Mv over `ServiceMngr` from `piker` with mods

Namely distinguishing service "IPC contexts" (opened in a
subactor via a `Portal`) from just local `trio.Task`s started
and managed under the `.service_n` (more or less wrapping in the
interface of a "task-manager" style nursery - aka a one-cancels-one
supervision start).

API changes from original (`piker`) impl,
- mk `.start_service_task()` do ONLY that, start a task with a wrapping
  cancel-scope and completion event.
  |_ ideally this gets factored-out/re-implemented using the
    task-manager/OCO-style-nursery from GH #363.
- change what was the impl of `.start_service_task()` to `.start_service_ctx()`
  since it more explicitly defines the functionality of entering
  `Portal.open_context()` with a wrapping cs and completion event inside
  a bg task (which syncs the ctx's lifetime with termination of the
  remote actor runtime).
- factor out what was a `.start_service_ctx()` closure to a new
  `_open_and_supervise_service_ctx()` mod-func holding the meat of
  the supervision logic.

`ServiceMngr` API brief,
- use `open_service_mngr()` and `get_service_mngr()` to acquire the
  actor-global singleton.
- `ServiceMngr.start_service()` and `.cancel_service()` which allow for
  straight forward mgmt of "service subactor daemons".
hilevel_serman
Tyler Goodlet 2024-12-11 12:38:35 -05:00
parent f08e888138
commit 46dbe6d2fc
2 changed files with 453 additions and 21 deletions

View File

@ -0,0 +1,26 @@
# tractor: structured concurrent "actors".
# Copyright 2024-eternity Tyler Goodlet.
# This program is free software: you can redistribute it and/or modify
# it under the terms of the GNU Affero General Public License as published by
# the Free Software Foundation, either version 3 of the License, or
# (at your option) any later version.
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# GNU Affero General Public License for more details.
# You should have received a copy of the GNU Affero General Public License
# along with this program. If not, see <https://www.gnu.org/licenses/>.
'''
High level design patterns, APIs and runtime extensions built on top
of the `tractor` runtime core.
'''
from ._service import (
open_service_mngr as open_service_mngr,
get_service_mngr as get_service_mngr,
ServiceMngr as ServiceMngr,
)

