takahe/stator/runner.py

299 wiersze
11 KiB
Python

import datetime
import logging
import os
import signal
import time
import uuid
from concurrent.futures import Future, ThreadPoolExecutor
from django.conf import settings
from django.db import close_old_connections
from django.utils import timezone
from core import sentry
from core.models import Config
from stator.models import StatorModel, Stats
logger = logging.getLogger(__name__)
class LoopingTimer:
"""
Triggers check() to be true once every `interval`.
"""
next_run: float | None = None
def __init__(self, interval: float, trigger_at_start=True):
self.interval = interval
self.trigger_at_start = trigger_at_start
def check(self) -> bool:
# See if it's our first time being called
if self.next_run is None:
# Set up the next call based on trigger_at_start
if self.trigger_at_start:
self.next_run = time.monotonic()
else:
self.next_run = time.monotonic() + self.interval
# See if it's time to run the next call
if time.monotonic() >= self.next_run:
self.next_run = time.monotonic() + self.interval
return True
return False
class StatorRunner:
"""
Runs tasks on models that are looking for state changes.
Designed to run either indefinitely, or just for a few seconds.
"""
def __init__(
self,
models: list[type[StatorModel]],
concurrency: int = getattr(settings, "STATOR_CONCURRENCY", 30),
concurrency_per_model: int = getattr(
settings, "STATOR_CONCURRENCY_PER_MODEL", 15
),
liveness_file: str | None = None,
schedule_interval: int = 60,
delete_interval: int = 30,
lock_expiry: int = 300,
run_for: int = 0,
):
self.models = models
self.runner_id = uuid.uuid4().hex
self.concurrency = concurrency
self.concurrency_per_model = concurrency_per_model
self.liveness_file = liveness_file
self.schedule_interval = schedule_interval
self.delete_interval = delete_interval
self.lock_expiry = lock_expiry
self.run_for = run_for
self.minimum_loop_delay = 0.5
self.maximum_loop_delay = 5
self.tasks: dict[tuple[str, str], Future] = {}
# Set up SIGALRM handler
signal.signal(signal.SIGALRM, self.alarm_handler)
def run(self):
sentry.set_takahe_app("stator")
self.handled = {}
self.started = time.monotonic()
self.executor = ThreadPoolExecutor(max_workers=self.concurrency)
self.loop_delay = self.minimum_loop_delay
self.scheduling_timer = LoopingTimer(self.schedule_interval)
self.deletion_timer = LoopingTimer(self.delete_interval)
# For the first time period, launch tasks
logger.info("Running main task loop")
try:
with sentry.configure_scope() as scope:
while True:
# See if we need to run cleaning
if self.scheduling_timer.check():
# Set up the watchdog timer (each time we do this the previous one is cancelled)
signal.alarm(self.schedule_interval * 2)
# Write liveness file if configured
if self.liveness_file:
with open(self.liveness_file, "w") as fh:
fh.write(str(int(time.time())))
# Refresh the config
self.load_config()
# Do scheduling (stale lock deletion and stats gathering)
self.run_scheduling()
# Clear the cleaning breadcrumbs/extra for the main part of the loop
sentry.scope_clear(scope)
self.clean_tasks()
# See if we need to add deletion tasks
if self.deletion_timer.check():
self.add_deletion_tasks()
# Fetch and run any new handlers we can fit
self.add_transition_tasks()
# Are we in limited run mode?
if (
self.run_for
and (time.monotonic() - self.started) > self.run_for
):
break
# Prevent busylooping, but also back off delay if we have
# no tasks
if self.tasks:
self.loop_delay = self.minimum_loop_delay
else:
self.loop_delay = min(
self.loop_delay * 1.5,
self.maximum_loop_delay,
)
time.sleep(self.loop_delay)
# Clear the Sentry breadcrumbs and extra for next loop
sentry.scope_clear(scope)
except KeyboardInterrupt:
pass
# Wait for tasks to finish
logger.info("Waiting for tasks to complete")
self.executor.shutdown()
# We're done
logger.info("Complete")
def alarm_handler(self, signum, frame):
"""
Called when SIGALRM fires, which means we missed a schedule loop.
