stricter_context_starting
Tyler Goodlet 2021-12-07 11:11:50 -05:00
parent 703dee8a59
commit faaecbf810
1 changed files with 37 additions and 0 deletions

View File

@ -0,0 +1,37 @@
Add cross-actor-task ``Context`` oriented error relay, a new
stream overrun error-signal ``StreamOverrun``, and support
disabling ``MsgStream`` backpressure as the default before a stream
is opened or by choice of the user.
We added stricter semantics around ``tractor.Context.open_stream():``
particularly to do with streams which are only opened at one end.
Previously, if only one end opened a stream there was no way for that
sender to know if msgs are being received until first, the feeder mem
chan on the receiver side hit a backpressure state and then that
condition delayed its msg loop processing task to eventually create
backpressure on the associated IPC transport. This is non-ideal in the
case where the receiver side never opened a stream by mistake since it
results in silent block of the sender and no adherence to the underlying
mem chan buffer size settings (which is still unsolved btw).
To solve this we add non-backpressure style message pushing inside
``Actor._push_result()`` by default and only use the backpressure
``trio.MemorySendChannel.send()`` call **iff** the local end of the
context has entered ``Context.open_stream():``. This way if the stream
was never opened but the mem chan is overrun, we relay back to the
sender a (new exception) ``SteamOverrun`` error which is raised in the
sender's scope with a special error message about the stream never
having been opened. Further, this behaviour (non-backpressure style
where senders can expect an error on overruns) can now be enabled with
``.open_stream(backpressure=False)`` and the underlying mem chan size
can be specified with a kwarg ``msg_buffer_size: int``.
Further bug fixes and enhancements in this changeset include:
- fix a race we were ignoring where if the callee task opened a context
it could enter ``Context.open_stream()`` before calling
``.started()``.
- Disallow calling ``Context.started()`` more then once.
- Enable ``Context`` linked tasks error relaying via the new
``Context._maybe_raise_from_remote_msg()`` which (for now) uses
a simple ``trio.Nursery.start_soon()`` to raise the error via closure
in the local scope.