Add initial bi-directional streaming

This mostly adds the api described in
https://github.com/goodboy/tractor/issues/53#issuecomment-806258798

The first draft summary:
- formalize bidir steaming using the `trio.Channel` style interface
  which we derive as a `MsgStream` type.
- add `Portal.open_context()` which provides a `trio.Nursery.start()`
  remote task invocation style for setting up and tearing down tasks
  contexts in remote actors.
- add a distinct `'started'` message to the ipc protocol to facilitate
  `Context.start()` with a first return value.
- for our `ReceiveMsgStream` type, don't cancel the remote task in
  `.aclose()`; this is now done explicitly by the surrounding `Context`
   usage: `Context.cancel()`.
- streams in either direction still use a `'yield'` message keeping the
  proto mostly symmetric without having to worry about which side is the
  caller / portal opener.
- subtlety: only allow sending a `'stop'` message during a 2-way
  streaming context from `ReceiveStream.aclose()`, detailed comment
  with explanation is included.

Relates to #53
bi_streaming
Tyler Goodlet 2021-05-01 15:10:03 -04:00
parent 55760b3fe0
commit 4371a0a898
3 changed files with 406 additions and 203 deletions

View File

@ -14,6 +14,7 @@ from types import ModuleType
import sys import sys
import os import os
from contextlib import ExitStack from contextlib import ExitStack
import warnings
import trio # type: ignore import trio # type: ignore
from trio_typing import TaskStatus from trio_typing import TaskStatus
@ -58,13 +59,37 @@ async def _invoke(
treat_as_gen = False treat_as_gen = False
cs = None cs = None
cancel_scope = trio.CancelScope() cancel_scope = trio.CancelScope()
ctx = Context(chan, cid, cancel_scope) ctx = Context(chan, cid, _cancel_scope=cancel_scope)
context = False
if getattr(func, '_tractor_stream_function', False): if getattr(func, '_tractor_stream_function', False):
# handle decorated ``@tractor.stream`` async functions # handle decorated ``@tractor.stream`` async functions
sig = inspect.signature(func)
params = sig.parameters
# compat with old api
kwargs['ctx'] = ctx kwargs['ctx'] = ctx
if 'ctx' in params:
warnings.warn(
"`@tractor.stream decorated funcs should now declare "
"a `stream` arg, `ctx` is now designated for use with "
"@tractor.context",
DeprecationWarning,
stacklevel=2,
)
elif 'stream' in params:
assert 'stream' in params
kwargs['stream'] = ctx
treat_as_gen = True treat_as_gen = True
elif getattr(func, '_tractor_context_function', False):
# handle decorated ``@tractor.context`` async function
kwargs['ctx'] = ctx
context = True
# errors raised inside this block are propgated back to caller # errors raised inside this block are propgated back to caller
try: try:
if not ( if not (
@ -102,26 +127,41 @@ async def _invoke(
# `StopAsyncIteration` system here for returning a final # `StopAsyncIteration` system here for returning a final
# value if desired # value if desired
await chan.send({'stop': True, 'cid': cid}) await chan.send({'stop': True, 'cid': cid})
# one way @stream func that gets treated like an async gen
elif treat_as_gen:
await chan.send({'functype': 'asyncgen', 'cid': cid})
# XXX: the async-func may spawn further tasks which push
# back values like an async-generator would but must
# manualy construct the response dict-packet-responses as
# above
with cancel_scope as cs:
task_status.started(cs)
await coro
if not cs.cancelled_caught:
# task was not cancelled so we can instruct the
# far end async gen to tear down
await chan.send({'stop': True, 'cid': cid})
elif context:
# context func with support for bi-dir streaming
await chan.send({'functype': 'context', 'cid': cid})
with cancel_scope as cs:
task_status.started(cs)
await chan.send({'return': await coro, 'cid': cid})
# if cs.cancelled_caught:
# # task was cancelled so relay to the cancel to caller
# await chan.send({'return': await coro, 'cid': cid})
else: else:
if treat_as_gen: # regular async function
await chan.send({'functype': 'asyncgen', 'cid': cid}) await chan.send({'functype': 'asyncfunc', 'cid': cid})
