Immediately retry any requests that have backed off when a server comes back online. (#12500)

Otherwise it can take up to a minute for any in-flight `/send` requests to be retried.
This commit is contained in:
Erik Johnston 2022-05-10 10:39:54 +01:00 committed by GitHub
parent ade3008821
commit 8dd3e0e084
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
6 changed files with 179 additions and 6 deletions

1
changelog.d/12500.misc Normal file
View file

@ -0,0 +1 @@
Immediately retry any requests that have backed off when a server comes back online.

View file

@ -73,7 +73,7 @@ from synapse.logging.context import make_deferred_yieldable, run_in_background
from synapse.logging.opentracing import set_tag, start_active_span, tags
from synapse.types import JsonDict
from synapse.util import json_decoder
from synapse.util.async_helpers import timeout_deferred
from synapse.util.async_helpers import AwakenableSleeper, timeout_deferred
from synapse.util.metrics import Measure
if TYPE_CHECKING:
@ -353,6 +353,13 @@ class MatrixFederationHttpClient:
self._cooperator = Cooperator(scheduler=schedule)
self._sleeper = AwakenableSleeper(self.reactor)
def wake_destination(self, destination: str) -> None:
"""Called when the remote server may have come back online."""
self._sleeper.wake(destination)
async def _send_request_with_optional_trailing_slash(
self,
request: MatrixFederationRequest,
@ -474,6 +481,8 @@ class MatrixFederationHttpClient:
self._store,
backoff_on_404=backoff_on_404,
ignore_backoff=ignore_backoff,
notifier=self.hs.get_notifier(),
replication_client=self.hs.get_replication_command_handler(),
)
method_bytes = request.method.encode("ascii")
@ -664,7 +673,9 @@ class MatrixFederationHttpClient:
delay,
)
await self.clock.sleep(delay)
# Sleep for the calculated delay, or wake up immediately
# if we get notified that the server is back up.
await self._sleeper.sleep(request.destination, delay * 1000)
retries_left -= 1
else:
raise

View file

@ -228,9 +228,7 @@ class Notifier:
# Called when there are new things to stream over replication
self.replication_callbacks: List[Callable[[], None]] = []
# Called when remote servers have come back online after having been
# down.
self.remote_server_up_callbacks: List[Callable[[str], None]] = []
self._federation_client = hs.get_federation_http_client()
self._third_party_rules = hs.get_third_party_event_rules()
@ -731,3 +729,7 @@ class Notifier:
# circular dependencies.
if self.federation_sender:
self.federation_sender.wake_destination(server)
# Tell the federation client about the fact the server is back up, so
# that any in flight requests can be immediately retried.
self._federation_client.wake_destination(server)

View file

@ -778,3 +778,60 @@ def delay_cancellation(awaitable: Awaitable[T]) -> Awaitable[T]:
new_deferred: "defer.Deferred[T]" = defer.Deferred(handle_cancel)
deferred.chainDeferred(new_deferred)
return new_deferred
class AwakenableSleeper:
"""Allows explicitly waking up deferreds related to an entity that are
currently sleeping.
"""
def __init__(self, reactor: IReactorTime) -> None:
self._streams: Dict[str, Set[defer.Deferred[None]]] = {}
self._reactor = reactor
def wake(self, name: str) -> None:
"""Wake everything related to `name` that is currently sleeping."""
stream_set = self._streams.pop(name, set())
for deferred in stream_set:
try:
with PreserveLoggingContext():
deferred.callback(None)
except Exception:
pass
async def sleep(self, name: str, delay_ms: int) -> None:
"""Sleep for the given number of milliseconds, or return if the given
`name` is explicitly woken up.
"""
# Create a deferred that gets called in N seconds
sleep_deferred: "defer.Deferred[None]" = defer.Deferred()
call = self._reactor.callLater(delay_ms / 1000, sleep_deferred.callback, None)
# Create a deferred that will get called if `wake` is called with
# the same `name`.
stream_set = self._streams.setdefault(name, set())
notify_deferred: "defer.Deferred[None]" = defer.Deferred()
stream_set.add(notify_deferred)
try:
# Wait for either the delay or for `wake` to be called.
await make_deferred_yieldable(
defer.DeferredList(
[sleep_deferred, notify_deferred],
fireOnOneCallback=True,
fireOnOneErrback=True,
consumeErrors=True,
)
)
finally:
# Clean up the state
curr_stream_set = self._streams.get(name)
if curr_stream_set is not None:
curr_stream_set.discard(notify_deferred)
if len(curr_stream_set) == 0:
self._streams.pop(name)
# Cancel the sleep if we were woken up
if call.active():
call.cancel()