View File

@ -21,18 +21,26 @@ and API.
'''
from __future__ import annotations
from contextlib import (
# asynccontextmanager as acm,
contextmanager as cm,
asynccontextmanager as acm,
# contextmanager as cm,
)
from collections import defaultdict
from dataclasses import (
dataclass,
field,
)
import functools
import inspect
from typing import (
Callable,
Any,
)
import tractor
import trio
from trio import TaskStatus
from tractor import (
log,
ActorNursery,
current_actor,
ContextCancelled,
@ -40,9 +48,7 @@ from tractor import (
Portal,
)
from ._util import (
log, # sub-sys logger
)
log = log.get_logger('tractor')
# TODO: implement a `@singleton` deco-API for wrapping the below
@ -93,11 +99,30 @@ from ._util import (
# TODO: implement a singleton deco-API for wrapping the below
# factory's impl for general actor-singleton use?
#
# @singleton
# async def open_service_mngr(
# **init_kwargs,
# ) -> ServiceMngr:
# '''
# Note this function body is invoke IFF no existing singleton instance already
# exists in this proc's memory.
# '''
# # setup
# yield ServiceMngr(**init_kwargs)
# # teardown
# TODO: singleton factory API instead of a class API
@cm
def open_service_mngr(
@acm
async def open_service_mngr(
*,
_singleton: list[ServiceMngr|None] = [None],
debug_mode: bool = False,
# NOTE; since default values for keyword-args are effectively
# module-vars/globals as per the note from,
# https://docs.python.org/3/tutorial/controlflow.html#default-argument-values
@ -106,27 +131,408 @@ def open_service_mngr(
# a difference when the default is a mutable object such as
# a list, dictionary, or instances of most classes"
#
_singleton: list[ServiceMngr|None] = [None],
**init_kwargs,
) -> ServiceMngr:
'''
Open a multi-subactor-as-service-daemon tree supervisor.
Open an actor-global "service-manager" for supervising a tree
of subactors and/or actor-global tasks.
The delivered `ServiceMngr` is a singleton instance for each
actor-process and is allocated on first open and never
The delivered `ServiceMngr` is singleton instance for each
actor-process, that is, allocated on first open and never
de-allocated unless explicitly deleted by al call to
`del_service_mngr()`.
'''
# TODO: factor this an allocation into
# a `._mngr.open_service_mngr()` and put in the
# once-n-only-once setup/`.__aenter__()` part!
# -[ ] how to make this only happen on the `mngr == None` case?
# |_ use `.trionics.maybe_open_context()` (for generic
# async-with-style-only-once of the factory impl, though
# what do we do for the allocation case?
# / `.maybe_open_nursery()` (since for this specific case
# it's simpler?) to activate
async with (
tractor.open_nursery() as an,
trio.open_nursery() as tn,
):
# impl specific obvi..
init_kwargs.update({
'actor_n': an,
'service_n': tn,
})
mngr: ServiceMngr|None
if (mngr := _singleton[0]) is None:
log.info('Allocating a new service mngr!')
mngr = _singleton[0] = ServiceMngr(**init_kwargs)
# TODO: put into `.__aenter__()` section of
# eventual `@singleton_acm` API wrapper.
#
# assign globally for future daemon/task creation
mngr.actor_n = an
mngr.service_n = tn
else:
assert (
mngr.actor_n
and
mngr.service_tn
)
log.info(
'Using extant service mngr!\n\n'
f'{mngr!r}\n' # it has a nice `.__repr__()` of services state
)
with mngr:
try:
# NOTE: this is a singleton factory impl specific detail
# which should be supported in the condensed
# `@singleton_acm` API?
mngr.debug_mode = debug_mode
yield mngr
finally:
# TODO: is this more clever/efficient?
# if 'samplerd' in mngr.service_tasks:
# await mngr.cancel_service('samplerd')
tn.cancel_scope.cancel()
def get_service_mngr() -> ServiceMngr:
'''
Try to get the singleton service-mngr for this actor presuming it
has already been allocated using,
.. code:: python
async with open_<@singleton_acm(func)>() as mngr`
... this block kept open ...
If not yet allocated raise a `ServiceError`.
'''
# https://stackoverflow.com/a/12627202
# https://docs.python.org/3/library/inspect.html#inspect.Signature
maybe_mngr: ServiceMngr|None = inspect.signature(
open_service_mngr
).parameters['_singleton'].default[0]
if maybe_mngr is None:
raise RuntimeError(
'Someone must allocate a `ServiceMngr` using\n\n'
'`async with open_service_mngr()` beforehand!!\n'
)
return maybe_mngr
# TODO: we need remote wrapping and a general soln:
# - factor this into a ``tractor.highlevel`` extension # pack for the
# library.
# - wrap a "remote api" wherein you can get a method proxy
# to the pikerd actor for starting services remotely!
# - prolly rename this to ActorServicesNursery since it spawns
# new actors and supervises them to completion?
@dataclass
class ServiceMngr:
'''
A multi-subactor-as-service manager.
Spawn, supervise and monitor service/daemon subactors in a SC
process tree.
'''
actor_n: ActorNursery
service_n: trio.Nursery
debug_mode: bool = False # tractor sub-actor debug mode flag
service_tasks: dict[
str,
tuple[
trio.CancelScope,
Context,
Portal,
trio.Event,
]
] = field(default_factory=dict)
# internal per-service task mutexs
_locks = defaultdict(trio.Lock)
# TODO, unify this interface with our `TaskManager` PR!
#
#
async def start_service_task(
self,
name: str,
# TODO: typevar for the return type of the target and then
# use it below for `ctx_res`?
fn: Callable,
