Allow wsrelay to fail without FATAL (#15191)

We have not identify the root cause of wsrelay failure but attempt to make wsrelay restart itself resulted in postgres and redis connection leak. We were not able to fully identify where the redis connection leak comes from so reverting back to failing and removing startsecs 30 will prevent wsrelay to FATAL
This commit is contained in:
Hao Liu 2024-05-20 19:34:12 -04:00 committed by GitHub
parent 7de350dc3e
commit fc9064e27f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 58 additions and 85 deletions

View File

@ -101,8 +101,9 @@ class Command(BaseCommand):
migrating = bool(executor.migration_plan(executor.loader.graph.leaf_nodes()))
connection.close() # Because of async nature, main loop will use new connection, so close this
except Exception as exc:
logger.warning(f'Error on startup of run_wsrelay (error: {exc}), retry in 10s...')
time.sleep(10)
time.sleep(10) # Prevent supervisor from restarting the service too quickly and the service to enter FATAL state
# sleeping before logging because logging rely on setting which require database connection...
logger.warning(f'Error on startup of run_wsrelay (error: {exc}), slept for 10s...')
return
# In containerized deployments, migrations happen in the task container,
@ -121,13 +122,14 @@ class Command(BaseCommand):
return
try:
my_hostname = Instance.objects.my_hostname()
my_hostname = Instance.objects.my_hostname() # This relies on settings.CLUSTER_HOST_ID which requires database connection
logger.info('Active instance with hostname {} is registered.'.format(my_hostname))
except RuntimeError as e:
# the CLUSTER_HOST_ID in the task, and web instance must match and
# ensure network connectivity between the task and web instance
logger.info('Unable to return currently active instance: {}, retry in 5s...'.format(e))
time.sleep(5)
time.sleep(10) # Prevent supervisor from restarting the service too quickly and the service to enter FATAL state
# sleeping before logging because logging rely on setting which require database connection...
logger.warning(f"Unable to return currently active instance: {e}, slept for 10s before return.")
return
if options.get('status'):
@ -166,12 +168,14 @@ class Command(BaseCommand):
WebsocketsMetricsServer().start()
while True:
try:
asyncio.run(WebSocketRelayManager().run())
except KeyboardInterrupt:
logger.info('Shutting down Websocket Relayer')
break
except Exception as e:
logger.exception('Error in Websocket Relayer, exception: {}. Restarting in 10 seconds'.format(e))
time.sleep(10)
try:
logger.info('Starting Websocket Relayer...')
websocket_relay_manager = WebSocketRelayManager()
asyncio.run(websocket_relay_manager.run())
except KeyboardInterrupt:
logger.info('Terminating Websocket Relayer')
except BaseException as e: # BaseException is used to catch all exceptions including asyncio.CancelledError
time.sleep(10) # Prevent supervisor from restarting the service too quickly and the service to enter FATAL state
# sleeping before logging because logging rely on setting which require database connection...
logger.warning(f"Encounter error while running Websocket Relayer {e}, slept for 10s...")
return

