import asyncio import collections import copy import logging import time import aiokafka.errors as Errors from aiokafka.client import ConnectionGroup, CoordinationType from aiokafka.coordinator.assignors.roundrobin import RoundRobinPartitionAssignor from aiokafka.coordinator.protocol import ConsumerProtocol from aiokafka.protocol.api import Response from aiokafka.protocol.commit import OffsetCommitRequest_v2 as OffsetCommitRequest from aiokafka.protocol.commit import OffsetFetchRequest_v1 as OffsetFetchRequest from aiokafka.protocol.group import ( HeartbeatRequest, JoinGroupRequest, JoinGroupResponse, JoinGroupResponse_v5, LeaveGroupRequest, SyncGroupRequest, ) from aiokafka.structs import OffsetAndMetadata, TopicPartition from aiokafka.util import create_future, create_task log = logging.getLogger(__name__) UNKNOWN_OFFSET = -1 class BaseCoordinator: def __init__( self, client, subscription, *, exclude_internal_topics=True, ): self._client = client self._exclude_internal_topics = exclude_internal_topics self._subscription = subscription self._metadata_snapshot = {} # Is updated by metadata listener self._cluster = client.cluster # update initial subscription state using currently known metadata self._handle_metadata_update(self._cluster) self._cluster.add_listener(self._handle_metadata_update) def _handle_metadata_update(self, cluster): subscription = self._subscription if subscription.subscribed_pattern: topics = [ topic for topic in cluster.topics(self._exclude_internal_topics) if subscription.subscribed_pattern.match(topic) ] if ( subscription.subscription is None or set(topics) != subscription.subscription.topics ): subscription.subscribe_from_pattern(topics) if ( subscription.partitions_auto_assigned() and self._group_subscription is not None ): metadata_snapshot = self._get_metadata_snapshot() if self._metadata_snapshot != metadata_snapshot: log.info( "Metadata for topic has changed from %s to %s. ", self._metadata_snapshot, metadata_snapshot, ) self._metadata_snapshot = metadata_snapshot self._on_metadata_change() def _get_metadata_snapshot(self): partitions_per_topic = {} for topic in self._group_subscription: partitions = self._cluster.partitions_for_topic(topic) or [] # Partitions are always from 0 to N, so no reason to check each # partition separately, only length is enough partitions_per_topic[topic] = len(partitions) return partitions_per_topic class NoGroupCoordinator(BaseCoordinator): """ When `group_id` consumer option is not used we don't have the functionality provided by Coordinator node in Kafka cluster, like committing offsets ( Kafka based offset storage) or automatic partition assignment between consumers. But `GroupCoordinator` class has some other responsibilities, that this class takes care of to avoid code duplication, like: * Static topic partition assignment when we subscribed to topic. Partition changes will be noticed by metadata update and assigned. * Pattern topic subscription. New topics will be noticed by metadata update and added to subscription. """ def __init__(self, *args, **kw): super().__init__(*args, **kw) # Reset all committed points, as the GroupCoordinator would self._reset_committed_task = create_task(self._reset_committed_routine()) def _on_metadata_change(self): self.assign_all_partitions() def assign_all_partitions(self, check_unknown=False): """Assign all partitions from subscribed topics to this consumer. If `check_unknown` we will raise UnknownTopicOrPartitionError if subscribed topic is not found in metadata response. """ partitions = [] for topic in self._subscription.subscription.topics: p_ids = self._cluster.partitions_for_topic(topic) if not p_ids: if check_unknown: raise Errors.UnknownTopicOrPartitionError() else: # We probably just changed subscription during metadata # update. No problem, lets wait for the next metadata # update continue partitions += [TopicPartition(topic, p_id) for p_id in p_ids] # If assignment did not change no need to reset it assignment = self._subscription.subscription.assignment if assignment is None or set(partitions) != assignment.tps: self._subscription.assign_from_subscribed(partitions) async def _reset_committed_routine(self): """Group coordinator will reset committed points to UNKNOWN_OFFSET if no commit is found for group. In the NoGroup mode we need to force it after each assignment """ event_waiter = None try: while True: if self._subscription.subscription is None: await self._subscription.wait_for_subscription() continue assignment = self._subscription.subscription.assignment if assignment is None: await self._subscription.wait_for_assignment() continue commit_refresh_needed = assignment.commit_refresh_needed commit_refresh_needed.clear() for tp in assignment.requesting_committed(): tp_state = assignment.state_value(tp) tp_state.update_committed(OffsetAndMetadata(UNKNOWN_OFFSET, "")) event_waiter = create_task(commit_refresh_needed.wait()) await asyncio.wait( [assignment.unassign_future, event_waiter], return_when=asyncio.FIRST_COMPLETED, ) if not event_waiter.done(): event_waiter.cancel() event_waiter = None except asyncio.CancelledError: pass # Just to make sure we properly close started tasks we cancel # event.wait() task if event_waiter is not None and not event_waiter.done(): event_waiter.cancel() event_waiter = None @property def _group_subscription(self): return self._subscription.subscription.topics async def close(self): self._reset_committed_task.cancel() await self._reset_committed_task self._reset_committed_task = None def check_errors(self): if self._reset_committed_task.done(): # pragma: no cover self._reset_committed_task.result() class GroupCoordinator(BaseCoordinator): """ GroupCoordinator implements group management for single group member by interacting with a designated Kafka broker (the coordinator). Group semantics are provided by extending this class. From a high level, Kafka's group management protocol consists of the following sequence of actions: 1. Group Registration: Group members register with the coordinator providing their own metadata (such as the set of topics they are interested in). 