This adds remote cancellation semantics to our `tractor.Context`
machinery to more closely match that of `trio.CancelScope` but
with operational differences to handle the nature of parallel tasks interoperating
across multiple memory boundaries:
- if an actor task cancels some context it has opened via
`Context.cancel()`, the remote (scope linked) task will be cancelled
using the normal `CancelScope` semantics of `trio` meaning the remote
cancel scope surrounding the far side task is cancelled and
`trio.Cancelled`s are expected to be raised in that scope as per
normal `trio` operation, and in the case where no error is raised
in that remote scope, a `ContextCancelled` error is raised inside the
runtime machinery and relayed back to the opener/caller side of the
context.
- if any actor task cancels a full remote actor runtime using
`Portal.cancel_actor()` the same semantics as above apply except every
other remote actor task which also has an open context with the actor
which was cancelled will also be sent a `ContextCancelled` **but**
with the `.canceller` field set to the uid of the original cancel
requesting actor.
This changeset also includes a more "proper" solution to the issue of
"allowing overruns" during streaming without attempting to implement any
form of IPC streaming backpressure. Implementing task-granularity
backpressure cross-process turns out to be more or less impossible
without augmenting out streaming protocol (likely at the cost of
performance). Further allowing overruns requires special care since
any blocking of the runtime RPC msg loop task effectively can block
control msgs such as cancels and stream terminations.
The implementation details per abstraction layer are as follows.
._streaming.Context:
- add a new contructor factor func `mk_context()` which provides
a strictly private init-er whilst allowing us to not have to define
an `.__init__()` on the type def.
- add public `.cancel_called` and `.cancel_called_remote` properties.
- general rename of what was the internal `._backpressure` var to
`._allow_overruns: bool`.
- move the old contents of `Actor._push_result()` into a new
`._deliver_msg()` allowing for better encapsulation of per-ctx
msg handling.
- always check for received 'error' msgs and process them with the new
`_maybe_cancel_and_set_remote_error()` **before** any msg delivery to
the local task, thus guaranteeing error and cancellation handling
despite any overflow handling.
- add a new `._drain_overflows()` task-method for use with new
`._allow_overruns: bool = True` mode.
- add back a `._scope_nursery: trio.Nursery` (allocated in
`Portal.open_context()`) who's sole purpose is to spawn a single task
which runs the above method; anything else is an error.
- augment `._deliver_msg()` to start a task and run the above method
when operating in no overrun mode; the task queues overflow msgs and
attempts to send them to the underlying mem chan using a blocking
`.send()` call.
- on context exit, any existing "drainer task" will be cancelled and
remaining overflow queued msgs are discarded with a warning.
- rename `._error` -> `_remote_error` and set it in a new method
`_maybe_cancel_and_set_remote_error()` which is called before
processing
- adjust `.result()` to always call `._maybe_raise_remote_err()` at its
start such that whenever a `ContextCancelled` arrives we do logic for
whether or not to immediately raise that error or ignore it due to the
current actor being the one who requested the cancel, by checking the
error's `.canceller` field.
- set the default value of `._result` to be `id(Context()` thus avoiding
conflict with any `.result()` actually being `False`..
._runtime.Actor:
- augment `.cancel()` and `._cancel_task()` and `.cancel_rpc_tasks()` to
take a `requesting_uid: tuple` indicating the source actor of every
cancellation request.
- pass through the new `Context._allow_overruns` through `.get_context()`
- call the new `Context._deliver_msg()` from `._push_result()` (since
the factoring out that method's contents).
._runtime._invoke:
- `TastStatus.started()` back a `Context` (unless an error is raised)
instead of the cancel scope to make it easy to set/get state on that
context for the purposes of cancellation and remote error relay.
- always raise any remote error via `Context._maybe_raise_remote_err()`
before doing any `ContextCancelled` logic.
- assign any `Context._cancel_called_remote` set by the `requesting_uid`
cancel methods (mentioned above) to the `ContextCancelled.canceller`.
._runtime.process_messages:
- always pass a `requesting_uid: tuple` to `Actor.cancel()` and
`._cancel_task` to that any corresponding `ContextCancelled.canceller`
can be set inside `._invoke()`.
