Regularly try to wake up dests instead of waiting for next PDU/EDU

changelog.d/15743.misc

@@ -0,0 +1 @@
+Regularly try to send transactions to other servers after they failed instead of waiting for a new event to be available before trying.

synapse/federation/sender/__init__.py

@@ -145,7 +145,6 @@
 from typing_extensions import Literal
 
 from twisted.internet import defer
-from twisted.internet.interfaces import IDelayedCall
 
 import synapse.metrics
 from synapse.api.presence import UserPresenceState
@@ -184,14 +183,18 @@
     "Total number of PDUs queued for sending across all destinations",
 )
 
-# Time (in s) after Synapse's startup that we will begin to wake up destinations
-# that have catch-up outstanding.
-CATCH_UP_STARTUP_DELAY_SEC = 15
+# Time (in s) to wait before trying to wake up destinations that have
+# catch-up outstanding. This will also be the delay applied at startup
+# before trying the same.
+# Please note that rate limiting still applies, so while the loop is
+# executed every X seconds the destinations may not be wake up because
+# they are being rate limited following previous attempt failures.
+WAKEUP_RETRY_PERIOD_SEC = 60
 
 # Time (in s) to wait in between waking up each destination, i.e. one destination
-# will be woken up every <x> seconds after Synapse's startup until we have woken
-# every destination has outstanding catch-up.
-CATCH_UP_STARTUP_INTERVAL_SEC = 5
+# will be woken up every <x> seconds until we have woken every destination
+# has outstanding catch-up.
+WAKEUP_INTERVAL_BETWEEN_DESTINATIONS_SEC = 5
 
 
 class AbstractFederationSender(metaclass=abc.ABCMeta):
@@ -415,12 +418,10 @@ def __init__(self, hs: "HomeServer"):
             / hs.config.ratelimiting.federation_rr_transactions_per_room_per_second
         )
 
-        # wake up destinations that have outstanding PDUs to be caught up
-        self._catchup_after_startup_timer: Optional[
-            IDelayedCall
-        ] = self.clock.call_later(
-            CATCH_UP_STARTUP_DELAY_SEC,
+        # Regularly wake up destinations that have outstanding PDUs to be caught up
+        self.clock.looping_call(
             run_as_background_process,
+            WAKEUP_RETRY_PERIOD_SEC * 1000.0,
             "wake_destinations_needing_catchup",
             self._wake_destinations_needing_catchup,
         )
@@ -966,7 +967,6 @@ async def _wake_destinations_needing_catchup(self) -> None:
 
             if not destinations_to_wake:
                 # finished waking all destinations!
-                self._catchup_after_startup_timer = None
                 break
 
             last_processed = destinations_to_wake[-1]
@@ -983,4 +983,4 @@ async def _wake_destinations_needing_catchup(self) -> None:
                     last_processed,
                 )
                 self.wake_destination(destination)
-                await self.clock.sleep(CATCH_UP_STARTUP_INTERVAL_SEC)
+                await self.clock.sleep(WAKEUP_INTERVAL_BETWEEN_DESTINATIONS_SEC)

tests/federation/test_federation_catch_up.py

@@ -431,28 +431,22 @@ def test_catch_up_on_synapse_startup(self) -> None:
         # ACT: call _wake_destinations_needing_catchup
 
         # patch wake_destination to just count the destinations instead
-        woken = []
+        woken = set()
 
         def wake_destination_track(destination: str) -> None:
-            woken.append(destination)
+            woken.add(destination)
 
         self.federation_sender.wake_destination = wake_destination_track  # type: ignore[assignment]
 
-        # cancel the pre-existing timer for _wake_destinations_needing_catchup
-        # this is because we are calling it manually rather than waiting for it
-        # to be called automatically
-        assert self.federation_sender._catchup_after_startup_timer is not None
-        self.federation_sender._catchup_after_startup_timer.cancel()
-
-        self.get_success(
-            self.federation_sender._wake_destinations_needing_catchup(), by=5.0
-        )
+        self.pump(by=5.0)
 
         # ASSERT (_wake_destinations_needing_catchup):
         # - all remotes are woken up, save for zzzerver
         self.assertNotIn("zzzerver", woken)
-        # - all destinations are woken exactly once; they appear once in woken.
-        self.assertCountEqual(woken, server_names[:-1])
+        # - all destinations are woken, potentially more than once, since the
+        # wake up is called regularly and we don't ack in this test that a transaction
+        # has been successfully sent.
+        self.assertCountEqual(woken, set(server_names[:-1]))
 
     def test_not_latest_event(self) -> None:
         """Test that we send the latest event in the room even if its not ours."""