
  • Call trio.to_thread.run_sync() using the abandon_on_cancel keyword argument instead of cancellable
  • Removed a checkpoint when exiting a task group
  • Renamed the cancellable argument in anyio.to_thread.run_sync() to abandon_on_cancel (and deprecated the old parameter name)
  • Bumped minimum version of Trio to v0.23
• Added support for voluntary thread cancellation via anyio.from_thread.check_cancelled()
• Bumped minimum version of trio to v0.23
• Exposed the ResourceGuard class in the public API (#627)
• Fixed RuntimeError: Runner is closed when running higher-scoped async generator fixtures in some cases (#619)
• Fixed discrepancy between asyncio and trio where reraising a cancellation exception in an except* block would incorrectly bubble out of its cancel scope (#634)

• BACKWARDS INCOMPATIBLE Replaced AnyIO's own ExceptionGroup class with the PEP 654 BaseExceptionGroup and ExceptionGroup
• BACKWARDS INCOMPATIBLE Changes to cancellation semantics:
  • Any exceptions raising out of a task groups are now nested inside an ExceptionGroup (or BaseExceptionGroup if one or more BaseException were included)
  • Fixed task group not raising a cancellation exception on asyncio at exit if no child tasks were spawned and an outer cancellation scope had been cancelled before
  • Ensured that exiting a TaskGroup always hits a yield point, regardless of whether there are running child tasks to be waited on
  • On asyncio, cancel scopes will defer cancelling tasks that are scheduled to resume with a finished future
  • On asyncio and Python 3.9/3.10, cancel scopes now only suppress cancellation exceptions if the cancel message matches the scope
  • Task groups on all backends now raise a single cancellation exception when an outer cancel scope is cancelled, and no exceptions other than cancellation exceptions are raised in the group
• BACKWARDS INCOMPATIBLE Changes the pytest plugin to run all tests and fixtures in the same task, allowing fixtures to set context variables for tests and other fixtures
• BACKWARDS INCOMPATIBLE Changed anyio.Path.relative_to() and anyio.Path.is_relative_to() to only accept one argument, as passing multiple arguments is deprecated as of Python 3.12
• BACKWARDS INCOMPATIBLE Dropped support for spawning tasks from old-style coroutine functions (@asyncio.coroutine)
• BACKWARDS INCOMPATIBLE The policy option on the asyncio backend was changed to loop_factory to accommodate asyncio.Runner
• Changed anyio.run() to use asyncio.Runner (or a back-ported version of it on Pythons older than 3.11) on the asyncio backend
• Dropped support for Python 3.7
• Added support for Python 3.12
• Bumped minimum version of trio to v0.22
• Added the anyio.Path.is_junction() and anyio.Path.walk() methods
• Added create_unix_datagram_socket and create_connected_unix_datagram_socket to create UNIX datagram sockets (PR by Jean Hominal)
• Fixed from_thread.run and from_thread.run_sync not setting sniffio on asyncio. As a result:
  • Fixed from_thread.run_sync failing when used to call sniffio-dependent functions on asyncio
  • Fixed from_thread.run failing when used to call sniffio-dependent functions on asyncio from a thread running trio or curio
  • Fixed deadlock when using from_thread.start_blocking_portal(backend="asyncio") in a thread running trio or curio (PR by Ganden Schaffner)
• Improved type annotations:
  • The item_type argument of create_memory_object_stream was deprecated. To indicate the item type handled by the stream, use create_memory_object_stream[T_Item]() instead. Type checking should no longer fail when annotating memory object streams with uninstantiable item types (PR by Ganden Schaffner)
• Added the CancelScope.cancelled_caught property which tells users if the cancel scope suppressed a cancellation exception
• Fixed fail_after() raising an unwarranted TimeoutError when the cancel scope was cancelled before reaching its deadline
• Fixed MemoryObjectReceiveStream.receive() causing the receiving task on asyncio to remain in a cancelled state if the operation was cancelled after an item was queued to be received by the task (but before the task could actually receive the item)
• Fixed TaskGroup.start() on asyncio not responding to cancellation from the outside
• Fixed tasks started from BlockingPortal not notifying synchronous listeners (concurrent.futures.wait()) when they're cancelled
• Removed unnecessary extra waiting cycle in Event.wait() on asyncio in the case where the event was not yet set
• Fixed processes spawned by anyio.to_process() being "lost" as unusable to the process pool when processes that have idled over 5 minutes are pruned at part of the to_process.run_sync() call, leading to increased memory consumption (PR by Anael Gorfinkel)

Changes since 4.0.0rc1:

• Fixed the type annotation of TaskGroup.start_soon() to accept any awaitables (already in v3.7.0 but was missing from 4.0.0rc1)
• Changed CancelScope to also consider the cancellation count (in addition to the cancel message) on asyncio to determine if a cancellation exception should be swallowed on scope exit, to combat issues where third party libraries catch the CancelledError and raise another, thus erasing the original cancel message
• Worked around a CPython bug that caused TLSListener.handle_handshake_error() on asyncio to log "NoneType: None" instead of the error (PR by Ganden Schaffner)
• Re-added the item_type argument to create_memory_object_stream() (but using it raises a deprecation warning and does nothing with regards to the static types of the returned streams)
• Fixed processes spawned by anyio.to_process() being "lost" as unusable to the process pool when processes that have idled over 5 minutes are pruned at part of the to_process.run_sync() call, leading to increased memory consumption (PR by Anael Gorfinkel)