KAFKA-3949: Avoid race condition when subscription changes during rebalance

Author: dpkp

kafka/cluster.py

@@ -291,6 +291,13 @@ def update_metadata(self, metadata):
         for listener in self._listeners:
             listener(self)
 
+        if self.need_all_topic_metadata:
+            # the listener may change the interested topics,
+            # which could cause another metadata refresh.
+            # If we have already fetched all topics, however,
+            # another fetch should be unnecessary.
+            self._need_update = False
+
     def add_listener(self, listener):
         """Add a callback function to be called on each metadata update"""
         self._listeners.add(listener)

kafka/consumer/fetcher.py

@@ -326,9 +326,6 @@ def fetched_records(self, max_records=None):
             max_records = self.config['max_poll_records']
         assert max_records > 0
 
-        if self._subscriptions.needs_partition_assignment:
-            return {}, False
-
         drained = collections.defaultdict(list)
         records_remaining = max_records
 
@@ -397,9 +394,6 @@ def _append(self, drained, part, max_records):
 
     def _message_generator(self):
         """Iterate over fetched_records"""
-        if self._subscriptions.needs_partition_assignment:
-            raise StopIteration('Subscription needs partition assignment')
-
         while self._next_partition_records or self._completed_fetches:
 
             if not self._next_partition_records:

kafka/consumer/group.py

@@ -644,6 +644,11 @@ def _poll_once(self, timeout_ms, max_records):
 
         timeout_ms = min(timeout_ms, self._coordinator.time_to_next_poll())
         self._client.poll(timeout_ms=timeout_ms)
+        # after the long poll, we should check whether the group needs to rebalance
+        # prior to returning data so that the group can stabilize faster
+        if self._coordinator.need_rejoin():
+            return {}
+
         records, _ = self._fetcher.fetched_records(max_records)
         return records
 
@@ -1055,6 +1060,11 @@ def _message_generator(self):
                 poll_ms = 0
             self._client.poll(timeout_ms=poll_ms)
 
+            # after the long poll, we should check whether the group needs to rebalance
+            # prior to returning data so that the group can stabilize faster
+            if self._coordinator.need_rejoin():
+                continue
+
             # We need to make sure we at least keep up with scheduled tasks,
             # like heartbeats, auto-commits, and metadata refreshes
             timeout_at = self._next_timeout()

kafka/consumer/subscription_state.py

@@ -68,7 +68,6 @@ def __init__(self, offset_reset_strategy='earliest'):
         self._group_subscription = set()
         self._user_assignment = set()
         self.assignment = dict()
-        self.needs_partition_assignment = False
         self.listener = None
 
         # initialize to true for the consumers to fetch offset upon starting up
@@ -172,7 +171,6 @@ def change_subscription(self, topics):
         log.info('Updating subscribed topics to: %s', topics)
         self.subscription = set(topics)
         self._group_subscription.update(topics)
-        self.needs_partition_assignment = True
 
         # Remove any assigned partitions which are no longer subscribed to
         for tp in set(self.assignment.keys()):
@@ -192,12 +190,12 @@ def group_subscribe(self, topics):
             raise IllegalStateError(self._SUBSCRIPTION_EXCEPTION_MESSAGE)
         self._group_subscription.update(topics)
 
-    def mark_for_reassignment(self):
+    def reset_group_subscription(self):
+        """Reset the group's subscription to only contain topics subscribed by this consumer."""
         if self._user_assignment:
             raise IllegalStateError(self._SUBSCRIPTION_EXCEPTION_MESSAGE)
         assert self.subscription is not None, 'Subscription required'
         self._group_subscription.intersection_update(self.subscription)
-        self.needs_partition_assignment = True
 
     def assign_from_user(self, partitions):
         """Manually assign a list of TopicPartitions to this consumer.
@@ -220,18 +218,17 @@ def assign_from_user(self, partitions):
         if self.subscription is not None:
             raise IllegalStateError(self._SUBSCRIPTION_EXCEPTION_MESSAGE)
 
-        self._user_assignment.clear()
-        self._user_assignment.update(partitions)
+        if self._user_assignment != set(partitions):
+            self._user_assignment = set(partitions)
 
-        for partition in partitions:
-            if partition not in self.assignment:
-                self._add_assigned_partition(partition)
+            for partition in partitions:
+                if partition not in self.assignment:
+                    self._add_assigned_partition(partition)
 
-        for tp in set(self.assignment.keys()) - self._user_assignment:
-            del self.assignment[tp]
+            for tp in set(self.assignment.keys()) - self._user_assignment:
+                del self.assignment[tp]
 