2. Group/Leader Selection: The coordinator (one of Kafka nodes) select the members of the group and chooses one member (one of client's) as the leader. 3. State Assignment: The leader receives metadata for all members and assigns partitions to them. 4. Group Stabilization: Each member receives the state assigned by the leader and begins processing. Between each phase coordinator awaits all clients to respond. If some do not respond in time - it will revoke their membership NOTE: Try to maintain same log messages and behaviour as Java and kafka-python clients: https://github.com/apache/kafka/blob/0.10.1.1/clients/src/main/java/\ org/apache/kafka/clients/consumer/internals/AbstractCoordinator.java https://github.com/apache/kafka/blob/0.10.1.1/clients/src/main/java/\ org/apache/kafka/clients/consumer/internals/ConsumerCoordinator.java """ def __init__( self, client, subscription, *, group_id="aiokafka-default-group", group_instance_id=None, session_timeout_ms=10000, heartbeat_interval_ms=3000, retry_backoff_ms=100, enable_auto_commit=True, auto_commit_interval_ms=5000, assignors=(RoundRobinPartitionAssignor,), exclude_internal_topics=True, max_poll_interval_ms=300000, rebalance_timeout_ms=30000, ): """Initialize the coordination manager. Parameters (see AIOKafkaConsumer) """ # Leader node will keep track of the whole group's metadata. self._group_subscription = None super().__init__( client, subscription, exclude_internal_topics=exclude_internal_topics, ) self._session_timeout_ms = session_timeout_ms self._heartbeat_interval_ms = heartbeat_interval_ms self._max_poll_interval = max_poll_interval_ms / 1000 self._rebalance_timeout_ms = rebalance_timeout_ms self._retry_backoff_ms = retry_backoff_ms self._assignors = assignors self._enable_auto_commit = enable_auto_commit self._auto_commit_interval_ms = auto_commit_interval_ms self.generation = OffsetCommitRequest.DEFAULT_GENERATION_ID self.member_id = JoinGroupRequest[0].UNKNOWN_MEMBER_ID self.group_id = group_id self._group_instance_id = group_instance_id self.coordinator_id = None # Coordination flags and futures self._performed_join_prepare = False self._rejoin_needed_fut = create_future() self._coordinator_dead_fut = create_future() self._coordination_task = create_task(self._coordination_routine()) # Will be started/stopped by coordination task self._heartbeat_task = None self._commit_refresh_task = None # Those are mostly unrecoverable exceptions, but user may perform an # action to handle those (for example add permission for this group). # Thus we set exception and pause coordination until user consumes it. self._pending_exception = None self._error_consumed_fut = None self._coordinator_lookup_lock = asyncio.Lock() # Will synchronize edits to TopicPartitionState.committed, as it may be # changed from user code by calling ``commit()``. self._commit_lock = asyncio.Lock() self._next_autocommit_deadline = ( time.monotonic() + auto_commit_interval_ms / 1000 ) # Will be set on close self._closing = create_future() def _on_metadata_change(self): self.request_rejoin() async def _send_req(self, request): """Send request to coordinator node. In case the coordinator is not ready a respective error will be raised. """ node_id = self.coordinator_id if node_id is None: raise Errors.GroupCoordinatorNotAvailableError() try: resp = await self._client.send( node_id, request, group=ConnectionGroup.COORDINATION ) except Errors.KafkaError as err: log.error( "Error sending %s to node %s [%s] -- marking coordinator dead", request.__class__.__name__, node_id, err, ) self.coordinator_dead() raise return resp def check_errors(self): """Check if coordinator is well and no authorization or unrecoverable errors occurred """ if self._coordination_task.done(): self._coordination_task.result() if self._error_consumed_fut is not None: self._error_consumed_fut.set_result(None) self._error_consumed_fut = None if self._pending_exception is not None: exc = self._pending_exception self._pending_exception = None raise exc def _push_error_to_user(self, exc): """Most critical errors are not something we can continue execution without user action. Well right now we just drop the Consumer, but java client would certainly be ok if we just poll another time, maybe it will need to rejoin, but not fail with GroupAuthorizationFailedError till the end of days... XXX: Research if we can't have the same error several times. For example if user gets GroupAuthorizationFailedError and adds permission for the group, would Consumer work right away or would still raise exception a few times? """ exc = copy.copy(exc) self._subscription.abort_waiters(exc) self._pending_exception = exc self._error_consumed_fut = create_future() return asyncio.wait( [self._error_consumed_fut, self._closing], return_when=asyncio.FIRST_COMPLETED, ) async def close(self): """Close the coordinator, leave the current group and reset local generation/memberId.""" if self._closing.done(): return self._closing.set_result(None) # We must let the coordination task properly finish all pending work if not self._coordination_task.done(): await self._coordination_task await self._stop_heartbeat_task() await self._stop_commit_offsets_refresh_task() await self._maybe_leave_group() def maybe_leave_group(self): task = create_task(self._maybe_leave_group()) return task async def _maybe_leave_group(self): if self.generation > 0 and self._group_instance_id is None: # this is a minimal effort attempt to leave the group. we do not # attempt any resending if the request fails or times out. # Note: do not send this leave request if we are running in static # partition assignment mode (when group_instance_id has been set). version = 0 if self._client.api_version < (0, 11, 0) else 1 request = LeaveGroupRequest[version](self.group_id, self.member_id) try: await self._send_req(request) except Errors.KafkaError as err: log.error("LeaveGroup request failed: %s", err) else: log.info("LeaveGroup request succeeded") self.reset_generation() def _lookup_assignor(self, name): for assignor in self._assignors: if assignor.name == name: return assignor return None async def _on_join_prepare(self, previous_assignment): self._subscription.begin_reassignment() self._group_subscription = None # commit offsets prior to rebalance if auto-commit enabled if previous_assignment is not None: try: await self._maybe_do_last_autocommit(previous_assignment) except Errors.KafkaError as err: # We would retry any retriable commit already log.error("OffsetCommit failed before join, ignoring: %s", err) revoked = previous_assignment.tps else: revoked = set() # execute the user's callback before rebalance log.info( "Revoking previously assigned partitions %s for group %s", revoked, self.group_id, ) if self._subscription.listener: try: res = self._subscription.listener.on_partitions_revoked(revoked) if asyncio.iscoroutine(res): await res except Exception: log.exception( "User provided subscription listener %s" " for group %s failed on_partitions_revoked", self._subscription.listener, self.group_id, ) async def _perform_assignment(self, response: Response): assignment_strategy = response.group_protocol members = response.members assignor = self._lookup_assignor(assignment_strategy) assert assignor, f"Invalid assignment protocol: {assignment_strategy}" member_metadata = {} all_subscribed_topics = set() for member in members: if isinstance(response, JoinGroupResponse_v5): member_id, group_instance_id, metadata_bytes = member elif isinstance( response, (JoinGroupResponse[0], JoinGroupResponse[1], JoinGroupResponse[2]), ): member_id, metadata_bytes = member else: raise RuntimeError("unknown protocol returned from assignment") # noqa: TRY004 metadata = ConsumerProtocol.METADATA.decode(metadata_bytes) member_metadata[member_id] = metadata all_subscribed_topics.update(metadata.subscription) # the leader will begin watching for changes to any of the topics # the group is interested in, which ensures that all metadata changes # will eventually be seen self._group_subscription = all_subscribed_topics if not self._subscription.subscribed_pattern: self._client.set_topics(self._group_subscription) # If somewhere we forced a metadata update (like in some `set_topics` # call) we should wait for it before performing assignment await self._client._maybe_wait_metadata() log.debug( "Performing assignment for group %s using strategy %s" " with subscriptions %s", self.group_id, assignor.name, member_metadata, ) assignments = assignor.assign(self._cluster, member_metadata) log.debug("Finished assignment for group %s: %s", self.group_id, assignments) # `set_topics()` will not trigger a metadata update if we only # removed some topics (client has all needed metadata already), # so we force a snapshot update. self._metadata_snapshot = self._get_metadata_snapshot() return assignments async def _on_join_complete( self, generation, member_id, protocol, member_assignment_bytes ): assignor = self._lookup_assignor(protocol) assert assignor, f"invalid assignment protocol: {protocol}" assignment = ConsumerProtocol.ASSIGNMENT.decode(member_assignment_bytes) # update partition assignment self._subscription.assign_from_subscribed(assignment.partitions()) # The await bellow can change subscription, remember the ongoing one. subscription = self._subscription.subscription # give the assignor a chance to update internal state # based on the received assignment assignor.on_assignment(assignment) # We need to start this task before callback to avoid deadlocks. # Callback can rely on something like ``Consumer.position()`` that # requires committed point to be refreshed. await self._stop_commit_offsets_refresh_task() self.start_commit_offsets_refresh_task(subscription.assignment) assigned = set(self._subscription.assigned_partitions()) log.info( "Setting newly assigned partitions %s for group %s", assigned, self.group_id ) # execute the user's callback after rebalance if self._subscription.listener: try: res = self._subscription.listener.on_partitions_assigned(assigned) if asyncio.iscoroutine(res): await res except Exception: log.exception( "User provided listener %s for group %s" " failed on partition assignment: %s", self._subscription.listener, self.group_id, assigned, ) def coordinator_dead(self): """Mark the current coordinator as dead. NOTE: this will not force a group rejoin. If new coordinator is able to recognize this member we will just continue with current generation. """ if self.coordinator_id is not None: log.warning( "Marking the coordinator dead (node %s)for group %s.", self.coordinator_id, self.group_id, ) self.coordinator_id = None self._coordinator_dead_fut.set_result(None) def reset_generation(self): """Coordinator did not recognize either generation or member_id. Will need to re-join the group. """ self.generation = OffsetCommitRequest.DEFAULT_GENERATION_ID self.member_id = JoinGroupRequest[0].UNKNOWN_MEMBER_ID self.request_rejoin() def request_rejoin(self): if not self._rejoin_needed_fut.done(): self._rejoin_needed_fut.set_result(None) def need_rejoin(self, subscription): """Check whether the group should be rejoined Returns: bool: True if consumer should rejoin group, False otherwise """ return subscription.assignment is None or self._rejoin_needed_fut.done() async def ensure_coordinator_known(self): """Block until the coordinator for this group is known.""" if self.coordinator_id is not None: return async with self._coordinator_lookup_lock: retry_backoff = self._retry_backoff_ms / 1000 while self.coordinator_id is None and not self._closing.done(): try: coordinator_id = await self._client.coordinator_lookup( CoordinationType.GROUP, self.group_id ) except Errors.GroupAuthorizationFailedError as err: new_err = Errors.GroupAuthorizationFailedError(self.group_id) raise new_err from err except Errors.KafkaError as err: log.error("Group Coordinator Request failed: %s", err) if err.retriable: await self._client.force_metadata_update() await asyncio.sleep(retry_backoff) continue else: raise # Try to connect to confirm that the connection can be # established. ready = await self._client.ready( coordinator_id, group=ConnectionGroup.COORDINATION ) if not ready: await asyncio.sleep(retry_backoff) continue self.coordinator_id = coordinator_id self._coordinator_dead_fut = create_future() log.info( "Discovered coordinator %s for group %s", self.coordinator_id, self.group_id, ) async def _coordination_routine(self): try: await self.__coordination_routine() except asyncio.CancelledError: # pragma: no cover raise except Exception as exc: log.exception("Unexpected error in coordinator routine") kafka_exc = Errors.KafkaError( f"Unexpected error during coordination {exc!r}" ) self._subscription.abort_waiters(kafka_exc) raise kafka_exc from exc async def __coordination_routine(self): """Main background task, that keeps track of changes in group coordination. This task will spawn/stop heartbeat task and perform autocommit in times it's safe to do so. """ subscription = self._subscription.subscription assignment = None while not self._closing.done(): # Check if there was a change to subscription if subscription is not None and not subscription.active: # The subscription can change few times, so we can not rely on # flags or topic lists. For example if user changes # subscription from X to Y and back to X we still need to # rejoin group. self.request_rejoin() subscription = self._subscription.subscription if subscription is None: await asyncio.wait( [self._subscription.wait_for_subscription(), self._closing], return_when=asyncio.FIRST_COMPLETED, ) if self._closing.done(): break subscription = self._subscription.subscription assert subscription is not None and subscription.active auto_assigned = self._subscription.partitions_auto_assigned() # Ensure active group try: await self.ensure_coordinator_known() if auto_assigned and self.need_rejoin(subscription): new_assignment = await self.ensure_active_group( subscription, assignment ) if new_assignment is None or not new_assignment.active: continue else: assignment = new_assignment else: assignment = subscription.assignment assert assignment is not None and assignment.active # We will only try to commit offsets once here. In error case # the returned wait_timeout will be ``retry_backoff``. In # success case time to next autocommit deadline. If autocommit # is disabled timeout will be ``None``, ie. no timeout. wait_timeout = await self._maybe_do_autocommit(assignment) except Errors.KafkaError as exc: await self._push_error_to_user(exc) continue futures = [ self._closing, # Will exit fast if close() called self._coordinator_dead_fut, subscription.unsubscribe_future, ] # In case of manual assignment this future will be always set and # we don't want a heavy loop here. # NOTE: metadata changes are for partition count and pattern # subscription, which is irrelevant in case of user assignment. if auto_assigned: futures.append(self._rejoin_needed_fut) # We should always watch for other task raising critical or # unexpected errors, so we attach those as futures too. We will # check them right after wait. if self._heartbeat_task: futures.append(self._heartbeat_task) if self._commit_refresh_task: futures.append(self._commit_refresh_task) done, _ = await asyncio.wait( futures, timeout=wait_timeout, return_when=asyncio.FIRST_COMPLETED ) # Handle exceptions in other background tasks for task in [self._heartbeat_task, self._commit_refresh_task]: if task and task.done(): exc = task.exception() if exc: await self._push_error_to_user(exc) # Closing finallization if assignment is not None: try: await self._maybe_do_last_autocommit(assignment) except Errors.KafkaError as err: # We did all we could, all we can is show this to user log.error("Failed to commit on finallization: %s", err) async def ensure_active_group(self, subscription, prev_assignment): # due to a race condition between the initial metadata # fetch and the initial rebalance, we need to ensure that # the metadata is fresh before joining initially. This # ensures that we have matched the pattern against the # cluster's topics at least once before joining. # Also the rebalance can be issued by another node, that # discovered a new topic, which is still unknown to this # one. if self._subscription.subscribed_pattern: await self._client.force_metadata_update() if not subscription.active: return None if not self._performed_join_prepare: # NOTE: We pass the previously used assignment here. await self._on_join_prepare(prev_assignment) self._performed_join_prepare = True # NOTE: we did not stop heartbeat task before to keep the # member alive during the callback, as it can commit offsets. # See the ``RebalanceInProgressError`` case in heartbeat # handling. await self._stop_heartbeat_task() # We will not attempt rejoin if there is no activity on consumer idle_time = self._subscription.fetcher_idle_time if