# XXX: the async-func may spawn further tasks which push with cancel_scope as cs:
# back values like an async-generator would but must task_status.started(cs)
# manualy construct the response dict-packet-responses as await chan.send({'return': await coro, 'cid': cid})
# above
with cancel_scope as cs:
task_status.started(cs)
await coro
if not cs.cancelled_caught:
# task was not cancelled so we can instruct the
# far end async gen to tear down
await chan.send({'stop': True, 'cid': cid})
else:
# regular async function
await chan.send({'functype': 'asyncfunc', 'cid': cid})
with cancel_scope as cs:
task_status.started(cs)
await chan.send({'return': await coro, 'cid': cid})
except (Exception, trio.MultiError) as err: except (Exception, trio.MultiError) as err:
@ -416,10 +456,10 @@ class Actor:
send_chan, recv_chan = self._cids2qs[(actorid, cid)] send_chan, recv_chan = self._cids2qs[(actorid, cid)]
assert send_chan.cid == cid # type: ignore assert send_chan.cid == cid # type: ignore
if 'stop' in msg: # if 'stop' in msg:
log.debug(f"{send_chan} was terminated at remote end") # log.debug(f"{send_chan} was terminated at remote end")
# indicate to consumer that far end has stopped # # indicate to consumer that far end has stopped
return await send_chan.aclose() # return await send_chan.aclose()
try: try:
log.debug(f"Delivering {msg} from {actorid} to caller {cid}") log.debug(f"Delivering {msg} from {actorid} to caller {cid}")
@ -427,6 +467,12 @@ class Actor:
await send_chan.send(msg) await send_chan.send(msg)
except trio.BrokenResourceError: except trio.BrokenResourceError:
# TODO: what is the right way to handle the case where the
# local task has already sent a 'stop' / StopAsyncInteration
# to the other side but and possibly has closed the local
# feeder mem chan? Do we wait for some kind of ack or just
# let this fail silently and bubble up (currently)?
# XXX: local consumer has closed their side # XXX: local consumer has closed their side
# so cancel the far end streaming task # so cancel the far end streaming task
log.warning(f"{send_chan} consumer is already closed") log.warning(f"{send_chan} consumer is already closed")
@ -506,6 +552,7 @@ class Actor:
if cid: if cid:
# deliver response to local caller/waiter # deliver response to local caller/waiter
await self._push_result(chan, cid, msg) await self._push_result(chan, cid, msg)
log.debug( log.debug(
f"Waiting on next msg for {chan} from {chan.uid}") f"Waiting on next msg for {chan} from {chan.uid}")
continue continue

View File

@ -312,11 +312,20 @@ class Portal:
ctx = Context(self.channel, cid, _portal=self) ctx = Context(self.channel, cid, _portal=self)
try: try:
async with ReceiveMsgStream(ctx, recv_chan, self) as rchan: # deliver receive only stream
async with ReceiveMsgStream(ctx, recv_chan) as rchan:
self._streams.add(rchan) self._streams.add(rchan)
yield rchan yield rchan
finally: finally:
# cancel the far end task on consumer close # cancel the far end task on consumer close
# NOTE: this is a special case since we assume that if using
# this ``.open_fream_from()`` api, the stream is one a one
# time use and we couple the far end tasks's lifetime to
# the consumer's scope; we don't ever send a `'stop'`
# message right now since there shouldn't be a reason to
# stop and restart the stream, right?
try: try:
await ctx.cancel() await ctx.cancel()
except trio.ClosedResourceError: except trio.ClosedResourceError:
@ -326,16 +335,55 @@ class Portal:
self._streams.remove(rchan) self._streams.remove(rchan)
# @asynccontextmanager @asynccontextmanager
# async def open_context( async def open_context(
# self, self,
# func: Callable, func: Callable,
# **kwargs, **kwargs,
# ) -> Context: ) -> Context:
# # TODO """Open an inter-actor task context.
# elif resptype == 'context': # context manager style setup/teardown
# # TODO likely not here though This is a synchronous API which allows for deterministic
# raise NotImplementedError setup/teardown of a remote task. The yielded ``Context`` further
allows for opening bidirectional streams - see
``Context.open_stream()``.
