Don't reply when we known that client is gone

In case of a broker restart/failover a reply queue can be
unreachable for short period the IncomingMessage.send_reply
will block for 60 seconds in this case or until rabbit recovers.

But in case of the reply queue is unreachable because the
rpc client is really gone, we can have a ton of reply to send
waiting 60 seconds.
This leads to a starvation of connection of the pool
The rpc server take to much time to send reply, other rpc client will
raise TimeoutError because their don't receive their replies in time.

This changes introduces an object cache that stores already known gone
client to not wait 60 seconds and hold a connection of the pool
Keeping 200 last gone rpc client for 1 minute is enough
and doesn't hold to much memory.

This also don't raise anymore a frightening exception when we can't send reply
to the rpc client. But just logging a info about missing exchange and
a warning about unsend reply.

Closes-bug: #1460652

Change-Id: I928b30c9b5f9ee007532ff703e136640b0e8aaf4

oslo_messaging/_drivers/amqp.py

@@ -263,3 +263,7 @@ def _add_unique_id(msg):
     """Add unique_id for checking duplicate messages."""
     unique_id = uuid.uuid4().hex
     msg.update({UNIQUE_ID: unique_id})
+
+
+class AMQPDestinationNotFound(Exception):
+    pass

oslo_messaging/_drivers/amqpdriver.py

@@ -19,6 +19,7 @@
 import threading
 import uuid
 
+import cachetools
 from six import moves
 
 import oslo_messaging
@@ -27,13 +28,15 @@
 from oslo_messaging._drivers import common as rpc_common
 from oslo_messaging._i18n import _
 from oslo_messaging._i18n import _LI
+from oslo_messaging._i18n import _LW
 
 LOG = logging.getLogger(__name__)
 
 
 class AMQPIncomingMessage(base.IncomingMessage):
 
-    def __init__(self, listener, ctxt, message, unique_id, msg_id, reply_q):
+    def __init__(self, listener, ctxt, message, unique_id, msg_id, reply_q,
+                 obsolete_reply_queues):
         super(AMQPIncomingMessage, self).__init__(listener, ctxt,
                                                   dict(message))
 
@@ -42,9 +45,15 @@ def __init__(self, listener, ctxt, message, unique_id, msg_id, reply_q):
         self.reply_q = reply_q
         self.acknowledge_callback = message.acknowledge
         self.requeue_callback = message.requeue
+        self._obsolete_reply_queues = obsolete_reply_queues
 
     def _send_reply(self, conn, reply=None, failure=None,
                     ending=False, log_failure=True):
+        if (self.reply_q and
+            not self._obsolete_reply_queues.reply_q_valid(self.reply_q,
+                                                          self.msg_id)):
+            return
+
         if failure:
             failure = rpc_common.serialize_remote_exception(failure,
                                                             log_failure)
@@ -60,15 +69,29 @@ def _send_reply(self, conn, reply=None, failure=None,
         # Otherwise use the msg_id for backward compatibility.
         if self.reply_q:
             msg['_msg_id'] = self.msg_id
-            conn.direct_send(self.reply_q, rpc_common.serialize_msg(msg))
+            try:
+                conn.direct_send(self.reply_q, rpc_common.serialize_msg(msg))
+            except rpc_amqp.AMQPDestinationNotFound:
+                self._obsolete_reply_queues.add(self.reply_q, self.msg_id)
         else:
+            # TODO(sileht): look at which version of oslo-incubator rpc
+            # send need this, but I guess this is older than icehouse
+            # if this is icehouse, we can drop this at M
+            # if this is havana, we can drop this now.
             conn.direct_send(self.msg_id, rpc_common.serialize_msg(msg))
 
     def reply(self, reply=None, failure=None, log_failure=True):
         if not self.msg_id:
             # NOTE(Alexei_987) not sending reply, if msg_id is empty
             #    because reply should not be expected by caller side
             return
+
+        # NOTE(sileht): return without hold the a connection if possible
+        if (self.reply_q and
+            not self._obsolete_reply_queues.reply_q_valid(self.reply_q,
+                                                          self.msg_id)):
+            return
+
         with self.listener.driver._get_connection(
                 rpc_amqp.PURPOSE_SEND) as conn:
             if self.listener.driver.send_single_reply:
@@ -92,6 +115,51 @@ def requeue(self):
         self.requeue_callback()
 
