forked from goodboy/tractor
1
0
Fork 0

Merge pull request #346 from goodboy/ipc_failure_while_streaming

Ipc failure while streaming
ipython_integration
goodboy 2023-01-29 15:02:54 -05:00 committed by GitHub
commit a777217674
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
12 changed files with 471 additions and 86 deletions

View File

@ -0,0 +1,151 @@
'''
Complex edge case where during real-time streaming the IPC tranport
channels are wiped out (purposely in this example though it could have
been an outage) and we want to ensure that despite being in debug mode
(or not) the user can sent SIGINT once they notice the hang and the
actor tree will eventually be cancelled without leaving any zombies.
'''
import trio
from tractor import (
open_nursery,
context,
Context,
MsgStream,
)
async def break_channel_silently_then_error(
stream: MsgStream,
):
async for msg in stream:
await stream.send(msg)
# XXX: close the channel right after an error is raised
# purposely breaking the IPC transport to make sure the parent
# doesn't get stuck in debug or hang on the connection join.
# this more or less simulates an infinite msg-receive hang on
# the other end.
await stream._ctx.chan.send(None)
assert 0
async def close_stream_and_error(
stream: MsgStream,
):
async for msg in stream:
await stream.send(msg)
# wipe out channel right before raising
await stream._ctx.chan.send(None)
await stream.aclose()
assert 0
@context
async def recv_and_spawn_net_killers(
ctx: Context,
break_ipc_after: bool | int = False,
) -> None:
'''
Receive stream msgs and spawn some IPC killers mid-stream.
'''
await ctx.started()
async with (
ctx.open_stream() as stream,
trio.open_nursery() as n,
):
async for i in stream:
print(f'child echoing {i}')
await stream.send(i)
if (
break_ipc_after
and i > break_ipc_after
):
'#################################\n'
'Simulating child-side IPC BREAK!\n'
'#################################'
n.start_soon(break_channel_silently_then_error, stream)
n.start_soon(close_stream_and_error, stream)
async def main(
debug_mode: bool = False,
start_method: str = 'trio',
# by default we break the parent IPC first (if configured to break
# at all), but this can be changed so the child does first (even if
# both are set to break).
break_parent_ipc_after: int | bool = False,
break_child_ipc_after: int | bool = False,
) -> None:
async with (
open_nursery(
start_method=start_method,
# NOTE: even debugger is used we shouldn't get
# a hang since it never engages due to broken IPC
debug_mode=debug_mode,
loglevel='warning',
) as an,
):
portal = await an.start_actor(
'chitty_hijo',
enable_modules=[__name__],
)
async with portal.open_context(
recv_and_spawn_net_killers,
break_ipc_after=break_child_ipc_after,
) as (ctx, sent):
async with ctx.open_stream() as stream:
for i in range(1000):
if (
break_parent_ipc_after
and i > break_parent_ipc_after
):
print(
'#################################\n'
'Simulating parent-side IPC BREAK!\n'
'#################################'
)
await stream._ctx.chan.send(None)
# it actually breaks right here in the
# mp_spawn/forkserver backends and thus the zombie
# reaper never even kicks in?
print(f'parent sending {i}')
await stream.send(i)
with trio.move_on_after(2) as cs:
# NOTE: in the parent side IPC failure case this
# will raise an ``EndOfChannel`` after the child
# is killed and sends a stop msg back to it's
# caller/this-parent.
rx = await stream.receive()
print(f"I'm a happy user and echoed to me is {rx}")
if cs.cancelled_caught:
# pretend to be a user seeing no streaming action
# thinking it's a hang, and then hitting ctl-c..
print("YOO i'm a user anddd thingz hangin..")
print(
"YOO i'm mad send side dun but thingz hangin..\n"
'MASHING CTlR-C Ctl-c..'
)
raise KeyboardInterrupt
if __name__ == '__main__':
trio.run(main)

View File

@ -0,0 +1,15 @@
Fixes to ensure IPC (channel) breakage doesn't result in hung actor
trees; the zombie reaping and general supervision machinery will always
clean up and terminate.
This includes not only the (mostly minor) fixes to solve these cases but
also a new extensive test suite in `test_advanced_faults.py` with an
accompanying highly configurable example module-script in
`examples/advanced_faults/ipc_failure_during_stream.py`. Tests ensure we
never get hang or zombies despite operating in debug mode and attempt to
simulate all possible IPC transport failure cases for a local-host actor
tree.
Further we simplify `Context.open_stream.__aexit__()` to just call
`MsgStream.aclose()` directly more or less avoiding a pure duplicate
code path.

