watcha-synapse/synapse/metrics/background_process_metrics.py

#
# This file is licensed under the Affero General Public License (AGPL) version 3.
#
# Copyright (C) 2023 New Vector, Ltd
#
# 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.
#
# See the GNU Affero General Public License for more details:
# <https://www.gnu.org/licenses/agpl-3.0.html>.
#
# Originally licensed under the Apache License, Version 2.0:
# <http://www.apache.org/licenses/LICENSE-2.0>.
#
# [This file includes modifications made by New Vector Limited]
#
#

import logging
import threading
from contextlib import nullcontext
from functools import wraps
from types import TracebackType
from typing import (
    TYPE_CHECKING,
    Any,
    Awaitable,
    Callable,
    Dict,
    Iterable,
    Optional,
    Set,
    Type,
    TypeVar,
    Union,
)

from prometheus_client import Metric
from prometheus_client.core import REGISTRY, Counter, Gauge
from typing_extensions import ParamSpec

from twisted.internet import defer

from synapse.logging.context import (
    ContextResourceUsage,
    LoggingContext,
    PreserveLoggingContext,
)
from synapse.logging.opentracing import SynapseTags, start_active_span
from synapse.metrics._types import Collector

if TYPE_CHECKING:
    import resource

    # Old versions don't have `LiteralString`
    from typing_extensions import LiteralString


logger = logging.getLogger(__name__)


_background_process_start_count = Counter(
    "synapse_background_process_start_count",
    "Number of background processes started",
    ["name"],
)

_background_process_in_flight_count = Gauge(
    "synapse_background_process_in_flight_count",
    "Number of background processes in flight",
    labelnames=["name"],
)

# we set registry=None in all of these to stop them getting registered with
# the default registry. Instead we collect them all via the CustomCollector,
# which ensures that we can update them before they are collected.
#
_background_process_ru_utime = Counter(
    "synapse_background_process_ru_utime_seconds",
    "User CPU time used by background processes, in seconds",
    ["name"],
    registry=None,
)

_background_process_ru_stime = Counter(
    "synapse_background_process_ru_stime_seconds",
    "System CPU time used by background processes, in seconds",
    ["name"],
    registry=None,
)

_background_process_db_txn_count = Counter(
    "synapse_background_process_db_txn_count",
    "Number of database transactions done by background processes",
    ["name"],
    registry=None,
)

_background_process_db_txn_duration = Counter(
    "synapse_background_process_db_txn_duration_seconds",
    (
        "Seconds spent by background processes waiting for database "
        "transactions, excluding scheduling time"
    ),
    ["name"],
    registry=None,
)

_background_process_db_sched_duration = Counter(
    "synapse_background_process_db_sched_duration_seconds",
    "Seconds spent by background processes waiting for database connections",
    ["name"],
    registry=None,
)

# map from description to a counter, so that we can name our logcontexts
# incrementally. (It actually duplicates _background_process_start_count, but
# it's much simpler to do so than to try to combine them.)
_background_process_counts: Dict[str, int] = {}

# Set of all running background processes that became active active since the
# last time metrics were scraped (i.e. background processes that performed some
# work since the last scrape.)
#
# We do it like this to handle the case where we have a large number of
# background processes stacking up behind a lock or linearizer, where we then
# only need to iterate over and update metrics for the process that have
# actually been active and can ignore the idle ones.
_background_processes_active_since_last_scrape: "Set[_BackgroundProcess]" = set()

# A lock that covers the above set and dict
_bg_metrics_lock = threading.Lock()


class _Collector(Collector):
    """A custom metrics collector for the background process metrics.

    Ensures that all of the metrics are up-to-date with any in-flight processes
    before they are returned.
    """

    def collect(self) -> Iterable[Metric]:
        global _background_processes_active_since_last_scrape

        # We swap out the _background_processes set with an empty one so that
        # we can safely iterate over the set without holding the lock.
        with _bg_metrics_lock:
            _background_processes_copy = _background_processes_active_since_last_scrape
            _background_processes_active_since_last_scrape = set()

        for process in _background_processes_copy:
            process.update_metrics()

        # now we need to run collect() over each of the static Counters, and
        # yield each metric they return.
        for m in (
            _background_process_ru_utime,
            _background_process_ru_stime,
            _background_process_db_txn_count,
            _background_process_db_txn_duration,
            _background_process_db_sched_duration,
        ):
            yield from m.collect()


REGISTRY.register(_Collector())


class _BackgroundProcess:
    def __init__(self, desc: str, ctx: LoggingContext):
        self.desc = desc
        self._context = ctx
        self._reported_stats: Optional[ContextResourceUsage] = None

    def update_metrics(self) -> None:
        """Updates the metrics with values from this process."""
        new_stats = self._context.get_resource_usage()
        if self._reported_stats is None:
            diff = new_stats
        else:
            diff = new_stats - self._reported_stats
        self._reported_stats = new_stats

        # For unknown reasons, the difference in times can be negative. See comment in
        # synapse.http.request_metrics.RequestMetrics.update_metrics.
        _background_process_ru_utime.labels(self.desc).inc(max(diff.ru_utime, 0))
        _background_process_ru_stime.labels(self.desc).inc(max(diff.ru_stime, 0))
        _background_process_db_txn_count.labels(self.desc).inc(diff.db_txn_count)
        _background_process_db_txn_duration.labels(self.desc).inc(
            diff.db_txn_duration_sec
        )
        _background_process_db_sched_duration.labels(self.desc).inc(
            diff.db_sched_duration_sec
        )


R = TypeVar("R")


def run_as_background_process(
    desc: "LiteralString",
    func: Callable[..., Awaitable[Optional[R]]],
    *args: Any,
    bg_start_span: bool = True,
    **kwargs: Any,
) -> "defer.Deferred[Optional[R]]":
    """Run the given function in its own logcontext, with resource metrics

    This should be used to wrap processes which are fired off to run in the
    background, instead of being associated with a particular request.