View File

@ -285,8 +285,6 @@ class WebSocketRelayManager(object):
except asyncio.CancelledError:
# Handle the case where the task was already cancelled by the time we got here.
pass
except Exception as e:
logger.warning(f"Failed to cancel relay connection for {hostname}: {e}")
del self.relay_connections[hostname]
@ -297,8 +295,6 @@ class WebSocketRelayManager(object):
self.stats_mgr.delete_remote_host_stats(hostname)
except KeyError:
pass
except Exception as e:
logger.warning(f"Failed to delete stats for {hostname}: {e}")
async def run(self):
event_loop = asyncio.get_running_loop()
@ -306,7 +302,6 @@ class WebSocketRelayManager(object):
self.stats_mgr = RelayWebsocketStatsManager(event_loop, self.local_hostname)
self.stats_mgr.start()
# Set up a pg_notify consumer for allowing web nodes to "provision" and "deprovision" themselves gracefully.
database_conf = deepcopy(settings.DATABASES['default'])
database_conf['OPTIONS'] = deepcopy(database_conf.get('OPTIONS', {}))
@ -318,79 +313,54 @@ class WebSocketRelayManager(object):
if 'PASSWORD' in database_conf:
database_conf['OPTIONS']['password'] = database_conf.pop('PASSWORD')
task = None
async_conn = await psycopg.AsyncConnection.connect(
dbname=database_conf['NAME'],
host=database_conf['HOST'],
user=database_conf['USER'],
port=database_conf['PORT'],
**database_conf.get("OPTIONS", {}),
)
# Managing the async_conn here so that we can close it if we need to restart the connection
async_conn = None
await async_conn.set_autocommit(True)
on_ws_heartbeat_task = event_loop.create_task(self.on_ws_heartbeat(async_conn))
# Establishes a websocket connection to /websocket/relay on all API servers
try:
while True:
if not task or task.done():
try:
# Try to close the connection if it's open
if async_conn:
try:
await async_conn.close()
except Exception as e:
logger.warning(f"Failed to close connection to database for pg_notify: {e}")
while True:
if on_ws_heartbeat_task.done():
raise Exception("on_ws_heartbeat_task has exited")
# and re-establish the connection
async_conn = await psycopg.AsyncConnection.connect(
dbname=database_conf['NAME'],
host=database_conf['HOST'],
user=database_conf['USER'],
port=database_conf['PORT'],
**database_conf.get("OPTIONS", {}),
)
await async_conn.set_autocommit(True)
future_remote_hosts = self.known_hosts.keys()
current_remote_hosts = self.relay_connections.keys()
deleted_remote_hosts = set(current_remote_hosts) - set(future_remote_hosts)
new_remote_hosts = set(future_remote_hosts) - set(current_remote_hosts)
# before creating the task that uses the connection
task = event_loop.create_task(self.on_ws_heartbeat(async_conn), name="on_ws_heartbeat")
logger.info("Creating `on_ws_heartbeat` task in event loop.")
# This loop handles if we get an advertisement from a host we already know about but
# the advertisement has a different IP than we are currently connected to.
for hostname, address in self.known_hosts.items():
if hostname not in self.relay_connections:
# We've picked up a new hostname that we don't know about yet.
continue
except Exception as e:
logger.warning(f"Failed to connect to database for pg_notify: {e}")
if address != self.relay_connections[hostname].remote_host:
deleted_remote_hosts.add(hostname)
new_remote_hosts.add(hostname)
future_remote_hosts = self.known_hosts.keys()
current_remote_hosts = self.relay_connections.keys()
deleted_remote_hosts = set(current_remote_hosts) - set(future_remote_hosts)
new_remote_hosts = set(future_remote_hosts) - set(current_remote_hosts)
# Delete any hosts with closed connections
for hostname, relay_conn in self.relay_connections.items():
if not relay_conn.connected:
deleted_remote_hosts.add(hostname)
# This loop handles if we get an advertisement from a host we already know about but
# the advertisement has a different IP than we are currently connected to.
for hostname, address in self.known_hosts.items():
if hostname not in self.relay_connections:
# We've picked up a new hostname that we don't know about yet.
continue
if deleted_remote_hosts:
logger.info(f"Removing {deleted_remote_hosts} from websocket broadcast list")
await asyncio.gather(*[self.cleanup_offline_host(h) for h in deleted_remote_hosts])
if address != self.relay_connections[hostname].remote_host:
deleted_remote_hosts.add(hostname)
new_remote_hosts.add(hostname)
if new_remote_hosts:
logger.info(f"Adding {new_remote_hosts} to websocket broadcast list")
# Delete any hosts with closed connections
for hostname, relay_conn in self.relay_connections.items():
if not relay_conn.connected:
deleted_remote_hosts.add(hostname)
for h in new_remote_hosts:
stats = self.stats_mgr.new_remote_host_stats(h)
relay_connection = WebsocketRelayConnection(name=self.local_hostname, stats=stats, remote_host=self.known_hosts[h])
relay_connection.start()
self.relay_connections[h] = relay_connection
if deleted_remote_hosts:
logger.info(f"Removing {deleted_remote_hosts} from websocket broadcast list")
await asyncio.gather(*[self.cleanup_offline_host(h) for h in deleted_remote_hosts])
if new_remote_hosts:
logger.info(f"Adding {new_remote_hosts} to websocket broadcast list")
for h in new_remote_hosts:
stats = self.stats_mgr.new_remote_host_stats(h)
relay_connection = WebsocketRelayConnection(name=self.local_hostname, stats=stats, remote_host=self.known_hosts[h])
relay_connection.start()
self.relay_connections[h] = relay_connection
await asyncio.sleep(settings.BROADCAST_WEBSOCKET_NEW_INSTANCE_POLL_RATE_SECONDS)
finally:
if async_conn:
logger.info("Shutting down db connection for wsrelay.")
try:
await async_conn.close()
except Exception as e:
logger.info(f"Failed to close connection to database for pg_notify: {e}")
await asyncio.sleep(settings.BROADCAST_WEBSOCKET_NEW_INSTANCE_POLL_RATE_SECONDS)

View File

@ -31,7 +31,6 @@ command = awx-manage run_wsrelay
directory = /var/lib/awx
{% endif %}
autorestart = true
startsecs = 30
stopasgroup=true
killasgroup=true
stdout_logfile=/dev/stdout