View File

@ -7,6 +7,7 @@ import os
import random
import signal
import platform
import pathlib
import time
import inspect
from functools import partial, wraps
@ -113,14 +114,21 @@ no_windows = pytest.mark.skipif(
)
def repodir():
"""Return the abspath to the repo directory.
"""
dirname = os.path.dirname
dirpath = os.path.abspath(
dirname(dirname(os.path.realpath(__file__)))
)
return dirpath
def repodir() -> pathlib.Path:
'''
Return the abspath to the repo directory.
'''
# 2 parents up to step up through tests/<repo_dir>
return pathlib.Path(__file__).parent.parent.absolute()
def examples_dir() -> pathlib.Path:
'''
Return the abspath to the examples directory as `pathlib.Path`.
'''
return repodir() / 'examples'
def pytest_addoption(parser):
@ -151,7 +159,7 @@ def loglevel(request):
@pytest.fixture(scope='session')
def spawn_backend(request):
def spawn_backend(request) -> str:
return request.config.option.spawn_backend

View File

@ -0,0 +1,193 @@
'''
Sketchy network blackoutz, ugly byzantine gens, puedes eschuchar la
cancelacion?..
'''
from functools import partial
import pytest
from _pytest.pathlib import import_path
import trio
import tractor
from conftest import (
examples_dir,
)
@pytest.mark.parametrize(
'debug_mode',
[False, True],
ids=['no_debug_mode', 'debug_mode'],
)
@pytest.mark.parametrize(
'ipc_break',
[
# no breaks
{
'break_parent_ipc_after': False,
'break_child_ipc_after': False,
},
# only parent breaks
{
'break_parent_ipc_after': 500,
'break_child_ipc_after': False,
},
# only child breaks
{
'break_parent_ipc_after': False,
'break_child_ipc_after': 500,
},
# both: break parent first
{
'break_parent_ipc_after': 500,
'break_child_ipc_after': 800,
},
# both: break child first
{
'break_parent_ipc_after': 800,
'break_child_ipc_after': 500,
},
],
ids=[
'no_break',
'break_parent',
'break_child',
'break_both_parent_first',
'break_both_child_first',
],
)
def test_ipc_channel_break_during_stream(
debug_mode: bool,
spawn_backend: str,
ipc_break: dict | None,
):
'''
Ensure we can have an IPC channel break its connection during
streaming and it's still possible for the (simulated) user to kill
the actor tree using SIGINT.
We also verify the type of connection error expected in the parent
depending on which side if the IPC breaks first.
'''
if spawn_backend != 'trio':
if debug_mode:
pytest.skip('`debug_mode` only supported on `trio` spawner')
# non-`trio` spawners should never hit the hang condition that
# requires the user to do ctl-c to cancel the actor tree.
expect_final_exc = trio.ClosedResourceError
mod = import_path(
examples_dir() / 'advanced_faults' / 'ipc_failure_during_stream.py',
root=examples_dir(),
)
expect_final_exc = KeyboardInterrupt
# when ONLY the child breaks we expect the parent to get a closed
# resource error on the next `MsgStream.receive()` and then fail out
# and cancel the child from there.
if (
# only child breaks
(
ipc_break['break_child_ipc_after']
and ipc_break['break_parent_ipc_after'] is False
)
# both break but, parent breaks first
or (
ipc_break['break_child_ipc_after'] is not False
and (
ipc_break['break_parent_ipc_after']
> ipc_break['break_child_ipc_after']
)
)
):
expect_final_exc = trio.ClosedResourceError
# when the parent IPC side dies (even if the child's does as well
# but the child fails BEFORE the parent) we expect the channel to be
# sent a stop msg from the child at some point which will signal the
# parent that the stream has been terminated.
# NOTE: when the parent breaks "after" the child you get this same
# case as well, the child breaks the IPC channel with a stop msg
# before any closure takes place.
elif (
# only parent breaks
(
ipc_break['break_parent_ipc_after']
and ipc_break['break_child_ipc_after'] is False
)
# both break but, child breaks first
or (
ipc_break['break_parent_ipc_after'] is not False
and (
ipc_break['break_child_ipc_after']
> ipc_break['break_parent_ipc_after']
)
)
):
expect_final_exc = trio.EndOfChannel
with pytest.raises(expect_final_exc):
trio.run(
partial(
mod.main,
debug_mode=debug_mode,
start_method=spawn_backend,
**ipc_break,
)
)
@tractor.context
async def break_ipc_after_started(
ctx: tractor.Context,
) -> None:
await ctx.started()
async with ctx.open_stream() as stream:
await stream.aclose()
await trio.sleep(0.2)
await ctx.chan.send(None)
print('child broke IPC and terminating')
def test_stream_closed_right_after_ipc_break_and_zombie_lord_engages():
'''
Verify that is a subactor's IPC goes down just after bringing up a stream
the parent can trigger a SIGINT and the child will be reaped out-of-IPC by
the localhost process supervision machinery: aka "zombie lord".
'''
async def main():
async with tractor.open_nursery() as n:
portal = await n.start_actor(
'ipc_breaker',
enable_modules=[__name__],
)
with trio.move_on_after(1):
async with (
portal.open_context(
break_ipc_after_started
) as (ctx, sent),
):
async with ctx.open_stream():
await trio.sleep(0.5)
print('parent waiting on context')
print('parent exited context')
raise KeyboardInterrupt
with pytest.raises(KeyboardInterrupt):
trio.run(main)

