def encode_list_offset_request(cls, payloads=()):
return kafka.protocol.offset.OffsetRequest[1](
replica_id=-1,
topics=[(
topic,
[(
partition,
payload.time)
for partition, payload in six.iteritems(topic_payloads)])
for topic, topic_payloads in six.iteritems(group_by_topic_and_partition(payloads))])
def _send_offset_requests(self, timestamps):
"""Fetch offsets for each partition in timestamps dict. This may send
request to multiple nodes, based on who is Leader for partition.
Arguments:
timestamps (dict): {TopicPartition: int} mapping of fetching
timestamps.
Returns:
Future: resolves to a mapping of retrieved offsets
"""
timestamps_by_node = collections.defaultdict(dict)
for partition, timestamp in six.iteritems(timestamps):
node_id = self._client.cluster.leader_for_partition(partition)
if node_id is None:
self._client.add_topic(partition.topic)
log.debug("Partition %s is unknown for fetching offset,"
" wait for metadata refresh", partition)
return Future().failure(Errors.StaleMetadata(partition))
elif node_id == -1:
log.debug("Leader for partition %s unavailable for fetching "
"offset, wait for metadata refresh", partition)
return Future().failure(
Errors.LeaderNotAvailableError(partition))
else:
timestamps_by_node[node_id][partition] = timestamp
# Aggregate results until we have all
list_offsets_future = Future()
responses = []
node_count = len(timestamps_by_node)
def on_success(value):
responses.append(value)
if len(responses) == node_count:
offsets = {}
for r in responses:
offsets.update(r)
list_offsets_future.success(offsets)
def on_fail(err):
if not list_offsets_future.is_done:
list_offsets_future.failure(err)
for node_id, timestamps in six.iteritems(timestamps_by_node):
_f = self._send_offset_request(node_id, timestamps)
_f.add_callback(on_success)
_f.add_errback(on_fail)
return list_offsets_future
def _perform_assignment(self, leader_id, assignment_strategy, members):
assignor = self._lookup_assignor(assignment_strategy)
assert assignor, 'Invalid assignment protocol: %s' % assignment_strategy
member_metadata = {}
all_subscribed_topics = set()
for member_id, metadata_bytes in members:
metadata = ConsumerProtocol.METADATA.decode(metadata_bytes)
member_metadata[member_id] = metadata
all_subscribed_topics.update(metadata.subscription) # pylint: disable-msg=no-member
# 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
# Because assignment typically happens within response callbacks,
# we cannot block on metadata updates here (no recursion into poll())
self._subscription.group_subscribe(all_subscribed_topics)
self._client.set_topics(self._subscription.group_subscription())
# 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._assignment_snapshot = self._metadata_snapshot
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)
group_assignment = {}
for member_id, assignment in six.iteritems(assignments):
group_assignment[member_id] = assignment
return group_assignment
def _raise_if_offset_out_of_range(self):
"""Check FetchResponses for offset out of range.
Raises:
OffsetOutOfRangeError: if any partition from previous FetchResponse
contains OffsetOutOfRangeError and the default_reset_policy is
None
"""
if not self._offset_out_of_range_partitions:
return
current_out_of_range_partitions = {}
# filter only the fetchable partitions
for partition, offset in six.iteritems(self._offset_out_of_range_partitions):
if not self._subscriptions.is_fetchable(partition):
log.debug("Ignoring fetched records for %s since it is no"
" longer fetchable", partition)
continue
position = self._subscriptions.assignment[partition].position
# ignore partition if the current position != offset in FetchResponse
# e.g. after seek()
if position is not None and offset == position:
current_out_of_range_partitions[partition] = position
self._offset_out_of_range_partitions.clear()
if current_out_of_range_partitions:
raise Errors.OffsetOutOfRangeError(current_out_of_range_partitions)
def fetchable_partitions(self):
"""Return set of TopicPartitions that should be Fetched."""
fetchable = set()
for partition, state in six.iteritems(self.assignment):
if state.is_fetchable():
fetchable.add(partition)
return fetchable
def all_consumed_offsets(self):
"""Returns consumed offsets as {TopicPartition: OffsetAndMetadata}"""
all_consumed = {}
for partition, state in six.iteritems(self.assignment):
if state.has_valid_position:
all_consumed[partition] = OffsetAndMetadata(state.position, '')
return all_consumed
def _produce_request(self, node_id, acks, timeout, batches):
"""Create a produce request from the given record batches.
Returns:
ProduceRequest (version depends on api_version)
"""
produce_records_by_partition = collections.defaultdict(dict)
for batch in batches:
topic = batch.topic_partition.topic
partition = batch.topic_partition.partition
# TODO: bytearray / memoryview
buf = batch.records.buffer()
produce_records_by_partition[topic][partition] = buf
if self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
return ProduceRequest[version](
required_acks=acks,
timeout=timeout,
topics=[(topic, list(partition_info.items()))
for topic, partition_info
in six.iteritems(produce_records_by_partition)]
)
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 = self._perform_assignment(response.leader_id,
response.group_protocol,
response.members)
except Exception as e:
return Future().failure(e)
request = SyncGroupRequest[0](
self.group_id,
self.generation,
self.member_id,
[(member_id,
assignment if isinstance(assignment, bytes) else assignment.encode())
for member_id, assignment in six.iteritems(group_assignment)])
log.debug("Sending leader SyncGroup for group %s to coordinator %s: %s",
self.group_id, self.coordinator_id, request)
return self._send_sync_group_request(request)
def refresh_committed_offsets_if_needed(self):
"""Fetch committed offsets for assigned partitions."""
if self._subscription.needs_fetch_committed_offsets:
offsets = self.fetch_committed_offsets(self._subscription.assigned_partitions())
for partition, offset in six.iteritems(offsets):
# verify assignment is still active
if self._subscription.is_assigned(partition):
self._subscription.assignment[partition].committed = offset.offset
self._subscription.needs_fetch_committed_offsets = False
def encode_offset_commit_request(cls, group, payloads):
"""
Encode an OffsetCommitRequest struct
Arguments:
group: string, the consumer group you are committing offsets for
payloads: list of OffsetCommitRequestPayload
"""
return kafka.protocol.commit.OffsetCommitRequest[0](
consumer_group=group,
topics=[(
topic,
[(
partition,
payload.offset,
payload.metadata)
for partition, payload in six.iteritems(topic_payloads)])
for topic, topic_payloads in six.iteritems(group_by_topic_and_partition(payloads))])
def encode_offset_commit_request_kafka(cls, group, payloads):
"""
Encode an OffsetCommitRequest struct
Arguments:
group: string, the consumer group you are committing offsets for
payloads: list of OffsetCommitRequestPayload
"""
return kafka.protocol.commit.OffsetCommitRequest[2](
consumer_group=group,
consumer_group_generation_id=kafka.protocol.commit.OffsetCommitRequest[2].DEFAULT_GENERATION_ID,
consumer_id='',
retention_time=kafka.protocol.commit.OffsetCommitRequest[2].DEFAULT_RETENTION_TIME,
topics=[(
topic,
[(
partition,
payload.offset,
payload.metadata)
for partition, payload in six.iteritems(topic_payloads)])
for topic, topic_payloads in six.iteritems(group_by_topic_and_partition(payloads))])
def _create_fetch_requests(self):
"""Create fetch requests for all assigned partitions, grouped by node.