View file

@ -14,13 +14,17 @@
import logging
import random
from types import TracebackType
from typing import Any, Optional, Type
from typing import TYPE_CHECKING, Any, Optional, Type
import synapse.logging.context
from synapse.api.errors import CodeMessageException
from synapse.storage import DataStore
from synapse.util import Clock
if TYPE_CHECKING:
from synapse.notifier import Notifier
from synapse.replication.tcp.handler import ReplicationCommandHandler
logger = logging.getLogger(__name__)
# the initial backoff, after the first transaction fails
@ -131,6 +135,8 @@ class RetryDestinationLimiter:
retry_interval: int,
backoff_on_404: bool = False,
backoff_on_failure: bool = True,
notifier: Optional["Notifier"] = None,
replication_client: Optional["ReplicationCommandHandler"] = None,
):
"""Marks the destination as "down" if an exception is thrown in the
context, except for CodeMessageException with code < 500.
@ -160,6 +166,9 @@ class RetryDestinationLimiter:
self.backoff_on_404 = backoff_on_404
self.backoff_on_failure = backoff_on_failure
self.notifier = notifier
self.replication_client = replication_client
def __enter__(self) -> None:
pass
@ -239,6 +248,19 @@ class RetryDestinationLimiter:
retry_last_ts,
self.retry_interval,
)
if self.notifier:
# Inform the relevant places that the remote server is back up.
self.notifier.notify_remote_server_up(self.destination)
if self.replication_client:
# If we're on a worker we try and inform master about this. The
# replication client doesn't hook into the notifier to avoid
# infinite loops where we send a `REMOTE_SERVER_UP` command to
# master, which then echoes it back to us which in turn pokes
# the notifier.
self.replication_client.send_remote_server_up(self.destination)
except Exception:
logger.exception("Failed to store destination_retry_timings")

View file

@ -28,6 +28,7 @@ from synapse.logging.context import (
make_deferred_yieldable,
)
from synapse.util.async_helpers import (
AwakenableSleeper,
ObservableDeferred,
concurrently_execute,
delay_cancellation,
@ -35,6 +36,7 @@ from synapse.util.async_helpers import (
timeout_deferred,
)
from tests.server import get_clock
from tests.unittest import TestCase
@ -496,3 +498,81 @@ class DelayCancellationTests(TestCase):
# logging context.
blocking_d.callback(None)
self.successResultOf(d)
class AwakenableSleeperTests(TestCase):
"Tests AwakenableSleeper"
def test_sleep(self):
reactor, _ = get_clock()
sleeper = AwakenableSleeper(reactor)
d = defer.ensureDeferred(sleeper.sleep("name", 1000))
reactor.pump([0.0])
self.assertFalse(d.called)
reactor.advance(0.5)
self.assertFalse(d.called)
reactor.advance(0.6)
self.assertTrue(d.called)
def test_explicit_wake(self):
reactor, _ = get_clock()
sleeper = AwakenableSleeper(reactor)
d = defer.ensureDeferred(sleeper.sleep("name", 1000))
reactor.pump([0.0])
self.assertFalse(d.called)
reactor.advance(0.5)
self.assertFalse(d.called)
sleeper.wake("name")
self.assertTrue(d.called)
reactor.advance(0.6)
def test_multiple_sleepers_timeout(self):
reactor, _ = get_clock()
sleeper = AwakenableSleeper(reactor)
d1 = defer.ensureDeferred(sleeper.sleep("name", 1000))
reactor.advance(0.6)
self.assertFalse(d1.called)
# Add another sleeper
d2 = defer.ensureDeferred(sleeper.sleep("name", 1000))
# Only the first sleep should time out now.
reactor.advance(0.6)
self.assertTrue(d1.called)
self.assertFalse(d2.called)
reactor.advance(0.6)
self.assertTrue(d2.called)
def test_multiple_sleepers_wake(self):
reactor, _ = get_clock()
sleeper = AwakenableSleeper(reactor)
d1 = defer.ensureDeferred(sleeper.sleep("name", 1000))
reactor.advance(0.5)
self.assertFalse(d1.called)
# Add another sleeper
d2 = defer.ensureDeferred(sleeper.sleep("name", 1000))
# Neither should fire yet
reactor.advance(0.3)
self.assertFalse(d1.called)
self.assertFalse(d2.called)
# Explicitly waking both up works
sleeper.wake("name")
self.assertTrue(d1.called)
self.assertTrue(d2.called)