allow_overruns: bool = False,
**ctx_kwargs,
) -> tuple[
trio.CancelScope,
Any,
trio.Event,
]:
async def _task_manager_start(
task_status: TaskStatus[
tuple[
trio.CancelScope,
trio.Event,
]
] = trio.TASK_STATUS_IGNORED,
) -> Any:
task_cs = trio.CancelScope()
task_complete = trio.Event()
with task_cs as cs:
task_status.started((
cs,
task_complete,
))
try:
await fn()
except trio.Cancelled as taskc:
log.cancel(
f'Service task for `{name}` was cancelled!\n'
# TODO: this would be a good spot to use
# a respawn feature Bo
)
raise taskc
finally:
task_complete.set()
(
cs,
complete,
) = await self.service_n.start(_task_manager_start)
# store the cancel scope and portal for later cancellation or
# retstart if needed.
self.service_tasks[name] = (
cs,
None,
None,
complete,
)
return (
cs,
complete,
)
async def start_service_ctx(
self,
name: str,
portal: Portal,
# TODO: typevar for the return type of the target and then
# use it below for `ctx_res`?
target: Callable,
**ctx_kwargs,
) -> tuple[
trio.CancelScope,
Context,
Any,
]:
cs, sub_ctx, complete, started = await self.service_n.start(
functools.partial(
self._open_and_supervise_service_ctx,
name=name,
target=target,
portal=portal,
**ctx_kwargs,
)
)
# store the cancel scope and portal for later cancellation or
# retstart if needed.
self.service_tasks[name] = (cs, sub_ctx, portal, complete)
return (
cs,
sub_ctx,
started,
)
async def _open_and_supervise_service_ctx(
self,
name: str,
target: Callable, # TODO, type for `@tractor.context` requirement
portal: Portal,
allow_overruns: bool = False,
task_status: TaskStatus[
tuple[
trio.CancelScope,
Context,
trio.Event,
Any,
]
] = trio.TASK_STATUS_IGNORED,
**ctx_kwargs,
) -> Any:
'''
Open a context in a service sub-actor, add to a stack
that gets unwound at ``pikerd`` teardown.
This allows for allocating long-running sub-services in our main
daemon and explicitly controlling their lifetimes.
'''
# TODO: use the ctx._scope directly here instead?
# -[ ] actually what semantics do we expect for this
# usage!?
with trio.CancelScope() as cs:
try:
async with portal.open_context(
target,
allow_overruns=allow_overruns,
**ctx_kwargs,
) as (ctx, started):
# unblock once the remote context has started
complete = trio.Event()
task_status.started((
cs,
ctx,
complete,
started,
))
log.info(
f'`pikerd` service {name} started with value {started}'
)
# wait on any context's return value
# and any final portal result from the
# sub-actor.
ctx_res: Any = await ctx.wait_for_result()
# NOTE: blocks indefinitely until cancelled
# either by error from the target context
# function or by being cancelled here by the
# surrounding cancel scope.
return (
await portal.wait_for_result(),
ctx_res,
)
except ContextCancelled as ctxe:
canceller: tuple[str, str] = ctxe.canceller
our_uid: tuple[str, str] = current_actor().uid
if (
canceller != portal.chan.uid
and
canceller != our_uid
):
log.cancel(
f'Actor-service `{name}` was remotely cancelled by a peer?\n'
# TODO: this would be a good spot to use
# a respawn feature Bo
f'-> Keeping `pikerd` service manager alive despite this inter-peer cancel\n\n'
f'cancellee: {portal.chan.uid}\n'
f'canceller: {canceller}\n'
)
else:
raise
finally:
# NOTE: the ctx MUST be cancelled first if we
# don't want the above `ctx.wait_for_result()` to
# raise a self-ctxc. WHY, well since from the ctx's
# perspective the cancel request will have
# arrived out-out-of-band at the `Actor.cancel()`
# level, thus `Context.cancel_called == False`,
# meaning `ctx._is_self_cancelled() == False`.
# with trio.CancelScope(shield=True):
# await ctx.cancel()
await portal.cancel_actor() # terminate (remote) sub-actor
complete.set() # signal caller this task is done
self.service_tasks.pop(name) # remove mngr entry
async def cancel_service(
self,
name: str,
) -> Any:
'''
Cancel the service task and actor for the given ``name``.
'''
log.info(f'Cancelling `pikerd` service {name}')
cs, sub_ctx, portal, complete = self.service_tasks[name]
# cs.cancel()
await sub_ctx.cancel()
await complete.wait()
if name in self.service_tasks:
# TODO: custom err?
# raise ServiceError(
raise RuntimeError(
f'Service actor for {name} not terminated and/or unknown?'
)
# assert name not in self.service_tasks, \
# f'Serice task for {name} not terminated?'
async def start_service(
self,
daemon_name: str,
ctx_ep: Callable, # kwargs must `partial`-ed in!
# ^TODO, type for `@tractor.context` deco-ed funcs!
debug_mode: bool = False,
**tractor_actor_kwargs,
) -> Context:
'''
Start a "service" task in a (new) sub-actor and manage its
lifetime indefinitely until termination.
Service actors can be cancelled (and thus shutdown) using
`.cancel_service()`.
'''
entry: tuple|None = self.service_tasks.get(daemon_name)
if entry:
(cs, sub_ctx, portal, complete) = entry
return sub_ctx
if daemon_name not in self.service_tasks:
portal: Portal = await self.actor_n.start_actor(
daemon_name,
debug_mode=( # maybe set globally during allocate
debug_mode
or
self.debug_mode
),
**tractor_actor_kwargs,
)
ctx_kwargs: dict[str, Any] = {}
if isinstance(ctx_ep, functools.partial):
ctx_kwargs: dict[str, Any] = ctx_ep.keywords
ctx_ep: Callable = ctx_ep.func
(
cs,
sub_ctx,
started,
) = await self.start_service_ctx(
name=daemon_name,
portal=portal,
target=ctx_ep,
**ctx_kwargs,
)
return sub_ctx