FetchRequests skipped if no leader, or node has requests in flight
Returns:
dict: {node_id: FetchRequest, ...} (version depends on api_version)
"""
# create the fetch info as a dict of lists of partition info tuples
# which can be passed to FetchRequest() via .items()
fetchable = collections.defaultdict(lambda: collections.defaultdict(list))
# avoid re-fetching pending offsets
pending = set()
for fetch_offset, tp, _ in self._records:
pending.add((tp, fetch_offset))
for partition in self._subscriptions.fetchable_partitions():
node_id = self._client.cluster.leader_for_partition(partition)
position = self._subscriptions.assignment[partition].position
# fetch if there is a leader, no in-flight requests, and no _records
if node_id is None or node_id == -1:
log.debug("No leader found for partition %s."
" Requesting metadata update", partition)
self._client.cluster.request_update()
elif ((partition, position) not in pending and
self._client.in_flight_request_count(node_id) == 0):
partition_info = (
partition.partition,
position,
self.config['max_partition_fetch_bytes']
)
fetchable[node_id][partition.topic].append(partition_info)
log.debug("Adding fetch request for partition %s at offset %d",
partition, position)
if self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
requests = {}
for node_id, partition_data in six.iteritems(fetchable):
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
partition_data.items())
return requests
def _init_fetches(self):
futures = []
for node_id, request in six.iteritems(self._create_fetch_requests()):
if self._client.ready(node_id):
log.debug("Sending FetchRequest to node %s", node_id)
future = self._client.send(node_id, request)
future.add_callback(self._handle_fetch_response, request, time.time())
future.add_errback(log.error, 'Fetch to node %s failed: %s', node_id)
futures.append(future)
self._fetch_futures.extend(futures)
self._clean_done_fetch_futures()
return futures
def available_partitions_for_topic(self, topic):
"""Return set of partitions with known leaders
Arguments:
topic (str): topic to check for partitions
Returns:
set: {partition (int), ...}
"""
if topic not in self._partitions:
return None
return set([partition for partition, metadata
in six.iteritems(self._partitions[topic])
if metadata.leader != -1])
def _send_offset_request(self, node_id, timestamps):
by_topic = collections.defaultdict(list)
for tp, timestamp in six.iteritems(timestamps):
if self.config['api_version'] >= (0, 10, 1):
data = (tp.partition, timestamp)
else:
data = (tp.partition, timestamp, 1)
by_topic[tp.topic].append(data)
if self.config['api_version'] >= (0, 10, 1):
request = OffsetRequest[1](-1, list(six.iteritems(by_topic)))
else:
request = OffsetRequest[0](-1, list(six.iteritems(by_topic)))
# Client returns a future that only fails on network issues
# so create a separate future and attach a callback to update it
# based on response error codes
future = Future()
_f = self._client.send(node_id, request)
_f.add_callback(self._handle_offset_response, future)
_f.add_errback(lambda e: future.failure(e))
return future
def _send_offset_request(self, node_id, timestamps):
by_topic = collections.defaultdict(list)
for tp, timestamp in six.iteritems(timestamps):
if self.config['api_version'] >= (0, 10, 1):
data = (tp.partition, timestamp)
else:
data = (tp.partition, timestamp, 1)
by_topic[tp.topic].append(data)
if self.config['api_version'] >= (0, 10, 1):
request = OffsetRequest[1](-1, list(six.iteritems(by_topic)))
else:
request = OffsetRequest[0](-1, list(six.iteritems(by_topic)))
# Client returns a future that only fails on network issues
# so create a separate future and attach a callback to update it
# based on response error codes
future = Future()
_f = self._client.send(node_id, request)
_f.add_callback(self._handle_offset_response, future)
_f.add_errback(lambda e: future.failure(e))
return future
def available_partitions_for_topic(self, topic):
"""Return set of partitions with known leaders
Arguments:
topic (str): topic to check for partitions
Returns:
set: {partition (int), ...}
"""
if topic not in self._partitions:
return None
return set([partition for partition, metadata
in six.iteritems(self._partitions[topic])
if metadata.leader != -1])
def assign(cls, cluster, member_metadata):
consumers_per_topic = collections.defaultdict(list)
for member, metadata in six.iteritems(member_metadata):
for topic in metadata.subscription:
consumers_per_topic[topic].append(member)
# construct {member_id: {topic: [partition, ...]}}
assignment = collections.defaultdict(dict)
for topic, consumers_for_topic in six.iteritems(consumers_per_topic):
partitions = cluster.partitions_for_topic(topic)
if partitions is None:
log.warning('No partition metadata for topic %s', topic)
continue
partitions = sorted(list(partitions))
partitions_for_topic = len(partitions)
consumers_for_topic.sort()
partitions_per_consumer = len(partitions) // len(consumers_for_topic)
consumers_with_extra = len(partitions) % len(consumers_for_topic)
for i in range(len(consumers_for_topic)):
start = partitions_per_consumer * i
start += min(i, consumers_with_extra)
length = partitions_per_consumer
if not i + 1 > consumers_with_extra:
length += 1
member = consumers_for_topic[i]
assignment[member][topic] = partitions[start:start+length]
protocol_assignment = {}
for member_id in member_metadata:
protocol_assignment[member_id] = ConsumerProtocolMemberAssignment(
cls.version,
sorted(assignment[member_id].items()),
b'')
return protocol_assignment
def assign(cls, cluster, member_metadata):
consumers_per_topic = collections.defaultdict(list)
for member, metadata in six.iteritems(member_metadata):
for topic in metadata.subscription:
consumers_per_topic[topic].append(member)
# construct {member_id: {topic: [partition, ...]}}
assignment = collections.defaultdict(dict)
for topic, consumers_for_topic in six.iteritems(consumers_per_topic):
partitions = cluster.partitions_for_topic(topic)
if partitions is None:
log.warning('No partition metadata for topic %s', topic)
continue
partitions = sorted(list(partitions))
partitions_for_topic = len(partitions)
consumers_for_topic.sort()
partitions_per_consumer = len(partitions) // len(consumers_for_topic)
consumers_with_extra = len(partitions) % len(consumers_for_topic)
for i in range(len(consumers_for_topic)):
start = partitions_per_consumer * i
start += min(i, consumers_with_extra)
length = partitions_per_consumer
if not i + 1 > consumers_with_extra:
length += 1
member = consumers_for_topic[i]
assignment[member][topic] = partitions[start:start+length]
protocol_assignment = {}
for member_id in member_metadata:
protocol_assignment[member_id] = ConsumerProtocolMemberAssignment(
cls.version,
sorted(assignment[member_id].items()),
b'')
return protocol_assignment
def test_conflicting_previous_assignments(mocker):
cluster = create_cluster(mocker, topics={'t'}, topics_partitions={0, 1})
subscriptions = {
'C1': {'t'},
'C2': {'t'},
}
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
# assume both C1 and C2 have partition 1 assigned to them in generation 1
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, [TopicPartition('t', 0),
TopicPartition('t', 0)], 1)
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
def _create_fetch_requests(self):
"""Create fetch requests for all assigned partitions, grouped by node.