-        self.needs_partition_assignment = False
-        self.needs_fetch_committed_offsets = True
+            self.needs_fetch_committed_offsets = True
 
     def assign_from_subscribed(self, assignments):
         """Update the assignment to the specified partitions
@@ -245,24 +242,25 @@ def assign_from_subscribed(self, assignments):
             assignments (list of TopicPartition): partitions to assign to this
                 consumer instance.
         """
-        if self.subscription is None:
+        if not self.partitions_auto_assigned():
             raise IllegalStateError(self._SUBSCRIPTION_EXCEPTION_MESSAGE)
 
         for tp in assignments:
             if tp.topic not in self.subscription:
                 raise ValueError("Assigned partition %s for non-subscribed topic." % str(tp))
+
+        # after rebalancing, we always reinitialize the assignment state
         self.assignment.clear()
         for tp in assignments:
             self._add_assigned_partition(tp)
-        self.needs_partition_assignment = False
+        self.needs_fetch_committed_offsets = True
         log.info("Updated partition assignment: %s", assignments)
 
     def unsubscribe(self):
         """Clear all topic subscriptions and partition assignments"""
         self.subscription = None
         self._user_assignment.clear()
         self.assignment.clear()
-        self.needs_partition_assignment = True
         self.subscribed_pattern = None
 
     def group_subscription(self):

kafka/coordinator/base.py

@@ -344,23 +344,25 @@ def _handle_join_failure(self, _):
     def ensure_active_group(self):
         """Ensure that the group is active (i.e. joined and synced)"""
         with self._lock:
-            if not self.need_rejoin():
-                return
-
-            # call on_join_prepare if needed. We set a flag to make sure that
-            # we do not call it a second time if the client is woken up before
-            # a pending rebalance completes.
-            if not self.rejoining:
-                self._on_join_prepare(self._generation.generation_id,
-                                      self._generation.member_id)
-                self.rejoining = True
-
             if self._heartbeat_thread is None:
                 self._start_heartbeat_thread()
 
             while self.need_rejoin():
                 self.ensure_coordinator_ready()
 
+                # call on_join_prepare if needed. We set a flag
+                # to make sure that we do not call it a second
+                # time if the client is woken up before a pending
+                # rebalance completes. This must be called on each
+                # iteration of the loop because an event requiring
+                # a rebalance (such as a metadata refresh which
+                # changes the matched subscription set) can occur
+                # while another rebalance is still in progress.
+                if not self.rejoining:
+                    self._on_join_prepare(self._generation.generation_id,
+                                          self._generation.member_id)
+                    self.rejoining = True
+
                 # ensure that there are no pending requests to the coordinator.
                 # This is important in particular to avoid resending a pending
                 # JoinGroup request.

kafka/coordinator/consumer.py

@@ -84,6 +84,8 @@ def __init__(self, client, subscription, metrics, **configs):
                 self.config[key] = configs[key]
 
         self._subscription = subscription
+        self._is_leader = False
+        self._joined_subscription = set()
         self._metadata_snapshot = self._build_metadata_snapshot(subscription, client.cluster)
         self._assignment_snapshot = None
         self._cluster = client.cluster
@@ -119,6 +121,7 @@ def __init__(self, client, subscription, metrics, **configs):
         self.consumer_sensors = ConsumerCoordinatorMetrics(
             metrics, self.config['metric_group_prefix'], self._subscription)
 
+        #self._handle_metadata_update(self._cluster)
         self._cluster.request_update()
         self._cluster.add_listener(WeakMethod(self._handle_metadata_update))
 
@@ -132,11 +135,22 @@ def protocol_type(self):
 
     def group_protocols(self):
         """Returns list of preferred (protocols, metadata)"""
-        topics = self._subscription.subscription
-        assert topics is not None, 'Consumer has not subscribed to topics'
+        if self._subscription.subscription is None:
+            raise Errors.IllegalStateError('Consumer has not subscribed to topics')
+        # dpkp note: I really dislike this.
+        # why? because we are using this strange method group_protocols,
+        # which is seemingly innocuous, to set internal state (_joined_subscription)
+        # that is later used to check whether metadata has changed since we joined a group
+        # but there is no guarantee that this method, group_protocols, will get called
+        # in the correct sequence or that it will only be called when we want it to be.
+        # So this really should be moved elsewhere, but I don't have the energy to
+        # work that out right now. If you read this at some later date after the mutable
+        # state has bitten you... I'm sorry! It mimics the java client, and that's the
+        # best I've got for now.
+        self._joined_subscription = set(self._subscription.subscription)
         metadata_list = []
         for assignor in self.config['assignors']:
-            metadata = assignor.metadata(topics)
+            metadata = assignor.metadata(self._joined_subscription)
             group_protocol = (assignor.name, metadata)
             metadata_list.append(group_protocol)
         return metadata_list
@@ -158,21 +172,29 @@ def _handle_metadata_update(self, cluster):
 