"""
# conduct target func method structural checks
if not inspect.iscoroutinefunction(func) and (
getattr(func, '_tractor_contex_function', False)
):
raise TypeError(
f'{func} must be an async generator function!')
fn_mod_path, fn_name = func_deats(func)
cid, recv_chan, functype, first_msg = await self._submit(
fn_mod_path, fn_name, kwargs)
assert functype == 'context'
msg = await recv_chan.receive()
try:
# the "first" value here is delivered by the callee's
# ``Context.started()`` call.
first = msg['started']
except KeyError:
assert msg.get('cid'), ("Received internal error at context?")
if msg.get('error'):
# raise the error message
raise unpack_error(msg, self.channel)
else:
raise
try:
ctx = Context(self.channel, cid, _portal=self)
yield ctx, first
finally:
await recv_chan.aclose()
await ctx.cancel()
@dataclass @dataclass

View File

@ -1,19 +1,195 @@
import inspect import inspect
from contextlib import contextmanager # , asynccontextmanager from contextlib import contextmanager, asynccontextmanager
from dataclasses import dataclass from dataclasses import dataclass
from typing import Any, Iterator, Optional from typing import Any, Iterator, Optional, Callable
import warnings import warnings
import trio import trio
from ._ipc import Channel from ._ipc import Channel
from ._exceptions import unpack_error from ._exceptions import unpack_error
from ._state import current_actor
from .log import get_logger from .log import get_logger
log = get_logger(__name__) log = get_logger(__name__)
# TODO: generic typing like trio's receive channel
# but with msgspec messages?
# class ReceiveChannel(AsyncResource, Generic[ReceiveType]):
class ReceiveMsgStream(trio.abc.ReceiveChannel):
"""A wrapper around a ``trio._channel.MemoryReceiveChannel`` with
special behaviour for signalling stream termination across an
inter-actor ``Channel``. This is the type returned to a local task
which invoked a remote streaming function using `Portal.run()`.
Termination rules:
- if the local task signals stop iteration a cancel signal is
relayed to the remote task indicating to stop streaming
- if the remote task signals the end of a stream, raise a
``StopAsyncIteration`` to terminate the local ``async for``
"""
def __init__(
self,
ctx: 'Context', # typing: ignore # noqa
rx_chan: trio.abc.ReceiveChannel,
) -> None:
self._ctx = ctx
self._rx_chan = rx_chan
self._shielded = False
# delegate directly to underlying mem channel
def receive_nowait(self):
return self._rx_chan.receive_nowait()
async def receive(self):
try:
msg = await self._rx_chan.receive()
return msg['yield']
except KeyError:
# internal error should never get here
assert msg.get('cid'), ("Received internal error at portal?")
# TODO: handle 2 cases with 3.10 match syntax
# - 'stop'
# - 'error'
# possibly just handle msg['stop'] here!
if msg.get('stop'):
log.debug(f"{self} was stopped at remote end")
# when the send is closed we assume the stream has
# terminated and signal this local iterator to stop
await self.aclose()
raise trio.EndOfChannel
# TODO: test that shows stream raising an expected error!!!
elif msg.get('error'):
# raise the error message
raise unpack_error(msg, self._ctx.chan)
else:
raise
except (trio.ClosedResourceError, StopAsyncIteration):