View File

@ -14,6 +14,7 @@ import itertools
from os import path
from typing import Optional
import platform
import pathlib
import sys
import time
@ -24,7 +25,10 @@ from pexpect.exceptions import (
EOF,
)
from conftest import repodir, _ci_env
from conftest import (
examples_dir,
_ci_env,
)
# TODO: The next great debugger audit could be done by you!
# - recurrent entry to breakpoint() from single actor *after* and an
@ -43,19 +47,13 @@ if platform.system() == 'Windows':
)
def examples_dir():
"""Return the abspath to the examples directory.
"""
return path.join(repodir(), 'examples', 'debugging/')
def mk_cmd(ex_name: str) -> str:
"""Generate a command suitable to pass to ``pexpect.spawn()``.
"""
return ' '.join(
['python',
path.join(examples_dir(), f'{ex_name}.py')]
)
'''
Generate a command suitable to pass to ``pexpect.spawn()``.
'''
script_path: pathlib.Path = examples_dir() / 'debugging' / f'{ex_name}.py'
return ' '.join(['python', str(script_path)])
# TODO: was trying to this xfail style but some weird bug i see in CI

View File

@ -12,17 +12,17 @@ import shutil
import pytest
from conftest import repodir
def examples_dir():
"""Return the abspath to the examples directory.
"""
return os.path.join(repodir(), 'examples')
from conftest import (
examples_dir,
)
@pytest.fixture
def run_example_in_subproc(loglevel, testdir, arb_addr):
def run_example_in_subproc(
loglevel: str,
testdir,
arb_addr: tuple[str, int],
):
@contextmanager
def run(script_code):
@ -32,8 +32,8 @@ def run_example_in_subproc(loglevel, testdir, arb_addr):
# on windows we need to create a special __main__.py which will
# be executed with ``python -m <modulename>`` on windows..
shutil.copyfile(
os.path.join(examples_dir(), '__main__.py'),
os.path.join(str(testdir), '__main__.py')
examples_dir() / '__main__.py',
str(testdir / '__main__.py'),
)
# drop the ``if __name__ == '__main__'`` guard onwards from
@ -88,6 +88,7 @@ def run_example_in_subproc(loglevel, testdir, arb_addr):
and f[0] != '_'
and 'debugging' not in p[0]
and 'integration' not in p[0]
and 'advanced_faults' not in p[0]
],
ids=lambda t: t[1],

