execlib/execlog/router.py

543 lines
23 KiB
Python

'''
Router
'''
import time
import asyncio
import logging
import inspect
import traceback
import threading
from pathlib import Path
from typing import Callable
from functools import partial
from colorama import Fore, Style
from collections import defaultdict
from concurrent.futures import ThreadPoolExecutor, wait, as_completed
from tqdm.auto import tqdm
from execlog.event import Event
from execlog.listener import Listener
logger = logging.getLogger(__name__)
class Router[E: Event]:
'''
Route events to registered callbacks
Generalized registration includes an endpoint (the origin of an event), a pattern (to
filter events at the endpoint), and a callback (to be executed if pattern is matched).
The Router _routes_ events to affiliated callbacks in a multi-threaded fashion. A
thread pool handles these jobs as events are submitted, typically by a composing
Listener. The Listener "hears" an event, and passes it on through to a Router to
further filter and delegate any matching follow-up jobs.
This base Router implements most of the registry and filter model. When events are
submitted for propagation, they are checked for matching routes. Events specify an
origin endpoint, which is used as the filter for attached routes. The event is then
subjected to the `filter` method, which checks if the event matches the registered
`pattern` under the originated `endpoint`. If so, the callback is scheduled for
execution, and the matching event is passed as its sole argument.
Subclasses are expected to implement (at least) the `filter` method. This function is
responsible for wrapping up the task-specific logic needed to determine if an event,
originating from a known endpoint, matches the callback-specific pattern. This method
needn't handle any other filter logic, like checking if the event originates from the
provided endpoint, as this is already handled by the outer look in `matching_routes`.
`get_listener` is a convenience method that instantiates and populates an affiliated
Listener over the register paths found in the Router. Listeners require a Router upon
instantiation so events can be propagated to available targets when they occur.
`get_listener()` is the recommended way to attain a Listener.
Note: on debouncing events
Previously, debouncing was handled by listeners. This logic has been generalized
and moved to this class, as it's general enough to be desired across various
Listener types. We also need unique, identifying info only available with a
`(endpoint, callback, pattern)` triple in order to debounce events in accordance
with their intended target.
Note: tracking events and serializing callback frames
Although not part of the original implementation, we now track which events have a
callback chain actively being executed, and prevent the same chain from being
started concurrently. If the callback chain is actively running for an event, and
that same event is submitted before this chain finishes, the request is simply
enqueued. The `clear_event` method is attached as a "done callback" to each job
future, and will re-submit the event once the active chain finishes.
While this could be interpreted as a harsh design choice, it helps prevent many
many thread conflicts (race conditions, writing to the same resources, etc) when
the same function is executed concurrently, many times over. Without waiting
completely for an event to be fully handled, later jobs may complete before
earlier ones, or interact with intermediate disk states (raw file writes, DB
inserts, etc), before the earliest call has had a chance to clean up.
'''
listener_cls = Listener[E]
def __init__(self, loop=None, workers=None):
'''
Parameters:
loop:
workers: number of workers to assign the thread pool when the event loop is
started. Defaults to `None`, which, when passed to
ThreadPoolExecutor, will by default use 5x the number of available
processors on the machine (which the docs claim is a reasonable
assumption given threads are more commonly leveraged for I/O work
rather than intense CPU operations). Given the intended context for
this class, this assumption aligns appropriately.
'''
self.loop = loop
self.workers = workers
self.routemap : dict[str, list[tuple]] = defaultdict(list)
self.post_callbacks = []
# track running jobs by event
self.running_events = defaultdict(set)
# debounce tracker
self.next_allowed_time = defaultdict(int)
# store prepped (e.g., delayed) callbacks
self.callback_registry = {}
# track event history
self.event_log = []
self._thread_pool = None
self._route_lock = threading.Lock()
@property
def thread_pool(self):
if self._thread_pool is None:
self._thread_pool = ThreadPoolExecutor(max_workers=self.workers)
return self._thread_pool
def register(
self,
endpoint,
callback: Callable,
pattern,
debounce=200,
delay=10,
**listener_kwargs,
):
'''
Register a route.