# XXX: this indicates that a `stop` message was
# sent by the far side of the underlying channel.
# Currently this is triggered by calling ``.aclose()`` on
# the send side of the channel inside
# ``Actor._push_result()``, but maybe it should be put here?
# to avoid exposing the internal mem chan closing mechanism?
# in theory we could instead do some flushing of the channel
# if needed to ensure all consumers are complete before
# triggering closure too early?
# Locally, we want to close this stream gracefully, by
# terminating any local consumers tasks deterministically.
# We **don't** want to be closing this send channel and not
# relaying a final value to remaining consumers who may not
# have been scheduled to receive it yet?
# lots of testing to do here
# when the send is closed we assume the stream has
# terminated and signal this local iterator to stop
await self.aclose()
# await self._ctx.send_stop()
raise StopAsyncIteration
except trio.Cancelled:
# relay cancels to the remote task
await self.aclose()
raise
@contextmanager
def shield(
self
) -> Iterator['ReceiveMsgStream']: # noqa
"""Shield this stream's underlying channel such that a local consumer task
can be cancelled (and possibly restarted) using ``trio.Cancelled``.
"""
self._shielded = True
yield self
self._shielded = False
async def aclose(self):
"""Cancel associated remote actor task and local memory channel
on close.
"""
# TODO: proper adherance to trio's `.aclose()` semantics:
# https://trio.readthedocs.io/en/stable/reference-io.html#trio.abc.AsyncResource.aclose
rx_chan = self._rx_chan
if rx_chan._closed:
log.warning(f"{self} is already closed")
return
# TODO: broadcasting to multiple consumers
# stats = rx_chan.statistics()
# if stats.open_receive_channels > 1:
# # if we've been cloned don't kill the stream
# log.debug(
# "there are still consumers running keeping stream alive")
# return
if self._shielded:
log.warning(f"{self} is shielded, portal channel being kept alive")
return
# NOTE: this is super subtle IPC messaging stuff:
# Relay stop iteration to far end **iff** we're
# in bidirectional mode. If we're only streaming
# *from* one side then that side **won't** have an
# entry in `Actor._cids2qs` (maybe it should though?).
# So any `yield` or `stop` msgs sent from the caller side
# will cause key errors on the callee side since there is
# no entry for a local feeder mem chan since the callee task
# isn't expecting messages to be sent by the caller.
# Thus, we must check that this context DOES NOT
# have a portal reference to ensure this is indeed the callee
# side and can relay a 'stop'. In the bidirectional case,
# `Context.open_stream()` will create the `Actor._cids2qs`
# entry from a call to `Actor.get_memchans()`.
if not self._ctx._portal:
# only for 2 way streams can we can send
# stop from the caller side
await self._ctx.send_stop()
# close the local mem chan
rx_chan.close()
# TODO: but make it broadcasting to consumers
# def clone(self):
# """Clone this receive channel allowing for multi-task
# consumption from the same channel.
# """
# return ReceiveStream(
# self._cid,
# self._rx_chan.clone(),
# self._portal,
# )
class MsgStream(ReceiveMsgStream, trio.abc.Channel):
"""
Bidirectional message stream for use within an inter-actor actor
``Context```.
"""
async def send(
self,
data: Any
) -> None:
await self._ctx.chan.send({'yield': data, 'cid': self._ctx.cid})
@dataclass(frozen=True) @dataclass(frozen=True)
class Context: class Context:
"""An IAC (inter-actor communication) context. """An IAC (inter-actor communication) context.
@ -31,6 +207,10 @@ class Context:
chan: Channel chan: Channel
cid: str cid: str
# TODO: should we have seperate types for caller vs. callee
# side contexts? The caller always opens a portal whereas the callee
# is always responding back through a context-stream
# only set on the caller side # only set on the caller side
_portal: Optional['Portal'] = None # type: ignore # noqa _portal: Optional['Portal'] = None # type: ignore # noqa
@ -57,46 +237,97 @@ class Context:
timeout quickly to sidestep 2-generals... timeout quickly to sidestep 2-generals...