    It returns a Deferred which completes when the function completes, but it doesn't
    follow the synapse logcontext rules, which makes it appropriate for passing to
    clock.looping_call and friends (or for firing-and-forgetting in the middle of a
    normal synapse async function).

    Args:
        desc: a description for this background process type
        func: a function, which may return a Deferred or a coroutine
        bg_start_span: Whether to start an opentracing span. Defaults to True.
            Should only be disabled for processes that will not log to or tag
            a span.
        args: positional args for func
        kwargs: keyword args for func

    Returns:
        Deferred which returns the result of func, or `None` if func raises.
        Note that the returned Deferred does not follow the synapse logcontext
        rules.
    """

    async def run() -> Optional[R]:
        with _bg_metrics_lock:
            count = _background_process_counts.get(desc, 0)
            _background_process_counts[desc] = count + 1

        _background_process_start_count.labels(desc).inc()
        _background_process_in_flight_count.labels(desc).inc()

        with BackgroundProcessLoggingContext(desc, count) as context:
            try:
                if bg_start_span:
                    ctx = start_active_span(
                        f"bgproc.{desc}", tags={SynapseTags.REQUEST_ID: str(context)}
                    )
                else:
                    ctx = nullcontext()  # type: ignore[assignment]
                with ctx:
                    return await func(*args, **kwargs)
            except Exception:
                logger.exception(
                    "Background process '%s' threw an exception",
                    desc,
                )
                return None
            finally:
                _background_process_in_flight_count.labels(desc).dec()

    with PreserveLoggingContext():
        # Note that we return a Deferred here so that it can be used in a
        # looping_call and other places that expect a Deferred.
        return defer.ensureDeferred(run())


P = ParamSpec("P")


def wrap_as_background_process(
    desc: "LiteralString",
) -> Callable[
    [Callable[P, Awaitable[Optional[R]]]],
    Callable[P, "defer.Deferred[Optional[R]]"],
]:
    """Decorator that wraps an asynchronous function `func`, returning a synchronous
    decorated function. Calling the decorated version runs `func` as a background
    process, forwarding all arguments verbatim.

    That is,

        @wrap_as_background_process
        def func(*args): ...
        func(1, 2, third=3)

    is equivalent to:

        def func(*args): ...
        run_as_background_process(func, 1, 2, third=3)

    The former can be convenient if `func` needs to be run as a background process in
    multiple places.
    """

    def wrap_as_background_process_inner(
        func: Callable[P, Awaitable[Optional[R]]]
    ) -> Callable[P, "defer.Deferred[Optional[R]]"]:
        @wraps(func)
        def wrap_as_background_process_inner_2(
            *args: P.args, **kwargs: P.kwargs
        ) -> "defer.Deferred[Optional[R]]":
            # type-ignore: mypy is confusing kwargs with the bg_start_span kwarg.
            #     Argument 4 to "run_as_background_process" has incompatible type
            #     "**P.kwargs"; expected "bool"
            # See https://github.com/python/mypy/issues/8862
            return run_as_background_process(desc, func, *args, **kwargs)  # type: ignore[arg-type]

        return wrap_as_background_process_inner_2

    return wrap_as_background_process_inner


class BackgroundProcessLoggingContext(LoggingContext):
    """A logging context that tracks in flight metrics for background
    processes.
    """

    __slots__ = ["_proc"]

    def __init__(self, name: str, instance_id: Optional[Union[int, str]] = None):
        """

        Args:
            name: The name of the background process. Each distinct `name` gets a
                separate prometheus time series.

            instance_id: an identifer to add to `name` to distinguish this instance of
                the named background process in the logs. If this is `None`, one is
                made up based on id(self).
        """
        if instance_id is None:
            instance_id = id(self)
        super().__init__("%s-%s" % (name, instance_id))
        self._proc: Optional[_BackgroundProcess] = _BackgroundProcess(name, self)

    def start(self, rusage: "Optional[resource.struct_rusage]") -> None:
        """Log context has started running (again)."""

        super().start(rusage)

        if self._proc is None:
            logger.error(
                "Background process re-entered without a proc: %s",
                self.name,
                stack_info=True,
            )
            return

        # We've become active again so we make sure we're in the list of active
        # procs. (Note that "start" here means we've become active, as opposed
        # to starting for the first time.)
        with _bg_metrics_lock:
            _background_processes_active_since_last_scrape.add(self._proc)

    def __exit__(
        self,
        type: Optional[Type[BaseException]],
        value: Optional[BaseException],
        traceback: Optional[TracebackType],
    ) -> None:
        """Log context has finished."""

        super().__exit__(type, value, traceback)

        if self._proc is None:
            logger.error(
                "Background process exited without a proc: %s",
                self.name,
                stack_info=True,
            )
            return

        # The background process has finished. We explicitly remove and manually
        # update the metrics here so that if nothing is scraping metrics the set
        # doesn't infinitely grow.
        with _bg_metrics_lock:
            _background_processes_active_since_last_scrape.discard(self._proc)

        self._proc.update_metrics()

        # Set proc to None to break the reference cycle.
        self._proc = None