Note: Listener arguments
Notice how listener_kwargs are accumulated instead of uniquely assigned to an
endpoint. This is generally acceptable as some listeners may allow various
configurations for the same endpoint. Note, however, for something like the
PathListener, this will have no effect. Registering the same endpoint multiple
times won't cause any errors, but the configuration options will only remain
for the last registered group.
(Update) The above remark about PathListener's is no longer, and likely never
was. Varying flag sets under the same endpoint do in fact have a cumulative
effect, and we need to be able disentangle events accordingly through
submitted event's `action` value.
Parameters:
endpoint:
callback: callable accepting an event to be executed if when a matching event
is received
pattern: hashable object to be used when filtering event (passed to inherited
`filter(...)`)
debounce:
delay:
'''
route_tuple = (callback, pattern, debounce, delay, listener_kwargs)
self.routemap[endpoint].append(route_tuple)
def submit(self, events:E | list[E], callbacks:list[Callable]|None=None):
'''
Handle a list of events. Each event is matched against the registered callbacks,
and those callbacks are ran concurrently (be it via a thread pool or an asyncio
loop).
'''
if type(events) is not list:
events = [events]
futures = []
for event in events:
future = self.submit_callback(self.submit_event, event, callbacks=callbacks)
future.add_done_callback(lambda f: self.clear_event(event, f))
futures.append(future)
return futures
def submit_event(self, event: E, callbacks:list[Callable]|None=None):
'''
Group up and submit all matching callbacks for `event`. All callbacks are ran
concurrently in their own threads, and this method blocks until all are completed.
In the outer `submit` context, this blocking method is itself ran in its own
thread, and the registered post-callbacks are attached to the completion of this
function, i.e., the finishing of all callbacks matching provided event.
Note that an event may not match any routes, in which case the method exits early.
An empty list is returned, and this shows up as the outer future's result. In this
case, the event is never considered "running," and the non-result picked up in
`clear_event` will ensure it exits right away (not even attempting to pop the
event from the running list, and for now not tracking it in the event log).
'''
if callbacks is None:
# ensure same thread gets all matching routes & sets debounce updates; else
# this may be split across threads mid-check, preventing one thread from
# handling the blocking of the entire group
with self._route_lock:
callbacks = self.matching_routes(event)
# stop early if no work to do
if len(callbacks) == 0:
return []
# enqueue requested/matched callbacks and exit if running
event_idx = self.event_index(event)
if event_idx in self.running_events:
self.queue_callbacks(event_idx, callbacks)
return []
# callbacks now computed, flush the running event
# note: a separate thread could queue valid callbacks since the running check;
# o/w we know the running index is empty
self.running_events[event_idx] = self.running_events[event_idx]
# submit matching callbacks and wait for them to complete
future_results = self.wait_on_callbacks(callbacks, event)
# finally call post event-group callbacks (only if some event callbacks were
# submitted), wait for them to complete
if future_results:
self.wait_on_futures([
self.submit_callback(post_callback, event, future_results)
for post_callback in self.post_callbacks
])
return future_results
def submit_callback(self, callback: Callable, *args, **kwargs):
'''
Note: this method is expected to return a future. Perform any event-based
filtering before submitting a callback with this method.
'''
if inspect.iscoroutinefunction(callback):
if self.loop is None:
self.loop = asyncio.new_event_loop()
asyncio.set_event_loop(self.loop)
#loop.run_in_executor(executor, loop.create_task, callback(event))
#future = self.loop.call_soon_threadsafe(
# self.loop.create_task,
future = asyncio.run_coroutine_threadsafe(
callback(*args, **kwargs),
self.loop,
)
else:
future = self.thread_pool.submit(
callback, *args, **kwargs
)
future.add_done_callback(handle_exception)
return future
def matching_routes(self, event: E, event_time=None):
'''
Return eligible matching routes for the provided event.
Note that we wait as late as possible before enqueuing matches if the event is in
fact already active in a frame. If this method were start filtering results while
the frame is active, and the frame were to finish before all matching callbacks
were determined, we would be perfectly happy to return all matches, and allow the
outer `submit_event` context to run them right away in a newly constructed frame.