View File

@ -251,7 +251,7 @@ def test_a_quadruple_example(time_quad_ex, ci_env, spawn_backend):
results, diff = time_quad_ex
assert results
this_fast = 6 if platform.system() in ('Windows', 'Darwin') else 2.666
this_fast = 6 if platform.system() in ('Windows', 'Darwin') else 3
assert diff < this_fast

View File

@ -497,6 +497,10 @@ class Portal:
f'actor: {uid}'
)
result = await ctx.result()
log.runtime(
f'Context {fn_name} returned '
f'value from callee `{result}`'
)
# though it should be impossible for any tasks
# operating *in* this scope to have survived
@ -518,12 +522,6 @@ class Portal:
f'task:{cid}\n'
f'actor:{uid}'
)
else:
log.runtime(
f'Context {fn_name} returned '
f'value from callee `{result}`'
)
# XXX: (MEGA IMPORTANT) if this is a root opened process we
# wait for any immediate child in debug before popping the
# context from the runtime msg loop otherwise inside

View File

@ -228,11 +228,11 @@ async def _invoke(
fname = func.__name__
if ctx._cancel_called:
msg = f'{fname} cancelled itself'
msg = f'`{fname}()` cancelled itself'
elif cs.cancel_called:
msg = (
f'{fname} was remotely cancelled by its caller '
f'`{fname}()` was remotely cancelled by its caller '
f'{ctx.chan.uid}'
)
@ -319,7 +319,7 @@ async def _invoke(
BrokenPipeError,
):
# if we can't propagate the error that's a big boo boo
log.error(
log.exception(
f"Failed to ship error to caller @ {chan.uid} !?"
)
@ -455,7 +455,7 @@ class Actor:
self._mods: dict[str, ModuleType] = {}
self.loglevel = loglevel
self._arb_addr = (
self._arb_addr: tuple[str, int] | None = (
str(arbiter_addr[0]),
int(arbiter_addr[1])
) if arbiter_addr else None
@ -488,7 +488,10 @@ class Actor:
self._parent_chan: Optional[Channel] = None
self._forkserver_info: Optional[
tuple[Any, Any, Any, Any, Any]] = None
self._actoruid2nursery: dict[Optional[tuple[str, str]], 'ActorNursery'] = {} # type: ignore # noqa
self._actoruid2nursery: dict[
tuple[str, str],
ActorNursery | None,
] = {} # type: ignore # noqa
async def wait_for_peer(
self, uid: tuple[str, str]
@ -826,7 +829,12 @@ class Actor:
if ctx._backpressure:
log.warning(text)
await send_chan.send(msg)
try:
await send_chan.send(msg)
except trio.BrokenResourceError:
# XXX: local consumer has closed their side
# so cancel the far end streaming task
log.warning(f"{chan} is already closed")
else:
try:
raise StreamOverrun(text) from None
@ -1371,10 +1379,12 @@ async def async_main(
actor.lifetime_stack.close()
# Unregister actor from the arbiter
if registered_with_arbiter and (
actor._arb_addr is not None
if (
registered_with_arbiter
and not actor.is_arbiter
):
failed = False
assert isinstance(actor._arb_addr, tuple)
with trio.move_on_after(0.5) as cs:
cs.shield = True
try:

View File

@ -260,7 +260,9 @@ async def soft_wait(
if proc.poll() is None: # type: ignore
log.warning(
f'Process still alive after cancel request:\n{uid}')
'Actor still alive after cancel request:\n'
f'{uid}'
)
n.cancel_scope.cancel()
raise