prev_assignment is not None and idle_time >= self._max_poll_interval: await asyncio.sleep(self._retry_backoff_ms / 1000) return None # We will only try to perform the rejoin once. If it fails, # we will spin this loop another time, checking for coordinator # and subscription changes. # NOTE: We do re-join in sync. The group rebalance will fail on # subscription change and coordinator failure by itself and # this way we don't need to worry about racing or cancellation # issues that could occur if re-join were to be a task. success = await self._do_rejoin_group(subscription) if success: self._performed_join_prepare = False self._start_heartbeat_task() return subscription.assignment return None def _start_heartbeat_task(self): if self._heartbeat_task is None: self._heartbeat_task = create_task(self._heartbeat_routine()) async def _stop_heartbeat_task(self): if self._heartbeat_task is not None: if not self._heartbeat_task.done(): self._heartbeat_task.cancel() await self._heartbeat_task self._heartbeat_task = None async def _heartbeat_routine(self): last_ok_heartbeat = time.monotonic() hb_interval = self._heartbeat_interval_ms / 1000 session_timeout = self._session_timeout_ms / 1000 retry_backoff = self._retry_backoff_ms / 1000 sleep_time = hb_interval # There is no point to heartbeat after Broker stopped recognizing # this consumer, so we stop after resetting generation. while self.member_id != JoinGroupRequest[0].UNKNOWN_MEMBER_ID: try: await asyncio.sleep(sleep_time) await self.ensure_coordinator_known() t0 = time.monotonic() success = await self._do_heartbeat() except asyncio.CancelledError: break # NOTE: We let all other errors propagate up to coordination # routine if success: last_ok_heartbeat = time.monotonic() sleep_time = max((0, hb_interval - last_ok_heartbeat + t0)) else: sleep_time = retry_backoff session_time = time.monotonic() - last_ok_heartbeat if session_time > session_timeout: # the session timeout has expired without seeing a successful # heartbeat, so we should probably make sure the coordinator # is still healthy. log.error("Heartbeat session expired - marking coordinator dead") self.coordinator_dead() # If consumer is idle (no records consumed) for too long we need # to leave the group idle_time = self._subscription.fetcher_idle_time if idle_time < self._max_poll_interval: sleep_time = min(sleep_time, self._max_poll_interval - idle_time) else: await self._maybe_leave_group() log.debug("Stopping heartbeat task") async def _do_heartbeat(self): version = 0 if self._client.api_version < (0, 11, 0) else 1 request = HeartbeatRequest[version]( self.group_id, self.generation, self.member_id ) log.debug( "Heartbeat: %s[%s] %s", self.group_id, self.generation, self.member_id ) # _send_req may fail with error like `RequestTimedOutError` # we need to catch it so coordinator_routine won't fail try: resp = await self._send_req(request) except Errors.KafkaError as err: log.error("Heartbeat send request failed: %s. Will retry.", err) return False error_type = Errors.for_code(resp.error_code) if error_type is Errors.NoError: log.debug( "Received successful heartbeat response for group %s", self.group_id ) return True if error_type in ( Errors.GroupCoordinatorNotAvailableError, Errors.NotCoordinatorForGroupError, ): log.warning( "Heartbeat failed for group %s: coordinator (node %s)" " is either not started or not valid", self.group_id, self.coordinator_id, ) self.coordinator_dead() elif error_type is Errors.RebalanceInProgressError: log.warning( "Heartbeat failed for group %s because it is rebalancing", self.group_id ) self.request_rejoin() # it is valid to continue heartbeating while the group is # rebalancing. This ensures that the coordinator keeps the # member in the group for as long as the duration of the # rebalance timeout. If we stop sending heartbeats, # however, then the session timeout may expire before we # can rejoin. return True elif error_type is Errors.IllegalGenerationError: log.warning( "Heartbeat failed for group %s: generation id is not current.", self.group_id, ) self.reset_generation() elif error_type is Errors.UnknownMemberIdError: log.warning( "Heartbeat failed: local member_id was not recognized;" " resetting and re-joining group" ) self.reset_generation() elif error_type is Errors.GroupAuthorizationFailedError: raise error_type(self.group_id) else: err = Errors.KafkaError( f"Unexpected exception in heartbeat task: {error_type()!r}" ) log.error("Heartbeat failed: %r", err) raise err return False def start_commit_offsets_refresh_task(self, assignment): if self._commit_refresh_task is not None: self._commit_refresh_task.cancel() self._commit_refresh_task = create_task( self._commit_refresh_routine(assignment) ) async def _stop_commit_offsets_refresh_task(self): # The previous task should end after assignment changed if self._commit_refresh_task is not None: if not self._commit_refresh_task.done(): self._commit_refresh_task.cancel() await self._commit_refresh_task self._commit_refresh_task = None async def _commit_refresh_routine(self, assignment): """Task that will do a commit cache refresh if someone is waiting for it. """ retry_backoff_ms = self._retry_backoff_ms / 1000 commit_refresh_needed = assignment.commit_refresh_needed event_waiter = None try: while assignment.active: commit_refresh_needed.clear() success = await self._maybe_refresh_commit_offsets(assignment) wait_futures = [assignment.unassign_future] if not success: timeout = retry_backoff_ms else: timeout = None event_waiter = create_task(commit_refresh_needed.wait()) wait_futures.append(event_waiter) await asyncio.wait( wait_futures, timeout=timeout, return_when=asyncio.FIRST_COMPLETED ) except asyncio.CancelledError: pass except Exception: # Reset event to