 
+class ObsoleteReplyQueuesCache(object):
+    """Cache of reply queue id that doesn't exists anymore.
+
+    NOTE(sileht): In case of a broker restart/failover
+    a reply queue can be unreachable for short period
+    the IncomingMessage.send_reply will block for 60 seconds
+    in this case or until rabbit recovers.
+
+    But in case of the reply queue is unreachable because the
+    rpc client is really gone, we can have a ton of reply to send
+    waiting 60 seconds.
+    This leads to a starvation of connection of the pool
+    The rpc server take to much time to send reply, other rpc client will
+    raise TimeoutError because their don't receive their replies in time.
+
+    This object cache stores already known gone client to not wait 60 seconds
+    and hold a connection of the pool.
+    Keeping 200 last gone rpc client for 1 minute is enough
+    and doesn't hold to much memory.
+    """
+
+    SIZE = 200
+    TTL = 60
+
+    def __init__(self):
+        self._lock = threading.RLock()
+        self._cache = cachetools.TTLCache(self.SIZE, self.TTL)
+
+    def reply_q_valid(self, reply_q, msg_id):
+        if reply_q in self._cache:
+            self._no_reply_log(reply_q, msg_id)
+            return False
+        return True
+
+    def add(self, reply_q, msg_id):
+        with self._lock:
+            self._cache.update({reply_q: msg_id})
+        self._no_reply_log(reply_q, msg_id)
+
+    def _no_reply_log(self, reply_q, msg_id):
+        LOG.warn(_LW("%(reply_queue)s doesn't exists, drop reply to "
+                     "%(msg_id)s"), {'reply_queue': reply_q,
+                                     'msg_id': msg_id})
+
+
 class AMQPListener(base.Listener):
 
     def __init__(self, driver, conn):
@@ -100,6 +168,7 @@ def __init__(self, driver, conn):
         self.msg_id_cache = rpc_amqp._MsgIdCache()
         self.incoming = []
         self._stopped = threading.Event()
+        self._obsolete_reply_queues = ObsoleteReplyQueuesCache()
 
     def __call__(self, message):
         ctxt = rpc_amqp.unpack_context(self.conf, message)
@@ -116,7 +185,8 @@ def __call__(self, message):
                                                  message,
                                                  unique_id,
                                                  ctxt.msg_id,
-                                                 ctxt.reply_q))
+                                                 ctxt.reply_q,
+                                                 self._obsolete_reply_queues))
 
     def poll(self, timeout=None):
         while not self._stopped.is_set():

oslo_messaging/_drivers/impl_rabbit.py

@@ -1081,8 +1081,17 @@ def _publish_and_retry_on_missing_exchange(self, exchange, msg,
                                  "retrying...") % {
                                      'exchange': exchange.name,
                                      'routing_key': routing_key})
-                    time.sleep(1)
+                    time.sleep(0.25)
                     continue
+                elif exc.code == 404:
+                    msg = _("The exchange %(exchange)s to send to "
+                            "%(routing_key)s still doesn't exist after "
+                            "%(duration)s sec abandonning...") % {
+                                'duration': duration,
+                                'exchange': exchange.name,
+                                'routing_key': routing_key}
+                    LOG.info(msg)
+                    raise rpc_amqp.AMQPDestinationNotFound(msg)
                 raise
 
     def direct_send(self, msg_id, msg):

oslo_messaging/tests/drivers/test_impl_rabbit.py

@@ -355,6 +355,11 @@ class TestSendReceive(test_utils.BaseTestCase):
         ('zero', dict(rx_id=False, reply=0)),
     ]
 