View File

@ -97,6 +97,9 @@ class ReceiveMsgStream(trio.abc.ReceiveChannel):
if self._eoc:
raise trio.EndOfChannel
if self._closed:
raise trio.ClosedResourceError('This stream was closed')
try:
msg = await self._rx_chan.receive()
return msg['yield']
@ -110,6 +113,9 @@ class ReceiveMsgStream(trio.abc.ReceiveChannel):
# - 'error'
# possibly just handle msg['stop'] here!
if self._closed:
raise trio.ClosedResourceError('This stream was closed')
if msg.get('stop') or self._eoc:
log.debug(f"{self} was stopped at remote end")
@ -189,7 +195,6 @@ class ReceiveMsgStream(trio.abc.ReceiveChannel):
return
self._eoc = True
self._closed = True
# NOTE: this is super subtle IPC messaging stuff:
# Relay stop iteration to far end **iff** we're
@ -206,29 +211,32 @@ class ReceiveMsgStream(trio.abc.ReceiveChannel):
# In the bidirectional case, `Context.open_stream()` will create
# the `Actor._cids2qs` entry from a call to
# `Actor.get_context()` and will send the stop message in
# ``__aexit__()`` on teardown so it **does not** need to be
# called here.
if not self._ctx._portal:
# Only for 2 way streams can we can send stop from the
# caller side.
try:
# NOTE: if this call is cancelled we expect this end to
# handle as though the stop was never sent (though if it
# was it shouldn't matter since it's unlikely a user
# will try to re-use a stream after attemping to close
# it).
with trio.CancelScope(shield=True):
await self._ctx.send_stop()
# `Actor.get_context()` and will call us here to send the stop
# msg in ``__aexit__()`` on teardown.
try:
# NOTE: if this call is cancelled we expect this end to
# handle as though the stop was never sent (though if it
# was it shouldn't matter since it's unlikely a user
# will try to re-use a stream after attemping to close
# it).
with trio.CancelScope(shield=True):
await self._ctx.send_stop()
except (
trio.BrokenResourceError,
trio.ClosedResourceError
):
# the underlying channel may already have been pulled
# in which case our stop message is meaningless since
# it can't traverse the transport.
log.debug(f'Channel for {self} was already closed')
except (
trio.BrokenResourceError,
trio.ClosedResourceError
):
# the underlying channel may already have been pulled
# in which case our stop message is meaningless since
# it can't traverse the transport.
ctx = self._ctx
log.warning(
f'Stream was already destroyed?\n'
f'actor: {ctx.chan.uid}\n'
f'ctx id: {ctx.cid}'
)
self._closed = True
# Do we close the local mem chan ``self._rx_chan`` ??!?
@ -271,7 +279,8 @@ class ReceiveMsgStream(trio.abc.ReceiveChannel):
self,
) -> AsyncIterator[BroadcastReceiver]:
'''Allocate and return a ``BroadcastReceiver`` which delegates
'''
Allocate and return a ``BroadcastReceiver`` which delegates
to this message stream.
This allows multiple local tasks to receive each their own copy
@ -593,23 +602,23 @@ class Context:
async with MsgStream(
ctx=self,
rx_chan=ctx._recv_chan,
) as rchan:
) as stream:
if self._portal:
self._portal._streams.add(rchan)
self._portal._streams.add(stream)
try:
self._stream_opened = True
# ensure we aren't cancelled before delivering
# the stream
# XXX: do we need this?
# ensure we aren't cancelled before yielding the stream
# await trio.lowlevel.checkpoint()
yield rchan
yield stream
# XXX: Make the stream "one-shot use". On exit, signal
# NOTE: Make the stream "one-shot use". On exit, signal
# ``trio.EndOfChannel``/``StopAsyncIteration`` to the
# far end.
await self.send_stop()
await stream.aclose()
finally:
if self._portal:

View File

@ -133,13 +133,13 @@ async def gather_contexts(
# deliver control once all managers have started up
await all_entered.wait()
# NOTE: order *should* be preserved in the output values
# since ``dict``s are now implicitly ordered.
yield tuple(unwrapped.values())
# we don't need a try/finally since cancellation will be triggered
# by the surrounding nursery on error.
parent_exit.set()
try:
yield tuple(unwrapped.values())
finally:
# NOTE: this is ABSOLUTELY REQUIRED to avoid
# the following wacky bug:
# <tractorbugurlhere>
parent_exit.set()
# Per actor task caching helpers.