continue in case user fixes the problem commit_refresh_needed.set() raise # Just to make sure we properly close started tasks we cancel # event.wait() task if event_waiter is not None and not event_waiter.done(): event_waiter.cancel() event_waiter = None async def _do_rejoin_group(self, subscription): rebalance = CoordinatorGroupRebalance( self, self.group_id, self.coordinator_id, subscription, self._assignors, self._session_timeout_ms, self._retry_backoff_ms, ) assignment = await rebalance.perform_group_join() if not subscription.active: log.debug( "Subscription changed during rebalance from %s to %s. " "Rejoining group.", subscription.topics, self._subscription.topics, ) return False if assignment is None: # wait backoff and try again await asyncio.sleep(self._retry_backoff_ms / 1000) return False protocol, member_assignment_bytes = assignment await self._on_join_complete( self.generation, self.member_id, protocol, member_assignment_bytes ) return True async def _maybe_do_autocommit(self, assignment): if not self._enable_auto_commit: return None now = time.monotonic() interval = self._auto_commit_interval_ms / 1000 backoff = self._retry_backoff_ms / 1000 if now > self._next_autocommit_deadline: try: async with self._commit_lock: await self._do_commit_offsets( assignment, assignment.all_consumed_offsets() ) except Errors.KafkaError as error: log.warning("Auto offset commit failed: %s", error) if self._is_commit_retriable(error): # Retry after backoff. self._next_autocommit_deadline = time.monotonic() + backoff return backoff else: raise # If we had an unrecoverable error we expect the user to handle it # from another source (say Fetcher, like authorization errors). self._next_autocommit_deadline = now + interval return max(0, self._next_autocommit_deadline - time.monotonic()) def _is_commit_retriable(self, error): # Java client raises CommitFailedError which is retriable and thus # masks those 3. We raise error that we got explicitly, so treat them # as retriable. return error.retriable or isinstance( error, ( Errors.UnknownMemberIdError, Errors.IllegalGenerationError, Errors.RebalanceInProgressError, ), ) async def _maybe_do_last_autocommit(self, assignment): if not self._enable_auto_commit: return await self.commit_offsets(assignment, assignment.all_consumed_offsets()) async def commit_offsets(self, assignment, offsets): """Commit specific offsets Arguments: offsets (dict {TopicPartition: OffsetAndMetadata}): what to commit Raises KafkaError on failure """ while True: await self.ensure_coordinator_known() try: async with self._commit_lock: await asyncio.shield(self._do_commit_offsets(assignment, offsets)) except ( Errors.UnknownMemberIdError, Errors.IllegalGenerationError, Errors.RebalanceInProgressError, ) as exc: raise Errors.CommitFailedError( "Commit cannot be completed since the group has already " "rebalanced and may have assigned the partitions " "to another member" ) from exc except Errors.KafkaError as err: if not err.retriable: raise else: # wait backoff and try again await asyncio.sleep(self._retry_backoff_ms / 1000) else: break async def _do_commit_offsets(self, assignment, offsets): # Fast return if nothing to commit if not offsets: return # create the offset commit request offset_data = collections.defaultdict(list) for tp, offset in offsets.items(): offset_data[tp.topic].append( (tp.partition, offset.offset, offset.metadata), ) request = OffsetCommitRequest( self.group_id, self.generation, self.member_id, OffsetCommitRequest.DEFAULT_RETENTION_TIME, list(offset_data.items()), ) log.debug( "Sending offset-commit request with %s for group %s to %s", offsets, self.group_id, self.coordinator_id, ) response = await self._send_req(request) errored = collections.OrderedDict() unauthorized_topics = set() for topic, partitions in response.topics: for partition, error_code in partitions: tp = TopicPartition(topic, partition) error_type = Errors.for_code(error_code) offset = offsets[tp] if error_type is Errors.NoError: log.debug("Committed offset %s for partition %s", offset, tp) elif error_type is Errors.GroupAuthorizationFailedError: log.error( "OffsetCommit failed for group %s - %s", self.group_id, error_type.__name__, ) errored[tp] = error_type(self.group_id) elif error_type is Errors.TopicAuthorizationFailedError: unauthorized_topics.add(topic) elif error_type in ( Errors.OffsetMetadataTooLargeError, Errors.InvalidCommitOffsetSizeError, ): # raise the error to the user log.info( "OffsetCommit failed for group %s on partition %s" " due to %s, will retry", self.group_id, tp, error_type.__name__, ) errored[tp] = error_type() elif error_type is Errors.GroupLoadInProgressError: # just retry log.info( "OffsetCommit failed for group %s because group is" " initializing (%s), will retry", self.group_id, error_type.__name__, ) errored[tp] = error_type() elif error_type in ( Errors.GroupCoordinatorNotAvailableError, Errors.NotCoordinatorForGroupError, Errors.RequestTimedOutError, ): log.info( "OffsetCommit failed for group %s due to a" " coordinator error (%s), will find new coordinator" " and retry", self.group_id, error_type.__name__, ) self.coordinator_dead() errored[tp] = error_type() elif error_type in ( Errors.UnknownMemberIdError, Errors.IllegalGenerationError, Errors.RebalanceInProgressError, ): # need to re-join group error = error_type(self.group_id) log.error( "OffsetCommit failed for group %s due to group" " error (%s), will rejoin", self.group_id, error, ) if error_type is Errors.RebalanceInProgressError: self.request_rejoin() else: self.reset_generation() errored[tp] = error else: log.error( "OffsetCommit failed for group %s on partition %s" " with offset %s: %s", self.group_id, tp, offset, error_type.