         # check if there are any changes to the metadata which should trigger
         # a rebalance
-        if self._subscription_metadata_changed(cluster):
-
-            if (self.config['api_version'] >= (0, 9)
-                and self.config['group_id'] is not None):
-
-                self._subscription.mark_for_reassignment()
-
-            # If we haven't got group coordinator support,
-            # just assign all partitions locally
-            else:
-                self._subscription.assign_from_subscribed([
-                    TopicPartition(topic, partition)
-                    for topic in self._subscription.subscription
-                    for partition in self._metadata_snapshot[topic]
-                ])
+        if self._subscription.partitions_auto_assigned():
+            metadata_snapshot = self._build_metadata_snapshot(self._subscription, cluster)
+            if self._metadata_snapshot != metadata_snapshot:
+                self._metadata_snapshot = metadata_snapshot
+
+                # If we haven't got group coordinator support,
+                # just assign all partitions locally
+                if self._auto_assign_all_partitions():
+                    self._subscription.assign_from_subscribed([
+                        TopicPartition(topic, partition)
+                        for topic in self._subscription.subscription
+                        for partition in self._metadata_snapshot[topic]
+                    ])
+
+    def _auto_assign_all_partitions(self):
+        # For users that use "subscribe" without group support,
+        # we will simply assign all partitions to this consumer
+        if self.config['api_version'] < (0, 9):
+            return True
+        elif self.config['group_id'] is None:
+            return True
+        else:
+            return False
 
     def _build_metadata_snapshot(self, subscription, cluster):
         metadata_snapshot = {}
@@ -181,16 +203,6 @@ def _build_metadata_snapshot(self, subscription, cluster):
             metadata_snapshot[topic] = set(partitions)
         return metadata_snapshot
 
-    def _subscription_metadata_changed(self, cluster):
-        if not self._subscription.partitions_auto_assigned():
-            return False
-
-        metadata_snapshot = self._build_metadata_snapshot(self._subscription, cluster)
-        if self._metadata_snapshot != metadata_snapshot:
-            self._metadata_snapshot = metadata_snapshot
-            return True
-        return False
-
     def _lookup_assignor(self, name):
         for assignor in self.config['assignors']:
             if assignor.name == name:
@@ -199,12 +211,10 @@ def _lookup_assignor(self, name):
 
     def _on_join_complete(self, generation, member_id, protocol,
                           member_assignment_bytes):
-        # if we were the assignor, then we need to make sure that there have
-        # been no metadata updates since the rebalance begin. Otherwise, we
-        # won't rebalance again until the next metadata change
-        if self._assignment_snapshot is not None and self._assignment_snapshot != self._metadata_snapshot:
-            self._subscription.mark_for_reassignment()
-            return
+        # only the leader is responsible for monitoring for metadata changes
+        # (i.e. partition changes)
+        if not self._is_leader:
+            self._assignment_snapshot = None
 
         assignor = self._lookup_assignor(protocol)
         assert assignor, 'Coordinator selected invalid assignment protocol: %s' % protocol
@@ -307,6 +317,7 @@ def _perform_assignment(self, leader_id, assignment_strategy, members):
         # keep track of the metadata used for assignment so that we can check
         # after rebalance completion whether anything has changed
         self._cluster.request_update()
+        self._is_leader = True
         self._assignment_snapshot = self._metadata_snapshot
 
         log.debug("Performing assignment for group %s using strategy %s"
@@ -338,18 +349,32 @@ def _on_join_prepare(self, generation, member_id):
                               " for group %s failed on_partitions_revoked",
                               self._subscription.listener, self.group_id)
 
-        self._assignment_snapshot = None
-        self._subscription.mark_for_reassignment()
+        self._is_leader = False
+        self._subscription.reset_group_subscription()
 
     def need_rejoin(self):
         """Check whether the group should be rejoined
 
         Returns:
             bool: True if consumer should rejoin group, False otherwise
         """
-        return (self._subscription.partitions_auto_assigned() and
-               (super(ConsumerCoordinator, self).need_rejoin() or
-                self._subscription.needs_partition_assignment))
+        if not self._subscription.partitions_auto_assigned():
+            return False
+
+        if self._auto_assign_all_partitions():
+            return False
+
+        # we need to rejoin if we performed the assignment and metadata has changed
+        if (self._assignment_snapshot is not None
+            and self._assignment_snapshot != self._metadata_snapshot):
+            return True
+
+        # we need to join if our subscription has changed since the last join
+        if (self._joined_subscription is not None
+            and self._joined_subscription != self._subscription.subscription):
+            return True
+
+        return super(ConsumerCoordinator, self).need_rejoin()
 
     def refresh_committed_offsets_if_needed(self):
         """Fetch committed offsets for assigned partitions."""