FetchRequests skipped if no leader, or node has requests in flight
Returns:
dict: {node_id: FetchRequest, ...} (version depends on api_version)
"""
# create the fetch info as a dict of lists of partition info tuples
# which can be passed to FetchRequest() via .items()
fetchable = collections.defaultdict(lambda: collections.defaultdict(list))
for partition in self._fetchable_partitions():
node_id = self._client.cluster.leader_for_partition(partition)
position = self._subscriptions.assignment[partition].position
# fetch if there is a leader and no in-flight requests
if node_id is None or node_id == -1:
log.debug("No leader found for partition %s."
" Requesting metadata update", partition)
self._client.cluster.request_update()
elif self._client.in_flight_request_count(node_id) == 0:
partition_info = (
partition.partition,
position,
self.config['max_partition_fetch_bytes']
)
fetchable[node_id][partition.topic].append(partition_info)
log.debug("Adding fetch request for partition %s at offset %d",
partition, position)
if self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
requests = {}
for node_id, partition_data in six.iteritems(fetchable):
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
partition_data.items())
return requests
def send_fetches(self):
"""Send FetchRequests for all assigned partitions that do not already have
an in-flight fetch or pending fetch data.
Returns:
List of Futures: each future resolves to a FetchResponse
"""
futures = []
for node_id, request in six.iteritems(self._create_fetch_requests()):
if self._client.ready(node_id):
log.debug("Sending FetchRequest to node %s", node_id)
future = self._client.send(node_id, request)
future.add_callback(self._handle_fetch_response, request, time.time())
future.add_errback(log.error, 'Fetch to node %s failed: %s', node_id)
futures.append(future)
self._fetch_futures.extend(futures)
self._clean_done_fetch_futures()
return futures
def _create_produce_requests(self, collated):
"""
Transfer the record batches into a list of produce requests on a
per-node basis.
Arguments:
collated: {node_id: [RecordBatch]}
Returns:
dict: {node_id: ProduceRequest} (version depends on api_version)
"""
requests = {}
for node_id, batches in six.iteritems(collated):
requests[node_id] = self._produce_request(
node_id, self.config['acks'],
self.config['request_timeout_ms'], batches)
return requests
def metadata(cls, topics):
if cls.member_assignment is None:
log.debug("No member assignment available")
user_data = b""
else:
log.debug(
"Member assignment is available, generating the metadata: generation {}"
.format(cls.generation))
partitions_by_topic = defaultdict(list)
for (topic_partition) in cls.member_assignment: # pylint: disable=not-an-iterable
partitions_by_topic[topic_partition.topic].append(
topic_partition.partition)
data = StickyAssignorUserDataV1(six.iteritems(partitions_by_topic),
cls.generation)
user_data = data.encode()
return ConsumerProtocolMemberMetadata(cls.version, list(topics),
user_data)
def send_fetches(self):
"""Send FetchRequests for all assigned partitions that do not already have
an in-flight fetch or pending fetch data.
Returns:
List of Futures: each future resolves to a FetchResponse
"""
futures = []
for node_id, request in six.iteritems(self._create_fetch_requests()):
if self._client.ready(node_id):
log.debug("Sending FetchRequest to node %s", node_id)
future = self._client.send(node_id, request)
future.add_callback(self._handle_fetch_response, request, time.time())
future.add_errback(log.error, 'Fetch to node %s failed: %s', node_id)
futures.append(future)
self._fetch_futures.extend(futures)
self._clean_done_fetch_futures()
return futures
def encode_offset_fetch_request(cls, group, payloads, from_kafka=False):
"""
Encode an OffsetFetchRequest struct. The request is encoded using
version 0 if from_kafka is false, indicating a request for Zookeeper
offsets. It is encoded using version 1 otherwise, indicating a request
for Kafka offsets.
Arguments:
group: string, the consumer group you are fetching offsets for
payloads: list of OffsetFetchRequestPayload
from_kafka: bool, default False, set True for Kafka-committed offsets
"""
version = 1 if from_kafka else 0
return kafka.protocol.commit.OffsetFetchRequest[version](
consumer_group=group,
topics=[(topic, list(topic_payloads.keys()))
for topic, topic_payloads in six.iteritems(
group_by_topic_and_partition(payloads))])
def send_fetches(self):
"""Send FetchRequests asynchronously for all assigned partitions.
Note: noop if there are unconsumed records internal to the fetcher
Returns:
List of Futures: each future resolves to a FetchResponse
"""
futures = []
for node_id, request in six.iteritems(self._create_fetch_requests()):
if self._client.ready(node_id):
log.debug("Sending FetchRequest to node %s", node_id)
future = self._client.send(node_id, request)
future.add_callback(self._handle_fetch_response, request, time.time())
future.add_errback(log.error, 'Fetch to node %s failed: %s', node_id)
futures.append(future)
self._fetch_futures.extend(futures)
self._clean_done_fetch_futures()
return futures
def encode_offset_fetch_request(cls, group, payloads, from_kafka=False):
"""
Encode an OffsetFetchRequest struct. The request is encoded using
version 0 if from_kafka is false, indicating a request for Zookeeper
offsets. It is encoded using version 1 otherwise, indicating a request
for Kafka offsets.
