amqp backend: polling now uses message.requeue instead of republishing

celery/backends/amqp.py

@@ -157,19 +157,21 @@ class AMQPBackend(BaseDictBackend):
         with self.app.pool.acquire_channel(block=True) as (_, channel):
             binding = self._create_binding(task_id)(channel)
             binding.declare()
-            latest, acc = None, None
-            for i in xrange(backlog_limit):
-                latest, acc = acc, binding.get(no_ack=True)
+            prev = latest = acc = None
+            for i in xrange(backlog_limit):  ## spool ffwd
+                prev, latest, acc = latest, acc, binding.get(no_ack=False)
                 if not acc:  # no more messages
                     break
+                if prev:
+                    # backends are not expected to keep history,
+                    # so we delete everything except the most recent state.
+                    prev.ack()
             else:
                 raise self.BacklogLimitExceeded(task_id)
 
             if latest:
-                # new state to report
-                self._republish(channel, task_id, latest.body,
-                                latest.content_type, latest.content_encoding)
                 payload = self._cache[task_id] = latest.payload
+                latest.requeue()
                 return payload
             else:
                 # no new state, use previous