__name__, ) errored[tp] = error_type() if errored: first_error, *_ = errored.values() raise first_error if unauthorized_topics: log.error( "OffsetCommit failed for unauthorized topics %s", unauthorized_topics ) raise Errors.TopicAuthorizationFailedError(unauthorized_topics) async def _maybe_refresh_commit_offsets(self, assignment): need_update = assignment.requesting_committed() if need_update: try: offsets = await self._do_fetch_commit_offsets(need_update) except Errors.KafkaError as err: if not err.retriable: raise else: log.debug("Failed to fetch committed offsets: %r", err) return False for tp in need_update: tp_state = assignment.state_value(tp) if tp in offsets: tp_state.update_committed(offsets[tp]) else: tp_state.update_committed(OffsetAndMetadata(UNKNOWN_OFFSET, "")) return True async def fetch_committed_offsets(self, partitions): """Fetch the current committed offsets for specified partitions Arguments: partitions (list of TopicPartition): partitions to fetch Returns: dict: {TopicPartition: OffsetAndMetadata} """ if not partitions: return {} while True: await self.ensure_coordinator_known() try: offsets = await self._do_fetch_commit_offsets(partitions) except Errors.KafkaError as err: if not err.retriable: raise else: # wait backoff and try again await asyncio.sleep(self._retry_backoff_ms / 1000) else: return offsets async def _do_fetch_commit_offsets(self, partitions): log.debug("Fetching committed offsets for partitions: %s", partitions) # construct the request partitions_by_topic = collections.defaultdict(list) for tp in partitions: partitions_by_topic[tp.topic].append(tp.partition) request = OffsetFetchRequest(self.group_id, list(partitions_by_topic.items())) response = await self._send_req(request) offsets = {} for topic, topic_partitions in response.topics: for partition, offset, metadata, error_code in topic_partitions: tp = TopicPartition(topic, partition) error_type = Errors.for_code(error_code) if error_type is not Errors.NoError: error = error_type() log.debug("Error fetching offset for %s: %s", tp, error) if error_type is Errors.GroupLoadInProgressError: # just retry raise error elif error_type is Errors.NotCoordinatorForGroupError: # re-discover the coordinator and retry self.coordinator_dead() raise error elif error_type is Errors.UnknownTopicOrPartitionError: log.warning("OffsetFetchRequest -- unknown topic %s", topic) continue elif error_type is Errors.GroupAuthorizationFailedError: raise error_type(self.group_id) else: log.error( "Unknown error fetching offsets for %s: %s", tp, error ) raise Errors.KafkaError(repr(error)) # record the position with the offset # (-1 indicates no committed offset to fetch) if offset == UNKNOWN_OFFSET: log.debug("No committed offset for partition %s", tp) else: offsets[tp] = OffsetAndMetadata(offset, metadata) return offsets class CoordinatorGroupRebalance: """ An adapter, that encapsulates rebalance logic and will have a copy of assigned topics, so we can detect assignment changes. This includes subscription pattern changes. On how to handle cases read in https://cwiki.apache.org/confluence/\ display/KAFKA/Kafka+Client-side+Assignment+Proposal """ def __init__( self, coordinator, group_id, coordinator_id, subscription, assignors, session_timeout_ms, retry_backoff_ms, ): self._coordinator = coordinator self.group_id = group_id self.coordinator_id = coordinator_id self._subscription = subscription self._assignors = assignors self._session_timeout_ms = session_timeout_ms self._retry_backoff_ms = retry_backoff_ms self._api_version = self._coordinator._client.api_version self._rebalance_timeout_ms = self._coordinator._rebalance_timeout_ms async def perform_group_join(self): """Join the group and return the assignment for the next generation. This function handles both JoinGroup and SyncGroup, delegating to _perform_assignment() if elected as leader by the coordinator node. Returns encoded-bytes assignment returned from the group leader """ # send a join group request to the coordinator log.info("(Re-)joining group %s", self.group_id) topics = self._subscription.topics metadata_list = [] for assignor in self._assignors: metadata = assignor.metadata(topics) if not isinstance(metadata, bytes): metadata = metadata.encode() group_protocol = (assignor.name, metadata) metadata_list.append(group_protocol) # for KIP-394 we may have to send a second join request try_join = True while try_join: try_join = False if self._api_version < (0, 10, 1): request = JoinGroupRequest[0]( self.group_id, self._session_timeout_ms, self._coordinator.member_id, ConsumerProtocol.PROTOCOL_TYPE, metadata_list, ) elif self._api_version < (0, 11, 0): request = JoinGroupRequest[1]( self.group_id, self._session_timeout_ms, self._rebalance_timeout_ms, self._coordinator.member_id, ConsumerProtocol.PROTOCOL_TYPE, metadata_list, ) elif self._api_version < (2, 3, 0): request = JoinGroupRequest[2]( self.group_id, self._session_timeout_ms, self._rebalance_timeout_ms, self._coordinator.member_id, ConsumerProtocol.PROTOCOL_TYPE, metadata_list, ) else: request = JoinGroupRequest[3]( self.group_id, self._session_timeout_ms, self._rebalance_timeout_ms, self._coordinator.member_id, self._coordinator._group_instance_id, ConsumerProtocol.PROTOCOL_TYPE, metadata_list, ) # create the request for the coordinator log.debug( "Sending JoinGroup (%s) to coordinator %s", request, self.coordinator_id, ) try: response = await self._coordinator._send_req(request) except Errors.KafkaError: # Return right away. It's a connection error, so backoff will be # handled by coordinator lookup return None if not self._subscription.active: # Subscription changed. Ignore response and restart group join return None error_type = Errors.for_code(response.error_code) if error_type is