The _very_ next thing that gets done is adding this event to the active event
tracker. Otherwise, matching is performed as usual, and eligible callbacks are
simply enqueued for the next event frame, which will be checked in the "done"
callback of the active frame. The logic here should mostly "seal up" any real
opportunities for error, e.g., a frame ending and popping off elements from
`running_events` half-way through their inserting at the end of this method, or
multiple threads checking for matching routes for the same event, and both coming
away with a non-empty set of matches to run. That last example highlights
precisely how the single event-frame model works: many threads might be running
this method at the same time, for the same event (which has fired rapidly), but
only one should be able to "secure the frame" and begin running the matching
callbacks. Making the "active frame check" both as late as possible and as close
to the event blocking stage in the tracker (in `submit_event`), we make the
ambiguity gap as small as possible (and almost certainly smaller than any
realistic I/O-bound event duplication).
Note: on event actions
The debounce reset is now only set if the event is successfully filtered. This
allows some middle ground when trying to depend on event actions: if the
action passes through, we block the whole range of actions until the debounce
window completes. Otherwise, the event remains open, only to be blocked by the
debounce on the first matching action.
'''
matches = []
endpoint = event.endpoint
name = event.name
#action = tuple(event.action) # should be more general
event_time = time.time()*1000 if event_time is None else event_time
for (callback, pattern, debounce, delay, listen_kwargs) in self.routemap[endpoint]:
#index = (endpoint, name, action, callback, pattern, debounce, delay)
index = (endpoint, name, callback, pattern, debounce, delay)
if event_time < self.next_allowed_time[index]:
# reject event
continue
if self.filter(event, pattern, **listen_kwargs):
# note that delayed callbacks are added
matches.append(self.get_delayed_callback(callback, delay, index))
# set next debounce
self.next_allowed_time[index] = event_time + debounce
match_text = Style.BRIGHT + Fore.GREEN + 'matched' + Fore.RESET
callback_name = str(callback)
if hasattr(callback, '__name__'):
callback_name = callback.__name__
logger.info(
f'Event [{name}] {match_text} [{pattern}] under [{endpoint}] for [{callback_name}]'
)
else:
match_text = Style.BRIGHT + Fore.RED + 'rejected' + Fore.RESET
logger.debug(
f'Event [{name}] {match_text} against [{pattern}] under [{endpoint}] for [{callback.__name__}]'
)
return matches
def get_delayed_callback(self, callback: Callable, delay: int|float, index):
'''
Parameters:
callback: function to wrap
delay: delay in ms
'''
if index not in self.callback_registry:
async def async_wrap(callback, *args, **kwargs):
await asyncio.sleep(delay/1000)
return await callback(*args, **kwargs)
def sync_wrap(callback, *args, **kwargs):
time.sleep(delay/1000)
return callback(*args, **kwargs)
wrapper = None
if inspect.iscoroutinefunction(callback): wrapper = async_wrap
else: wrapper = sync_wrap
self.callback_registry[index] = partial(wrapper, callback)
return self.callback_registry[index]
def wait_on_futures(self, futures):
'''
Block until all futures in `futures` are complete. Return collected results as a
list, and log warnings when a future fails.
'''
future_results = []
for future in as_completed(futures):
try:
future_results.append(future.result())
except Exception as e:
logger.warning(f"Router callback job failed with exception {e}")
return future_results
def wait_on_callbacks(self, callbacks: list[Callable], event: E, *args, **kwargs):
'''
Overridable by inheriting classes based on callback structure
'''
return self.wait_on_futures([
self.submit_callback(callback, event, *args, **kwargs)
for callback in callbacks
])
def queue_callbacks(self, event_idx, callbacks: list[Callable]):
'''
Overridable by inheriting classes based on callback structure
'''
self.running_events[event_idx].update(callbacks)
def filter(self, event: E, pattern, **listen_kwargs) -> bool:
'''
Determine if a given event matches the providedpattern
Parameters:
event:
pattern:
listen_kwargs:
'''
raise NotImplementedError
def add_post_callback(self, callback: Callable):
self.post_callbacks.append(callback)
def get_listener(self, listener_cls=None):
'''
Create a new Listener to manage watched routes and their callbacks.
'''
if listener_cls is None:
listener_cls = self.listener_cls
if listener_cls is None:
raise ValueError('No Listener class provided')
listener = listener_cls(self)
return self.extend_listener(listener)
def extend_listener(self, listener):
'''
Extend a provided Listener object with the Router instance's `listener_kwargs`.