Arguments:
group: string, the consumer group you are fetching offsets for
payloads: list of OffsetFetchRequestPayload
from_kafka: bool, default False, set True for Kafka-committed offsets
"""
version = 1 if from_kafka else 0
return kafka.protocol.commit.OffsetFetchRequest[version](
consumer_group=group,
topics=[(
topic,
list(topic_payloads.keys()))
for topic, topic_payloads in six.iteritems(group_by_topic_and_partition(payloads))])
def test_kafka_consumer_max_bytes_simple(self):
self.send_messages(0, range(100, 200))
self.send_messages(1, range(200, 300))
# Start a consumer
consumer = self.kafka_consumer(
auto_offset_reset='earliest', fetch_max_bytes=300)
seen_partitions = set([])
for i in range(10):
poll_res = consumer.poll(timeout_ms=100)
for partition, msgs in six.iteritems(poll_res):
for msg in msgs:
seen_partitions.add(partition)
# Check that we fetched at least 1 message from both partitions
self.assertEqual(
seen_partitions, set([
TopicPartition(self.topic, 0), TopicPartition(self.topic, 1)]))
consumer.close()
def test_kafka_consumer_max_bytes_simple(self):
self.send_messages(0, range(100, 200))
self.send_messages(1, range(200, 300))
# Start a consumer
consumer = self.kafka_consumer(
auto_offset_reset='earliest', fetch_max_bytes=300)
seen_partitions = set([])
for i in range(10):
poll_res = consumer.poll(timeout_ms=100)
for partition, msgs in six.iteritems(poll_res):
for msg in msgs:
seen_partitions.add(partition)
# Check that we fetched at least 1 message from both partitions
self.assertEqual(
seen_partitions, set([
TopicPartition(self.topic, 0), TopicPartition(self.topic, 1)]))
consumer.close()
def _perform_assignment(self, leader_id, assignment_strategy, members):
assignor = self._lookup_assignor(assignment_strategy)
assert assignor, "Invalid assignment protocol: %s" % (
assignment_strategy, )
member_metadata = {}
all_subscribed_topics = set()
for member_id, metadata_bytes in members:
metadata = ConsumerProtocol.METADATA.decode(metadata_bytes)
member_metadata[member_id] = metadata
all_subscribed_topics.update(
metadata.subscription) # pylint: disable-msg=no-member
# 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
# Because assignment typically happens within response callbacks,
# we cannot block on metadata updates here (no recursion into poll())
self._subscription.group_subscribe(all_subscribed_topics)
self._client.set_topics(self._subscription.group_subscription())
# 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"
" 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)
group_assignment = {}
for member_id, assignment in six.iteritems(assignments):
group_assignment[member_id] = assignment
return group_assignment
def metrics(self, raw=False):
"""Get metrics on producer performance.
This is ported from the Java Producer, for details see:
https://kafka.apache.org/documentation/#producer_monitoring
Warning:
This is an unstable interface. It may change in future
releases without warning.
"""
if raw:
return self._metrics.metrics.copy()
metrics = {}
for k, v in six.iteritems(self._metrics.metrics.copy()):
if k.group not in metrics:
metrics[k.group] = {}
if k.name not in metrics[k.group]:
metrics[k.group][k.name] = {}
metrics[k.group][k.name] = v.value()
return metrics
def metrics(self, raw=False):
"""Get metrics on producer performance.
This is ported from the Java Producer, for details see:
https://kafka.apache.org/documentation/#producer_monitoring
Warning:
This is an unstable interface. It may change in future
releases without warning.
"""
if raw:
return self._metrics.metrics.copy()
metrics = {}
for k, v in six.iteritems(self._metrics.metrics.copy()):
if k.group not in metrics:
metrics[k.group] = {}
if k.name not in metrics[k.group]:
metrics[k.group][k.name] = {}
metrics[k.group][k.name] = v.value()
return metrics
def _is_balanced(self):
"""Determines if the current assignment is a balanced one"""
if (len(self.current_assignment[
self._get_consumer_with_least_subscriptions()]) >=
len(self.current_assignment[
self._get_consumer_with_most_subscriptions()]) - 1):
# if minimum and maximum numbers of partitions assigned to consumers differ by at most one return true
return True
# create a mapping from partitions to the consumer assigned to them
all_assigned_partitions = {}
for consumer_id, consumer_partitions in six.iteritems(
self.current_assignment):
for partition in consumer_partitions:
if partition in all_assigned_partitions:
log.error(
"{} is assigned to more than one consumer.".format(
partition))
all_assigned_partitions[partition] = consumer_id
# for each consumer that does not have all the topic partitions it can get
# make sure none of the topic partitions it could but did not get cannot be moved to it
# (because that would break the balance)
for consumer, _ in self.sorted_current_subscriptions:
consumer_partition_count = len(self.current_assignment[consumer])
# skip if this consumer already has all the topic partitions it can get
if consumer_partition_count == len(
self.consumer_to_all_potential_partitions[consumer]):
continue
# otherwise make sure it cannot get any more
for partition in self.consumer_to_all_potential_partitions[
consumer]:
if partition not in self.current_assignment[consumer]:
other_consumer = all_assigned_partitions[partition]
other_consumer_partition_count = len(
self.current_assignment[other_consumer])
if consumer_partition_count < other_consumer_partition_count:
return False
return True
def _produce_request(self, node_id, acks, timeout, batches):
"""Create a produce request from the given record batches.