Just exit as we're likely deadlocked.
"""
logger.warning("Watchdog timeout exceeded")
os._exit(2)
def load_config(self):
"""
Refreshes config from the DB
"""
Config.system = Config.load_system()
def run_scheduling(self):
"""
Deletes stale locks for models, and submits their stats.
"""
with sentry.start_transaction(op="task", name="stator.run_scheduling"):
for model in self.models:
with sentry.start_span(description=model._meta.label_lower):
num = self.handled.get(model._meta.label_lower, 0)
if num or settings.DEBUG:
logger.info(
f"{model._meta.label_lower}: Scheduling ({num} handled)"
)
self.submit_stats(model)
model.transition_clean_locks()
def submit_stats(self, model: type[StatorModel]):
"""
Pop some statistics into the database from our local info for the given model
"""
stats_instance = Stats.get_for_model(model)
if stats_instance.model_label in self.handled:
stats_instance.add_handled(self.handled[stats_instance.model_label])
del self.handled[stats_instance.model_label]
stats_instance.set_queued(model.transition_ready_count())
stats_instance.trim_data()
stats_instance.save()
def add_transition_tasks(self, call_inline=False):
"""
Adds a transition thread for as many instances as we can, given capacity
and batch size limits.
"""
# Calculate space left for tasks
space_remaining = self.concurrency - len(self.tasks)
# Fetch new tasks
for model in self.models:
if space_remaining > 0:
for instance in model.transition_get_with_lock(
number=min(space_remaining, self.concurrency_per_model),
lock_expiry=(
timezone.now() + datetime.timedelta(seconds=self.lock_expiry)
),
):
key = (model._meta.label_lower, instance.pk)
# Don't run two threads for the same thing
if key in self.tasks:
continue
if call_inline:
task_transition(instance, in_thread=False)
else:
self.tasks[key] = self.executor.submit(
task_transition, instance
)
self.handled[model._meta.label_lower] = (
self.handled.get(model._meta.label_lower, 0) + 1
)
space_remaining -= 1
def add_deletion_tasks(self, call_inline=False):
"""
Adds a deletion thread for each model
"""
# Yes, this potentially goes over the capacity limit - it's fine.
for model in self.models:
if model.state_graph.deletion_states:
if call_inline:
task_deletion(model, in_thread=False)
else:
self.tasks[
model._meta.label_lower, "__delete__"
] = self.executor.submit(task_deletion, model)
def clean_tasks(self):
"""
Removes any tasks that are done and handles exceptions if they
raised them.
"""
for key, task in list(self.tasks.items()):
if task.done():
del self.tasks[key]
try:
task.result()
except BaseException as e:
logger.exception(e)
def run_single_cycle(self):
"""
Testing entrypoint to advance things just one cycle, and allow errors
to propagate out.
"""
self.add_deletion_tasks(call_inline=True)
self.add_transition_tasks(call_inline=True)
def task_transition(instance: StatorModel, in_thread: bool = True):
"""
Runs one state transition/action.
"""
task_name = f"stator.task_transition:{instance._meta.label_lower}#{{id}} from {instance.state}"
started = time.monotonic()
with sentry.start_transaction(op="task", name=task_name):
sentry.set_context(
"instance",
{
"model": instance._meta.label_lower,
"pk": instance.pk,
"state": instance.state,
"state_age": instance.state_age,
},
)
result = instance.transition_attempt()
duration = time.monotonic() - started
if result:
logger.info(
f"{instance._meta.label_lower}: {instance.pk}: {instance.state} -> {result} ({duration:.2f}s)"
)
else:
logger.info(
f"{instance._meta.label_lower}: {instance.pk}: {instance.state} unchanged ({duration:.2f}s)"
)
if in_thread:
close_old_connections()
def task_deletion(model: type[StatorModel], in_thread: bool = True):
"""
Runs one model deletion set.
"""
# Loop, running deletions every second, until there are no more to do
while True:
deleted = model.transition_delete_due()
if not deleted:
break
logger.info(f"{model._meta.label_lower}: Deleted {deleted} stale items")
time.sleep(1)
if in_thread:
close_old_connections()