""" """
assert self._portal, ( if self._portal: # caller side:
"No portal found, this is likely a callee side context") if not self._portal:
raise RuntimeError(
"No portal found, this is likely a callee side context"
)
cid = self.cid cid = self.cid
with trio.move_on_after(0.5) as cs: with trio.move_on_after(0.5) as cs:
cs.shield = True cs.shield = True
log.warning(
f"Cancelling stream {cid} to "
f"{self._portal.channel.uid}")
# NOTE: we're telling the far end actor to cancel a task
# corresponding to *this actor*. The far end local channel
# instance is passed to `Actor._cancel_task()` implicitly.
await self._portal.run_from_ns('self', '_cancel_task', cid=cid)
if cs.cancelled_caught:
# XXX: there's no way to know if the remote task was indeed
# cancelled in the case where the connection is broken or
# some other network error occurred.
if not self._portal.channel.connected():
log.warning( log.warning(
"May have failed to cancel remote task " f"Cancelling stream {cid} to "
f"{cid} for {self._portal.channel.uid}") f"{self._portal.channel.uid}")
# NOTE: we're telling the far end actor to cancel a task
# corresponding to *this actor*. The far end local channel
# instance is passed to `Actor._cancel_task()` implicitly.
await self._portal.run_from_ns('self', '_cancel_task', cid=cid)
if cs.cancelled_caught:
# XXX: there's no way to know if the remote task was indeed
# cancelled in the case where the connection is broken or
# some other network error occurred.
# if not self._portal.channel.connected():
if not self.chan.connected():
log.warning(
"May have failed to cancel remote task "
f"{cid} for {self._portal.channel.uid}")
else:
# ensure callee side
assert self._cancel_scope
# TODO: should we have an explicit cancel message
# or is relaying the local `trio.Cancelled` as an
# {'error': trio.Cancelled, cid: "blah"} enough?
# This probably gets into the discussion in
# https://github.com/goodboy/tractor/issues/36
self._cancel_scope.cancel()
# TODO: do we need a restart api?
# async def restart(self) -> None: # async def restart(self) -> None:
# # TODO
# pass # pass
# @asynccontextmanager @asynccontextmanager
# async def open_stream( async def open_stream(
# self, self,
# ) -> AsyncContextManager: ) -> MsgStream:
# # TODO # TODO
# pass
actor = current_actor()
# here we create a mem chan that corresponds to the
# far end caller / callee.
# NOTE: in one way streaming this only happens on the
# caller side inside `Actor.send_cmd()` so if you try
# to send a stop from the caller to the callee in the
# single-direction-stream case you'll get a lookup error
# currently.
_, recv_chan = actor.get_memchans(
self.chan.uid,
self.cid
)
async with MsgStream(ctx=self, rx_chan=recv_chan) as rchan:
if self._portal:
self._portal._streams.add(rchan)
try:
yield rchan
finally:
await self.send_stop()
if self._portal:
self._portal._streams.add(rchan)
async def started(self, value: Any) -> None:
if self._portal:
raise RuntimeError(
f"Caller side context {self} can not call started!")
await self.chan.send({'started': value, 'cid': self.cid})
def stream(func): def stream(func: Callable) -> Callable:
"""Mark an async function as a streaming routine with ``@stream``. """Mark an async function as a streaming routine with ``@stream``.
""" """
# annotate
func._tractor_stream_function = True func._tractor_stream_function = True
sig = inspect.signature(func) sig = inspect.signature(func)
params = sig.parameters params = sig.parameters
if 'stream' not in params and 'ctx' in params: if 'stream' not in params and 'ctx' in params:
@ -114,147 +345,24 @@ def stream(func):
): ):
raise TypeError( raise TypeError(
"The first argument to the stream function " "The first argument to the stream function "
f"{func.__name__} must be `ctx: tractor.Context`" f"{func.__name__} must be `ctx: tractor.Context` "
"(Or ``to_trio`` if using ``asyncio`` in guest mode)."
