Fix bug where 5s delays would occasionally happen. ()

This only affects deployments using workers.
This commit is contained in:
Erik Johnston 2023-02-24 14:39:50 +00:00 committed by GitHub
parent 8219525b66
commit b2357a898c
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 80 additions and 0 deletions
changelog.d
synapse/replication/tcp
tests/replication/tcp

1
changelog.d/15150.bugfix Normal file
View File

@ -0,0 +1 @@
Fix a bug introduced in Synapse 1.76 where 5s delays would occasionally occur in deployments using workers.

View File

@ -238,6 +238,24 @@ class ReplicationStreamer:
except Exception:
logger.exception("Failed to replicate")
# The last token we send may not match the current
# token, in which case we want to send out a `POSITION`
# to tell other workers the actual current position.
if updates[-1][0] < current_token:
logger.info(
"Sending position: %s -> %s",
stream.NAME,
current_token,
)
self.command_handler.send_command(
PositionCommand(
stream.NAME,
self._instance_name,
updates[-1][0],
current_token,
)
)
logger.debug("No more pending updates, breaking poke loop")
finally:
self.pending_updates = False

View File

@ -141,3 +141,64 @@ class ChannelsTestCase(BaseMultiWorkerStreamTestCase):
self.get_success(ctx_worker1.__aexit__(None, None, None))
self.assertTrue(d.called)
def test_wait_for_stream_position_rdata(self) -> None:
"""Check that wait for stream position correctly waits for an update
from the correct instance, when RDATA is sent.
"""
store = self.hs.get_datastores().main
cmd_handler = self.hs.get_replication_command_handler()
data_handler = self.hs.get_replication_data_handler()
worker1 = self.make_worker_hs(
"synapse.app.generic_worker",
extra_config={
"worker_name": "worker1",
"run_background_tasks_on": "worker1",
"redis": {"enabled": True},
},
)
cache_id_gen = worker1.get_datastores().main._cache_id_gen
assert cache_id_gen is not None
self.replicate()
# First, make sure the master knows that `worker1` exists.
initial_token = cache_id_gen.get_current_token()
cmd_handler.send_command(
PositionCommand("caches", "worker1", initial_token, initial_token)
)
self.replicate()
# `wait_for_stream_position` should only return once master receives a
# notification that `next_token2` has persisted.
ctx_worker1 = cache_id_gen.get_next_mult(2)
next_token1, next_token2 = self.get_success(ctx_worker1.__aenter__())
d = defer.ensureDeferred(
data_handler.wait_for_stream_position("worker1", "caches", next_token2)
)
self.assertFalse(d.called)
# Insert an entry into the cache stream with token `next_token1`, but
# not `next_token2`.
self.get_success(
store.db_pool.simple_insert(
table="cache_invalidation_stream_by_instance",
values={
"stream_id": next_token1,
"instance_name": "worker1",
"cache_func": "foo",
"keys": [],
"invalidation_ts": 0,
},
)
)
# Finish the context manager, triggering the data to be sent to master.
self.get_success(ctx_worker1.__aexit__(None, None, None))
# Master should get told about `next_token2`, so the deferred should
# resolve.
self.assertTrue(d.called)