Returns:
ProduceRequest (version depends on api_version)
"""
produce_records_by_partition = collections.defaultdict(dict)
for batch in batches:
topic = batch.topic_partition.topic
partition = batch.topic_partition.partition
buf = batch.records.buffer()
produce_records_by_partition[topic][partition] = buf
kwargs = {}
if self.config['api_version'] >= (2, 1):
version = 7
elif self.config['api_version'] >= (2, 0):
version = 6
elif self.config['api_version'] >= (1, 1):
version = 5
elif self.config['api_version'] >= (1, 0):
version = 4
elif self.config['api_version'] >= (0, 11):
version = 3
kwargs = dict(transactional_id=None)
elif self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
return ProduceRequest[version](
required_acks=acks,
timeout=timeout,
topics=[(topic, list(partition_info.items()))
for topic, partition_info in six.iteritems(
produce_records_by_partition)],
**kwargs)
def test_send_broker_unaware_request_fail(self, load_metadata, conn):
mocked_conns = {
('kafka01', 9092): MagicMock(),
('kafka02', 9092): MagicMock()
}
for val in mocked_conns.values():
mock_conn(val, success=False)
def mock_get_conn(host, port, afi):
return mocked_conns[(host, port)]
conn.side_effect = mock_get_conn
client = SimpleClient(hosts=['kafka01:9092', 'kafka02:9092'])
req = KafkaProtocol.encode_metadata_request()
with self.assertRaises(KafkaUnavailableError):
client._send_broker_unaware_request(payloads=['fake request'],
encoder_fn=MagicMock(return_value='fake encoded message'),
decoder_fn=lambda x: x)
for key, conn in six.iteritems(mocked_conns):
conn.send.assert_called_with('fake encoded message')
def test_new_subscription(mocker):
cluster = create_cluster(mocker,
topics={'t1', 't2', 't3', 't4'},
topics_partitions={0})
subscriptions = defaultdict(set)
for i in range(3):
for j in range(i, 3 * i - 2 + 1):
subscriptions['C{}'.format(i)].add('t{}'.format(j))
member_metadata = make_member_metadata(subscriptions)
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
subscriptions['C0'].add('t1')
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(topics, [])
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
assert StickyPartitionAssignor._latest_partition_movements.are_sticky()
def test_sticky_reassignment_after_one_consumer_added(mocker):
cluster = create_cluster(mocker,
topics={'t'},
topics_partitions=set(range(20)))
subscriptions = defaultdict(set)
for i in range(1, 10):
subscriptions['C{}'.format(i)] = {'t'}
member_metadata = make_member_metadata(subscriptions)
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
subscriptions['C10'] = {'t'}
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics,
assignment[member].partitions() if member in assignment else [])
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
assert StickyPartitionAssignor._latest_partition_movements.are_sticky()
def test_sticky_large_assignment_with_multiple_consumers_leaving(mocker):
n_topics = 40
n_consumers = 200
all_topics = set(['t{}'.format(i) for i in range(1, n_topics + 1)])
partitions = dict([(t, set(range(1,
randint(0, 10) + 1)))
for t in all_topics])
cluster = create_cluster(mocker,
topics=all_topics,
topic_partitions_lambda=lambda t: partitions[t])
subscriptions = defaultdict(set)
for i in range(1, n_consumers + 1):
for j in range(0, randint(1, 20)):
subscriptions['C{}'.format(i)].add('t{}'.format(
randint(1, n_topics)))
member_metadata = make_member_metadata(subscriptions)
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, assignment[member].partitions())
for i in range(50):
member = 'C{}'.format(randint(1, n_consumers))
if member in subscriptions:
del subscriptions[member]
del member_metadata[member]
assignment = StickyPartitionAssignor.assign(cluster, member_metadata)
verify_validity_and_balance(subscriptions, assignment)
assert StickyPartitionAssignor._latest_partition_movements.are_sticky()
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 = self._perform_assignment(
response.leader_id, response.group_protocol, response.members)
except Exception as e:
return Future().failure(e)
version = 0 if self.config["api_version"] < (0, 11, 0) else 1
request = SyncGroupRequest[version](
self.group_id,
self._generation.generation_id,
self._generation.member_id,
[(
member_id,
assignment
if isinstance(assignment, bytes) else assignment.encode(),
) for member_id, assignment in six.iteritems(group_assignment)],
)
log.debug(
"Sending leader SyncGroup for group %s to coordinator %s: %s",
self.group_id,
self.coordinator_id,
request,
)
return self._send_sync_group_request(request)
def test_send_broker_unaware_request_fail(self, load_metadata, conn):
mocked_conns = {
('kafka01', 9092): MagicMock(),
('kafka02', 9092): MagicMock()
}
for val in mocked_conns.values():
mock_conn(val, success=False)
def mock_get_conn(host, port, afi):
return mocked_conns[(host, port)]
conn.side_effect = mock_get_conn
client = SimpleClient(hosts=['kafka01:9092', 'kafka02:9092'])
req = KafkaProtocol.encode_metadata_request()
with self.assertRaises(KafkaUnavailableError):
client._send_broker_unaware_request(
payloads=['fake request'],
encoder_fn=MagicMock(return_value='fake encoded message'),
decoder_fn=lambda x: x)
for key, conn in six.iteritems(mocked_conns):
conn.send.assert_called_with('fake encoded message')
def test_sticky_add_remove_topic_two_consumers(mocker):
cluster = create_cluster(mocker,
topics={'t1', 't2'},
topics_partitions={0, 1, 2})
subscriptions = {
'C1': {'t1'},
'C2': {'t1'},
}
member_metadata = make_member_metadata(subscriptions)
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [0, 2])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [1])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
subscriptions = {
'C1': {'t1', 't2'},
'C2': {'t1', 't2'},
}
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, sticky_assignment[member].partitions())
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [0, 2]), ('t2', [1])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [1]), ('t2', [0, 2])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
subscriptions = {
'C1': {'t2'},
'C2': {'t2'},
}
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, sticky_assignment[member].partitions())
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t2', [1])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t2', [0, 2])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
def test_sticky_assignor2(mocker):
"""
Given: there are three consumers C0, C1, C2,
and three topics t0, t1, t2, with 1, 2, and 3 partitions respectively.
Therefore, the partitions are t0p0, t1p0, t1p1, t2p0, t2p1, t2p2.
C0 is subscribed to t0;
C1 is subscribed to t0, t1;
and C2 is subscribed to t0, t1, t2.
Then: perform the assignment
Expected: the assignment is
- C0 [t0p0]
- C1 [t1p0, t1p1]
- C2 [t2p0, t2p1, t2p2]
Then: remove C0 and perform the assignment
Expected: the assignment is
- C1 [t0p0, t1p0, t1p1]
- C2 [t2p0, t2p1, t2p2]
"""
partitions = {'t0': {0}, 't1': {0, 1}, 't2': {0, 1, 2}}
cluster = create_cluster(mocker,
topics={'t0', 't1', 't2'},
topic_partitions_lambda=lambda t: partitions[t])
subscriptions = {
'C0': {'t0'},
'C1': {'t0', 't1'},
'C2': {'t0', 't1', 't2'},
}
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(topics, [])
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C0':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [0])], b''),
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [0, 1])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t2', [0, 1, 2])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
del subscriptions['C0']
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, sticky_assignment[member].partitions())
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [0]), ('t1', [0, 1])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t2', [0, 1, 2])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
def make_member_metadata(subscriptions):
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(topics, [])
return member_metadata
def _send_offset_commit_request(self, offsets):
"""Commit offsets for the specified list of topics and partitions.
This is a non-blocking call which returns a request future that can be
polled in the case of a synchronous commit or ignored in the
asynchronous case.