Errors.MemberIdRequired: self._coordinator.member_id = response.member_id try_join = True if error_type is Errors.NoError: log.debug("Join group response %s", response) self._coordinator.member_id = response.member_id self._coordinator.generation = response.generation_id protocol = response.group_protocol log.info( "Joined group '%s' (generation %s) with member_id %s", self.group_id, response.generation_id, response.member_id, ) if response.leader_id == response.member_id: log.info( "Elected group leader -- performing partition" " assignments using %s", protocol, ) assignment_bytes = await self._on_join_leader(response) else: assignment_bytes = await self._on_join_follower() if assignment_bytes is None: return None return (protocol, assignment_bytes) elif error_type is Errors.GroupLoadInProgressError: # Backoff and retry log.debug( "Attempt to join group %s rejected since coordinator %s" " is loading the group.", self.group_id, self.coordinator_id, ) await asyncio.sleep(self._retry_backoff_ms / 1000) elif error_type is Errors.UnknownMemberIdError: # reset the member id and retry immediately self._coordinator.reset_generation() log.debug( "Attempt to join group %s failed due to unknown member id", self.group_id, ) elif error_type in ( Errors.GroupCoordinatorNotAvailableError, Errors.NotCoordinatorForGroupError, ): # Coordinator changed we should be able to find it immediately err = error_type() self._coordinator.coordinator_dead() log.debug( "Attempt to join group %s failed due to obsolete " "coordinator information: %s", self.group_id, err, ) elif error_type in ( Errors.InconsistentGroupProtocolError, Errors.InvalidSessionTimeoutError, Errors.InvalidGroupIdError, ): err = error_type() log.error("Attempt to join group failed due to fatal error: %s", err) raise err elif error_type is Errors.GroupAuthorizationFailedError: raise error_type(self.group_id) else: err = error_type() log.error( "Unexpected error in join group '%s' response: %s", self.group_id, err ) raise Errors.KafkaError(repr(err)) return None async def _on_join_follower(self): # send follower's sync group with an empty assignment if self._api_version < (2, 3, 0): version = 0 if self._api_version < (0, 11, 0) else 1 request = SyncGroupRequest[version]( self.group_id, self._coordinator.generation, self._coordinator.member_id, [], ) else: request = SyncGroupRequest[2]( self.group_id, self._coordinator.generation, self._coordinator.member_id, self._coordinator._group_instance_id, [], ) log.debug( "Sending follower SyncGroup for group %s to coordinator %s: %s", self.group_id, self.coordinator_id, request, ) return await self._send_sync_group_request(request) async def _on_join_leader(self, response): """ Perform leader synchronization and send back the assignment for the group via SyncGroupRequest Arguments: response (JoinResponse): broker response to parse Returns: Future: resolves to member assignment encoded-bytes """ try: group_assignment = await self._coordinator._perform_assignment(response) except Exception as e: raise Errors.KafkaError(repr(e)) from e assignment_req = [] for member_id, assignment in group_assignment.items(): if not isinstance(assignment, bytes): assignment = assignment.encode() assignment_req.append((member_id, assignment)) if self._api_version < (2, 3, 0): version = 0 if self._api_version < (0, 11, 0) else 1 request = SyncGroupRequest[version]( self.group_id, self._coordinator.generation, self._coordinator.member_id, assignment_req, ) else: request = SyncGroupRequest[2]( self.group_id, self._coordinator.generation, self._coordinator.member_id, self._coordinator._group_instance_id, assignment_req, ) log.debug( "Sending leader SyncGroup for group %s to coordinator %s: %s", self.group_id, self.coordinator_id, request, ) return await self._send_sync_group_request(request) async def _send_sync_group_request(self, request): # We need to reset the rejoin future right after the assignment to # capture metadata changes after join group was performed. We do not # set it directly after JoinGroup to avoid a false rejoin in case # ``_perform_assignment()`` does a metadata update. self._coordinator._rejoin_needed_fut = create_future() req_generation = self._coordinator.generation req_member_id = self._coordinator.member_id try: response = await self._coordinator._send_req(request) except Errors.KafkaError: # We lost connection to coordinator. No need to try and finish this # group join, just rejoin again. self._coordinator.request_rejoin() return None error_type = Errors.for_code(response.error_code) if error_type is Errors.NoError: log.info( "Successfully synced group %s with generation %s", self.group_id, self._coordinator.generation, ) # make sure the right member_id/generation is set in case they changed # while the rejoin was taking place self._coordinator.generation = req_generation self._coordinator.member_id = req_member_id return response.member_assignment # Error case self._coordinator.request_rejoin() if error_type is Errors.RebalanceInProgressError: log.debug( "SyncGroup for group %s failed due to group rebalance", self.group_id ) elif error_type in (Errors.UnknownMemberIdError, Errors.IllegalGenerationError): err = error_type() log.debug("SyncGroup for group %s failed due to %s,", self.group_id, err) self._coordinator.reset_generation() elif error_type in ( Errors.GroupCoordinatorNotAvailableError, Errors.NotCoordinatorForGroupError, ): err = error_type() log.debug("SyncGroup for group %s failed due to %s", self.group_id, err) self._coordinator.coordinator_dead() elif error_type is Errors.GroupAuthorizationFailedError: raise error_type(self.group_id) else: err = error_type() log.error("Unexpected error from SyncGroup: %s", err) raise Errors.KafkaError(repr(err)) return None