) )
return func return func
class ReceiveMsgStream(trio.abc.ReceiveChannel): def context(func: Callable) -> Callable:
"""A wrapper around a ``trio._channel.MemoryReceiveChannel`` with """Mark an async function as a streaming routine with ``@context``.
special behaviour for signalling stream termination across an
inter-actor ``Channel``. This is the type returned to a local task
which invoked a remote streaming function using `Portal.run()`.
Termination rules:
- if the local task signals stop iteration a cancel signal is
relayed to the remote task indicating to stop streaming
- if the remote task signals the end of a stream, raise a
``StopAsyncIteration`` to terminate the local ``async for``
""" """
def __init__( # annotate
self, func._tractor_context_function = True
ctx: Context,
rx_chan: trio.abc.ReceiveChannel,
portal: 'Portal', # type: ignore # noqa
) -> None:
self._ctx = ctx
self._rx_chan = rx_chan
self._portal = portal
self._shielded = False
# delegate directly to underlying mem channel sig = inspect.signature(func)
def receive_nowait(self): params = sig.parameters
return self._rx_chan.receive_nowait() if 'ctx' not in params:
raise TypeError(
async def receive(self): "The first argument to the context function "
try: f"{func.__name__} must be `ctx: tractor.Context`"
msg = await self._rx_chan.receive() )
return msg['yield'] return func
except KeyError:
# internal error should never get here
assert msg.get('cid'), ("Received internal error at portal?")
# TODO: handle 2 cases with 3.10 match syntax
# - 'stop'
# - 'error'
# possibly just handle msg['stop'] here!
# TODO: test that shows stream raising an expected error!!!
if msg.get('error'):
# raise the error message
raise unpack_error(msg, self._portal.channel)
except (trio.ClosedResourceError, StopAsyncIteration):
# XXX: this indicates that a `stop` message was
# sent by the far side of the underlying channel.
# Currently this is triggered by calling ``.aclose()`` on
# the send side of the channel inside
# ``Actor._push_result()``, but maybe it should be put here?
# to avoid exposing the internal mem chan closing mechanism?
# in theory we could instead do some flushing of the channel
# if needed to ensure all consumers are complete before
# triggering closure too early?
# Locally, we want to close this stream gracefully, by
# terminating any local consumers tasks deterministically.
# We **don't** want to be closing this send channel and not
# relaying a final value to remaining consumers who may not
# have been scheduled to receive it yet?
# lots of testing to do here
# when the send is closed we assume the stream has
# terminated and signal this local iterator to stop
await self.aclose()
raise StopAsyncIteration
except trio.Cancelled:
# relay cancels to the remote task
await self.aclose()
raise
@contextmanager
def shield(
self
) -> Iterator['ReceiveMsgStream']: # noqa
"""Shield this stream's underlying channel such that a local consumer task
can be cancelled (and possibly restarted) using ``trio.Cancelled``.
"""
self._shielded = True
yield self
self._shielded = False
async def aclose(self):
"""Cancel associated remote actor task and local memory channel
on close.
"""
rx_chan = self._rx_chan
if rx_chan._closed:
log.warning(f"{self} is already closed")
return
# stats = rx_chan.statistics()
# if stats.open_receive_channels > 1:
# # if we've been cloned don't kill the stream
# log.debug(
# "there are still consumers running keeping stream alive")
# return
if self._shielded:
log.warning(f"{self} is shielded, portal channel being kept alive")
return
# close the local mem chan
rx_chan.close()
# cancel surrounding IPC context
await self._ctx.cancel()
# TODO: but make it broadcasting to consumers
# def clone(self):
# """Clone this receive channel allowing for multi-task
# consumption from the same channel.
# """
# return ReceiveStream(
# self._cid,
# self._rx_chan.clone(),
# self._portal,
# )
# class MsgStream(ReceiveMsgStream, trio.abc.Channel):
# """
# Bidirectional message stream for use within an inter-actor actor
# ``Context```.
# """
# async def send(
# self,
# data: Any
# ) -> None:
# await self._ctx.chan.send({'yield': data, 'cid': self._ctx.cid})