Arguments:
offsets (dict of {TopicPartition: OffsetAndMetadata}): what should
be committed
Returns:
Future: indicating whether the commit was successful or not
"""
assert self.config['api_version'] >= (0, 8, 1), 'Unsupported Broker API'
assert all(map(lambda k: isinstance(k, TopicPartition), offsets))
assert all(map(lambda v: isinstance(v, OffsetAndMetadata),
offsets.values()))
if not offsets:
log.debug('No offsets to commit')
return Future().success(True)
elif self.coordinator_unknown():
return Future().failure(Errors.GroupCoordinatorNotAvailableError)
node_id = self.coordinator_id
# create the offset commit request
offset_data = collections.defaultdict(dict)
for tp, offset in six.iteritems(offsets):
offset_data[tp.topic][tp.partition] = offset
if self.config['api_version'] >= (0, 9):
request = OffsetCommitRequest[2](
self.group_id,
self.generation,
self.member_id,
OffsetCommitRequest[2].DEFAULT_RETENTION_TIME,
[(
topic, [(
partition,
offset.offset,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
elif self.config['api_version'] >= (0, 8, 2):
request = OffsetCommitRequest[1](
self.group_id, -1, '',
[(
topic, [(
partition,
offset.offset,
-1,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
elif self.config['api_version'] >= (0, 8, 1):
request = OffsetCommitRequest[0](
self.group_id,
[(
topic, [(
partition,
offset.offset,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
log.debug("Sending offset-commit request with %s for group %s to %s",
offsets, self.group_id, node_id)
future = Future()
_f = self._client.send(node_id, request)
_f.add_callback(self._handle_offset_commit_response, offsets, future, time.time())
_f.add_errback(self._failed_request, node_id, request, future)
return future
def _send_broker_aware_request(self, payloads, encoder_fn, decoder_fn):
"""
Group a list of request payloads by topic+partition and send them to
the leader broker for that partition using the supplied encode/decode
functions
Arguments:
payloads: list of object-like entities with a topic (str) and
partition (int) attribute; payloads with duplicate topic-partitions
are not supported.
encode_fn: a method to encode the list of payloads to a request body,
must accept client_id, correlation_id, and payloads as
keyword arguments
decode_fn: a method to decode a response body into response objects.
The response objects must be object-like and have topic
and partition attributes
Returns:
List of response objects in the same order as the supplied payloads
"""
# encoders / decoders do not maintain ordering currently
# so we need to keep this so we can rebuild order before returning
original_ordering = [(p.topic, p.partition) for p in payloads]
# Connection errors generally mean stale metadata
# although sometimes it means incorrect api request
# Unfortunately there is no good way to tell the difference
# so we'll just reset metadata on all errors to be safe
refresh_metadata = False
# For each broker, send the list of request payloads
# and collect the responses and errors
payloads_by_broker = self._payloads_by_broker(payloads)
responses = {}
def failed_payloads(payloads):
for payload in payloads:
topic_partition = (str(payload.topic), payload.partition)
responses[(topic_partition)] = FailedPayloadsError(payload)
# For each BrokerConnection keep the real socket so that we can use
# a select to perform unblocking I/O
connections_by_future = {}
for broker, broker_payloads in six.iteritems(payloads_by_broker):
if broker is None:
failed_payloads(broker_payloads)
continue
host, port, afi = get_ip_port_afi(broker.host)
try:
conn = self._get_conn(host, broker.port, afi)
except ConnectionError:
refresh_metadata = True
failed_payloads(broker_payloads)
continue
request = encoder_fn(payloads=broker_payloads)
future = conn.send(request)
if future.failed():
refresh_metadata = True
failed_payloads(broker_payloads)
continue
if not request.expect_response():
for payload in broker_payloads:
topic_partition = (str(payload.topic), payload.partition)
responses[topic_partition] = None
continue
connections_by_future[future] = (conn, broker)
conn = None
while connections_by_future:
futures = list(connections_by_future.keys())
# block until a socket is ready to be read
sockets = [
conn._sock
for future, (conn, _) in six.iteritems(connections_by_future)
if not future.is_done and conn._sock is not None
]
if sockets:
read_socks, _, _ = select.select(sockets, [], [])
for future in futures:
if not future.is_done:
conn, _ = connections_by_future[future]
conn.recv()
continue
_, broker = connections_by_future.pop(future)
if future.failed():
refresh_metadata = True
failed_payloads(payloads_by_broker[broker])
else:
for payload_response in decoder_fn(future.value):
topic_partition = (str(payload_response.topic),
payload_response.partition)
responses[topic_partition] = payload_response
if refresh_metadata:
self.reset_all_metadata()
# Return responses in the same order as provided
return [responses[tp] for tp in original_ordering]
def _fetch(self):
# Create fetch request payloads for all the partitions
partitions = dict((p, self.buffer_size)
for p in self.fetch_offsets.keys())
while partitions:
requests = []
for partition, buffer_size in six.iteritems(partitions):
requests.append(FetchRequestPayload(self.topic, partition,
self.fetch_offsets[partition],
buffer_size))
# Send request
responses = self.client.send_fetch_request(
requests,
max_wait_time=int(self.fetch_max_wait_time),
min_bytes=self.fetch_min_bytes,
fail_on_error=False
)
retry_partitions = {}
for resp in responses:
try:
check_error(resp)
except UnknownTopicOrPartitionError:
log.error('UnknownTopicOrPartitionError for %s:%d',
resp.topic, resp.partition)
self.client.reset_topic_metadata(resp.topic)
raise
except NotLeaderForPartitionError:
log.error('NotLeaderForPartitionError for %s:%d',
resp.topic, resp.partition)
self.client.reset_topic_metadata(resp.topic)
continue
except OffsetOutOfRangeError:
log.warning('OffsetOutOfRangeError for %s:%d. '
'Resetting partition offset...',
resp.topic, resp.partition)
self.reset_partition_offset(resp.partition)
# Retry this partition
retry_partitions[resp.partition] = partitions[resp.partition]
continue
except FailedPayloadsError as e:
log.warning('FailedPayloadsError for %s:%d',
e.payload.topic, e.payload.partition)
# Retry this partition
retry_partitions[e.payload.partition] = partitions[e.payload.partition]
continue
partition = resp.partition
buffer_size = partitions[partition]
# Check for partial message
if resp.messages and isinstance(resp.messages[-1].message, PartialMessage):
# If buffer is at max and all we got was a partial message
# raise ConsumerFetchSizeTooSmall
if (self.max_buffer_size is not None and
buffer_size == self.max_buffer_size and
len(resp.messages) == 1):
log.error('Max fetch size %d too small', self.max_buffer_size)
raise ConsumerFetchSizeTooSmall()
if self.max_buffer_size is None:
buffer_size *= 2
else:
buffer_size = min(buffer_size * 2, self.max_buffer_size)
log.warning('Fetch size too small, increase to %d (2x) '
'and retry', buffer_size)
retry_partitions[partition] = buffer_size
resp.messages.pop()
for message in resp.messages:
if message.offset < self.fetch_offsets[partition]:
log.debug('Skipping message %s because its offset is less than the consumer offset',
message)
continue
# Put the message in our queue
self.queue.put((partition, message))
self.fetch_offsets[partition] = message.offset + 1
partitions = retry_partitions
def list_consumer_group_offsets(self, group_id, group_coordinator_id=None,
partitions=None):
"""Fetch Consumer Group Offsets.