'''
for endpoint, route_tuples in self.routemap.items():
for route_tuple in route_tuples:
listen_kwargs = route_tuple[-1]
listener.listen(endpoint, **listen_kwargs)
return listener
def stop_event(self, event):
'''
Pop event out of the running events tracker and return it.
'''
event_idx = self.event_index(event)
return self.running_events.pop(event_idx, None)
def clear_event(self, event: E, future):
'''
Clear an event. Pops the passed event out of `running_events`, and the request
counter is >0, the event is re-submitted.
This method is attached as a "done" callback to the main event wrapping job
`submit_event`. The `future` given to this method is one to which it was
attached as this "done" callback. This method should only be called when that
`future` is finished running (or failed). If any jobs were submitted in the
wrapper task, the future results here should be non-empty. We use this fact to
filter out non-work threads that call this method. Because even the
`matching_routes` check is threaded, we can't wait to see an event has no work to
schedule, and thus can't prevent this method being attached as a "done" callback.
The check for results from the passed future allows us to know when in fact a
valid frame has finished, and a resubmission may be on the table.
'''
result = future.result()
if not result: return
self.event_log.append((event, result))
queued_callbacks = self.stop_event(event)
# resubmit event if some queued work remains
if queued_callbacks and len(queued_callbacks) > 0:
logger.debug(
f'Event [{event.name}] resubmitted with [{len(queued_callbacks)}] queued callbacks'
)
self.submit(event, callbacks=queued_callbacks)
def event_index(self, event):
return event[:2]
class ChainRouter[E: Event](Router[E]):
'''
Routes events to registered callbacks
'''
def __init__(self, ordered_routers):
super().__init__()
self.ordered_routers = []
for router in ordered_routers:
self.add_router(router)
self.running_events = defaultdict(lambda: defaultdict(set))
def add_router(self, router):
'''
TODO: allow positional insertion in ordered list
Note: the `routemap` extensions here shouldn't be necessary, since 1) route maps
show up only in `matching_routes`, and 2) `matching_routes` is only invoked in
`submit_event`, which is totally overwritten for the ChainRouter type. All events
are routed through to individual Routers, and which point their route maps are
used.
'''
self.ordered_routers.append(router)
for endpoint, routelist in router.routemap.items():
self.routemap[endpoint].extend(routelist)
def matching_routes(self, event: E, event_time=None):
'''
Colloquial `callbacks` now used as a dict of lists of callbacks, indexed by
router, and only having keys for routers with non-empty callback lists.
'''
if event_time is None:
event_time = time.time()*1000
route_map = {}
for router in self.ordered_routers:
router_matches = router.matching_routes(event, event_time)
if router_matches:
route_map[router] = router_matches
return route_map
def wait_on_callbacks(self, callbacks, event: E, *args, **kwargs):
'''
Note: relies on order of callbacks dict matching that of `ordered_routers`, which
should happen in `matching_routes`
'''
results = {}
for router, callback_list in callbacks.items():
router_results = router.submit_event(event, callbacks=callback_list)
results[router] = router_results
return results
def queue_callbacks(self, event_idx, callbacks):
for router, callback_list in callbacks.items():
self.running_events[event_idx][router].update(callback_list)
def stop_event(self, event):
'''
Sub-routers do not get a "done" callback for their `submit_event` jobs, as they
would if they handled their own event submissions. They will, however, set the
submitted event as "running." We can't rely on sub-routers' "done" callbacks to
"unset" the running event, because the disconnect between the thread completing
and execution of that callback may take too long.
Instead, we explicitly unset the running event for each of the constituent
sub-routers at the _same time_ we handle the ChainRouter's notion of event's
ending.
'''
event_idx = self.event_index(event)
for router in self.ordered_routers:
rq_callbacks = router.running_events.pop(event_idx, [])
assert len(rq_callbacks) == 0
return self.running_events.pop(event_idx, None)
def get_listener(self, listener_cls=None):
if listener_cls is None:
for router in self.ordered_routers:
if router.listener_cls is not None:
listener_cls = router.listener_cls
break
listener = super().get_listener(listener_cls)
for router in self.ordered_routers:
router.extend_listener(listener)
return listener
def handle_exception(future):
try:
future.result()
except Exception as e:
print(f"Exception occurred: {e}")
traceback.print_exc()