Turns out stuff was totally broken in these cases because we're either
closing the underlying mem chan too early or not handling the
"allow_overruns" mode's cancellation correctly..
To handle both remote cancellation this adds `ContextCanceled.canceller:
tuple` the uid of the cancel requesting actor and is expected to be set
by the runtime when servicing any remote cancel request. This makes it
possible for `ContextCancelled` receivers to know whether "their actor
runtime" is the source of the cancellation.
Also add an explicit `RemoteActor.src_actor_uid` which better formalizes
the notion of "which remote actor" the error originated from.
Both of these new attrs are expected to be packed in the `.msgdata` when
the errors are loaded locally.
These will verify new changes to the runtime/messaging core which allows
us to adopt an "ignore cancel if requested by us" style handling of
`ContextCancelled` more like how `trio` does with
`trio.Nursery.cancel_scope.cancel()`. We now expect
a `ContextCancelled.canceller: tuple` which is set to the actor uid of
the actor which requested the cancellation which eventually resulted in
the remote error-msg.
Also adds some experimental tweaks to the "backpressure" test which it
turns out is very problematic in coordination with context cancellation
since blocking on the feed mem chan to some task will block the ipc msg
loop and thus handling of cancellation.. More to come to both the test
and core to address this hopefully since right now this test is failing.
Makes the broadcast test suite not hang xD, and is our expected default
behaviour. Also removes a ton of commented legacy cruft from before the
refactor to remove the `.receive()` recursion and fixes some typing.
Oh right, and in the case where there's only one subscriber left we warn
log about it since in theory we could actually entirely unwind the
bcaster back to the original underlying, though not sure if that's sane
or works for some use cases (like wanting to have some other subscriber
get added dynamically later).
Since one-way streaming can be accomplished by just *not* sending on one
side (and/or thus wrapping such usage in a more restrictive API), we
just drop the recv-only parent type. The only method different was
`MsgStream.send()`, now merged in. Further in usage of `.subscribe()`
we monkey patch the underlying stream's `.send()` onto the delivered
broadcast receiver so that subscriber tasks can two-way stream as though
using the stream directly.
This allows us to more definitively drop `tractor.open_stream_from()` in
the longer run if we so choose as well; note currently this will
potentially create an issue if a caller tries to `.send()` on such a one
way stream.
Driven by a bug found in `piker` where we'd get an inf recursion error
due to `BroadcastReceiver.receive()` being called when consumer tasks
are awoken but no value is ready to `.nowait_receive()`.
This new rework takes an approach closer to the interface and internals
of `trio.MemoryReceiveChannel` particularly in terms of,
- implementing a `BroadcastReceiver.receive_nowait()` and using it
within the async `.receive()`.
- failing over to an internal `._receive_from_underlying()` when the
`_nowait()` call raises `trio.WouldBlock`.
- adding `BroadcastState.statistics()` for debugging and testing
dropping recursion from `.receive()`.
We weren't doing this originally I *think* just because of the path
dependent nature of the way the code was developed (originally being
mega pedantic about one-way vs. bidirectional streams) but, it doesn't
seem like there's any issue just calling the stream's `.aclose()`; also
have the benefit of just being less code and logic checks B)
With the new fancy `_pytest.pathlib.import_path()` we can do real
parametrization of the example-script-module code and thus configure
whether the child, parent, or both silently break the IPC connection.
Parametrize the test for all the above mentioned cases as well as the
case where the IPC never breaks but we still simulate the user hammering
ctl-c / SIGINT to terminate the actor tree. Adjust expected errors based
on each case and heavily document each of these.
Use a task nursery in the subactor to spawn tasks which cancel the IPC
channel mid stream to simulate the most concurrent case we're likely to
see. Make `main()` accept a `debug_mode: bool` for parametrization. Fill
out detailed comments/docs on this example.
When backpressure is used and a feeder mem chan breaks during msg
delivery (usually because the IPC allocating task already terminated)
instead of raising we simply warn as we do for the non-backpressure
case.
Also, add a proper `Actor.is_arbiter` test inside `._invoke()` to avoid
doing an arbiter-registry lookup if the current actor **is** the
registrar.
Tyler Goodlet
goodboy