Note:
This does not verify that the group_id or partitions actually exist
in the cluster.
As soon as any error is encountered, it is immediately raised.
:param group_id: The consumer group id name for which to fetch offsets.
:param group_coordinator_id: The node_id of the group's coordinator
broker. If set to None, will query the cluster to find the group
coordinator. Explicitly specifying this can be useful to prevent
that extra network round trip if you already know the group
coordinator. Default: None.
:param partitions: A list of TopicPartitions for which to fetch
offsets. On brokers >= 0.10.2, this can be set to None to fetch all
known offsets for the consumer group. Default: None.
:return dictionary: A dictionary with TopicPartition keys and
OffsetAndMetada values. Partitions that are not specified and for
which the group_id does not have a recorded offset are omitted. An
offset value of `-1` indicates the group_id has no offset for that
TopicPartition. A `-1` can only happen for partitions that are
explicitly specified.
"""
group_offsets_listing = {}
if group_coordinator_id is None:
group_coordinator_id = self._find_group_coordinator_id(group_id)
version = self._matching_api_version(OffsetFetchRequest)
if version <= 3:
if partitions is None:
if version <= 1:
raise ValueError(
"""OffsetFetchRequest_v{} requires specifying the
partitions for which to fetch offsets. Omitting the
partitions is only supported on brokers >= 0.10.2.
For details, see KIP-88.""".format(version))
topics_partitions = None
else:
# transform from [TopicPartition("t1", 1), TopicPartition("t1", 2)] to [("t1", [1, 2])]
topics_partitions_dict = defaultdict(set)
for topic, partition in partitions:
topics_partitions_dict[topic].add(partition)
topics_partitions = list(six.iteritems(topics_partitions_dict))
request = OffsetFetchRequest[version](group_id, topics_partitions)
response = self._send_request_to_node(group_coordinator_id, request)
if version > 1: # OffsetFetchResponse_v1 lacks a top-level error_code
error_type = Errors.for_code(response.error_code)
if error_type is not Errors.NoError:
# optionally we could retry if error_type.retriable
raise error_type(
"Request '{}' failed with response '{}'."
.format(request, response))
# transform response into a dictionary with TopicPartition keys and
# OffsetAndMetada values--this is what the Java AdminClient returns
for topic, partitions in response.topics:
for partition, offset, metadata, error_code in partitions:
error_type = Errors.for_code(error_code)
if error_type is not Errors.NoError:
raise error_type(
"Unable to fetch offsets for group_id {}, topic {}, partition {}"
.format(group_id, topic, partition))
group_offsets_listing[TopicPartition(topic, partition)] = OffsetAndMetadata(offset, metadata)
else:
raise NotImplementedError(
"Support for OffsetFetch v{} has not yet been added to KafkaAdminClient."
.format(version))
return group_offsets_listing
def run_once(self):
"""Run a single iteration of sending."""
while self._topics_to_add:
self._client.add_topic(self._topics_to_add.pop())
# get the list of partitions with data ready to send
result = self._accumulator.ready(self._metadata)
ready_nodes, next_ready_check_delay, unknown_leaders_exist = result
# if there are any partitions whose leaders are not known yet, force
# metadata update
if unknown_leaders_exist:
log.debug('Unknown leaders exist, requesting metadata update')
self._metadata.request_update()
# remove any nodes we aren't ready to send to
not_ready_timeout = 999999999
for node in list(ready_nodes):
if not self._client.ready(node):
log.debug('Node %s not ready; delaying produce of accumulated batch', node)
ready_nodes.remove(node)
not_ready_timeout = min(not_ready_timeout,
self._client.connection_delay(node))
# create produce requests
batches_by_node = self._accumulator.drain(
self._metadata, ready_nodes, self.config['max_request_size'])
if self.config['guarantee_message_order']:
# Mute all the partitions drained
for batch_list in six.itervalues(batches_by_node):
for batch in batch_list:
self._accumulator.muted.add(batch.topic_partition)
expired_batches = self._accumulator.abort_expired_batches(
self.config['request_timeout_ms'], self._metadata)
for expired_batch in expired_batches:
self._sensors.record_errors(expired_batch.topic_partition.topic, expired_batch.record_count)
self._sensors.update_produce_request_metrics(batches_by_node)
requests = self._create_produce_requests(batches_by_node)
# If we have any nodes that are ready to send + have sendable data,
# poll with 0 timeout so this can immediately loop and try sending more
# data. Otherwise, the timeout is determined by nodes that have
# partitions with data that isn't yet sendable (e.g. lingering, backing
# off). Note that this specifically does not include nodes with
# sendable data that aren't ready to send since they would cause busy
# looping.
poll_timeout_ms = min(next_ready_check_delay * 1000, not_ready_timeout)
if ready_nodes:
log.debug("Nodes with data ready to send: %s", ready_nodes) # trace
log.debug("Created %d produce requests: %s", len(requests), requests) # trace
poll_timeout_ms = 0
for node_id, request in six.iteritems(requests):
batches = batches_by_node[node_id]
log.debug('Sending Produce Request: %r', request)
(self._client.send(node_id, request)
.add_callback(
self._handle_produce_response, node_id, time.time(), batches)
.add_errback(
self._failed_produce, batches, node_id))
# if some partitions are already ready to be sent, the select time
# would be 0; otherwise if some partition already has some data
# accumulated but not ready yet, the select time will be the time
# difference between now and its linger expiry time; otherwise the
# select time will be the time difference between now and the
# metadata expiry time
self._client.poll(poll_timeout_ms, sleep=True)
def _send_broker_aware_request(self, payloads, encoder_fn, decoder_fn):
"""
Group a list of request payloads by topic+partition and send them to
the leader broker for that partition using the supplied encode/decode
functions
Arguments:
payloads: list of object-like entities with a topic (str) and
partition (int) attribute; payloads with duplicate topic-partitions
are not supported.
encode_fn: a method to encode the list of payloads to a request body,
must accept client_id, correlation_id, and payloads as
keyword arguments
decode_fn: a method to decode a response body into response objects.