+    _reply_fail = [
+        ('reply_success', dict(reply_failure_404=False)),
+        ('reply_failure', dict(reply_failure_404=True)),
+    ]
+
     _failure = [
         ('success', dict(failure=False)),
         ('failure', dict(failure=True, expected=False)),
@@ -376,11 +381,14 @@ def generate_scenarios(cls):
         cls.scenarios = testscenarios.multiply_scenarios(cls._n_senders,
                                                          cls._context,
                                                          cls._reply,
+                                                         cls._reply_fail,
                                                          cls._failure,
                                                          cls._timeout,
                                                          cls._reply_ending)
 
     def test_send_receive(self):
+        self.config(kombu_reconnect_timeout=0.5,
+                    group="oslo_messaging_rabbit")
         self.config(heartbeat_timeout_threshold=0,
                     group="oslo_messaging_rabbit")
         self.config(send_single_reply=self.send_single_reply,
@@ -409,16 +417,21 @@ def stub_error(msg, *a, **kw):
 
         def send_and_wait_for_reply(i):
             try:
+                if self.reply_failure_404:
+                    timeout = 0.01
+                else:
+                    timeout = self.timeout
                 replies.append(driver.send(target,
                                            self.ctxt,
                                            {'tx_id': i},
                                            wait_for_reply=True,
-                                           timeout=self.timeout))
+                                           timeout=timeout))
                 self.assertFalse(self.failure)
                 self.assertIsNone(self.timeout)
             except (ZeroDivisionError, oslo_messaging.MessagingTimeout) as e:
                 replies.append(e)
-                self.assertTrue(self.failure or self.timeout is not None)
+                self.assertTrue(self.failure or self.timeout is not None
+                                or self.reply_failure_404)
 
         while len(senders) < self.n_senders:
             senders.append(threading.Thread(target=send_and_wait_for_reply,
@@ -438,6 +451,18 @@ def send_and_wait_for_reply(i):
         if len(order) > 1:
             order[-1], order[-2] = order[-2], order[-1]
 
+        if self.reply_failure_404:
+            start = time.time()
+            # NOTE(sileht): Simulate a rpc client restart
+            # By returning a ExchangeNotFound when we try to
+            # send reply
+            exc = (driver._reply_q_conn.connection.
+                   connection.channel_errors[0]())
+            exc.code = 404
+            self.useFixture(mockpatch.Patch(
+                'kombu.messaging.Producer.publish',
+                side_effect=exc))
+
         for i in order:
             if self.timeout is None:
                 if self.failure:
@@ -451,11 +476,19 @@ def send_and_wait_for_reply(i):
                     msgs[i].reply({'rx_id': i})
                 else:
                     msgs[i].reply(self.reply)
+            elif self.reply_failure_404:
+                msgs[i].reply({})
             senders[i].join()
 
+        if self.reply_failure_404:
+            # NOTE(sileht) all reply fail, first take
+            # kombu_reconnect_timeout seconds to fail
+            # next immediatly fail
+            self.assertAlmostEqual(0.5, time.time() - start, 1)
+
         self.assertEqual(len(senders), len(replies))
         for i, reply in enumerate(replies):
-            if self.timeout is not None:
+            if self.timeout is not None or self.reply_failure_404:
                 self.assertIsInstance(reply, oslo_messaging.MessagingTimeout)
             elif self.failure:
                 self.assertIsInstance(reply, ZeroDivisionError)

requirements.txt

@@ -13,6 +13,7 @@ stevedore>=1.5.0 # Apache-2.0
 
 # for jsonutils
 six>=1.9.0
+cachetools>=1.0.0 # MIT License
 
 # FIXME(markmc): remove this when the drivers no longer
 # import eventlet