The response objects must be object-like and have topic
and partition attributes
Returns:
List of response objects in the same order as the supplied payloads
"""
# encoders / decoders do not maintain ordering currently
# so we need to keep this so we can rebuild order before returning
original_ordering = [(p.topic, p.partition) for p in payloads]
# Connection errors generally mean stale metadata
# although sometimes it means incorrect api request
# Unfortunately there is no good way to tell the difference
# so we'll just reset metadata on all errors to be safe
refresh_metadata = False
# For each broker, send the list of request payloads
# and collect the responses and errors
payloads_by_broker = self._payloads_by_broker(payloads)
responses = {}
def failed_payloads(payloads):
for payload in payloads:
topic_partition = (str(payload.topic), payload.partition)
responses[(topic_partition)] = FailedPayloadsError(payload)
# For each BrokerConnection keep the real socket so that we can use
# a select to perform unblocking I/O
connections_by_future = {}
for broker, broker_payloads in six.iteritems(payloads_by_broker):
if broker is None:
failed_payloads(broker_payloads)
continue
host, port, afi = get_ip_port_afi(broker.host)
try:
conn = self._get_conn(host, broker.port, afi)
except ConnectionError:
refresh_metadata = True
failed_payloads(broker_payloads)
continue
request = encoder_fn(payloads=broker_payloads)
future = conn.send(request)
if future.failed():
refresh_metadata = True
failed_payloads(broker_payloads)
continue
if not request.expect_response():
for payload in broker_payloads:
topic_partition = (str(payload.topic), payload.partition)
responses[topic_partition] = None
continue
connections_by_future[future] = (conn, broker)
conn = None
while connections_by_future:
futures = list(connections_by_future.keys())
# block until a socket is ready to be read
sockets = [
conn._sock
for future, (conn, _) in six.iteritems(connections_by_future)
if not future.is_done and conn._sock is not None]
if sockets:
read_socks, _, _ = select.select(sockets, [], [])
for future in futures:
if not future.is_done:
conn, _ = connections_by_future[future]
conn.recv()
continue
_, broker = connections_by_future.pop(future)
if future.failed():
refresh_metadata = True
failed_payloads(payloads_by_broker[broker])
else:
for payload_response in decoder_fn(future.value):
topic_partition = (str(payload_response.topic),
payload_response.partition)
responses[topic_partition] = payload_response
if refresh_metadata:
self.reset_all_metadata()
# Return responses in the same order as provided
return [responses[tp] for tp in original_ordering]
def test_sticky_assignor1(mocker):
"""
Given: there are three consumers C0, C1, C2,
four topics t0, t1, t2, t3, and each topic has 2 partitions,
resulting in partitions t0p0, t0p1, t1p0, t1p1, t2p0, t2p1, t3p0, t3p1.
Each consumer is subscribed to all three topics.
Then: perform fresh assignment
Expected: the assignment is
- C0: [t0p0, t1p1, t3p0]
- C1: [t0p1, t2p0, t3p1]
- C2: [t1p0, t2p1]
Then: remove C1 consumer and perform the reassignment
Expected: the new assignment is
- C0 [t0p0, t1p1, t2p0, t3p0]
- C2 [t0p1, t1p0, t2p1, t3p1]
"""
cluster = create_cluster(mocker,
topics={'t0', 't1', 't2', 't3'},
topics_partitions={0, 1})
subscriptions = {
'C0': {'t0', 't1', 't2', 't3'},
'C1': {'t0', 't1', 't2', 't3'},
'C2': {'t0', 't1', 't2', 't3'},
}
member_metadata = make_member_metadata(subscriptions)
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C0':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [0]), ('t1', [1]),
('t3', [0])], b''),
'C1':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [1]), ('t2', [0]),
('t3', [1])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t1', [0]), ('t2', [1])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
del subscriptions['C1']
member_metadata = {}
for member, topics in six.iteritems(subscriptions):
member_metadata[member] = StickyPartitionAssignor._metadata(
topics, sticky_assignment[member].partitions())
sticky_assignment = StickyPartitionAssignor.assign(cluster,
member_metadata)
expected_assignment = {
'C0':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [0]), ('t1', [1]),
('t2', [0]), ('t3', [0])], b''),
'C2':
ConsumerProtocolMemberAssignment(StickyPartitionAssignor.version,
[('t0', [1]), ('t1', [0]),
('t2', [1]), ('t3', [1])], b''),
}
assert_assignment(sticky_assignment, expected_assignment)
def _create_fetch_requests(self):
"""Create fetch requests for all assigned partitions, grouped by node.
FetchRequests skipped if no leader, or node has requests in flight
Returns:
dict: {node_id: FetchRequest, ...} (version depends on api_version)
"""
# create the fetch info as a dict of lists of partition info tuples
# which can be passed to FetchRequest() via .items()
fetchable = collections.defaultdict(
lambda: collections.defaultdict(list))
for partition in self._fetchable_partitions():
node_id = self._client.cluster.leader_for_partition(partition)
# advance position for any deleted compacted messages if required
if self._subscriptions.assignment[
partition].last_offset_from_message_batch:
next_offset_from_batch_header = self._subscriptions.assignment[
partition].last_offset_from_message_batch + 1
if next_offset_from_batch_header > self._subscriptions.assignment[
partition].position:
log.debug(
"Advance position for partition %s from %s to %s (last message batch location plus one)"
" to correct for deleted compacted messages",
partition,
self._subscriptions.assignment[partition].position,
next_offset_from_batch_header)
self._subscriptions.assignment[
partition].position = next_offset_from_batch_header
position = self._subscriptions.assignment[partition].position
# fetch if there is a leader and no in-flight requests
if node_id is None or node_id == -1:
log.debug(
"No leader found for partition %s."
" Requesting metadata update", partition)
self._client.cluster.request_update()
elif self._client.in_flight_request_count(node_id) == 0:
partition_info = (partition.partition, position,
self.config['max_partition_fetch_bytes'])
fetchable[node_id][partition.topic].append(partition_info)
log.debug("Adding fetch request for partition %s at offset %d",
partition, position)
else:
log.log(
0,
"Skipping fetch for partition %s because there is an inflight request to node %s",
partition, node_id)
if self.config['api_version'] >= (0, 11, 0):
version = 4
elif self.config['api_version'] >= (0, 10, 1):
version = 3
elif self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
requests = {}
for node_id, partition_data in six.iteritems(fetchable):
if version < 3:
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
partition_data.items())
else:
# As of version == 3 partitions will be returned in order as
# they are requested, so to avoid starvation with
# `fetch_max_bytes` option we need this shuffle
# NOTE: we do have partition_data in random order due to usage
# of unordered structures like dicts, but that does not
# guarantee equal distribution, and starting in Python3.6
# dicts retain insert order.
partition_data = list(partition_data.items())
random.shuffle(partition_data)
if version == 3:
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
self.config['fetch_max_bytes'],
partition_data)
else:
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
self.config['fetch_max_bytes'],
self._isolation_level,
partition_data)
return requests
def _create_fetch_requests(self):
"""Create fetch requests for all assigned partitions, grouped by node.
FetchRequests skipped if no leader, or node has requests in flight
Returns:
dict: {node_id: FetchRequest, ...} (version depends on api_version)
"""
# create the fetch info as a dict of lists of partition info tuples
# which can be passed to FetchRequest() via .items()
fetchable = collections.defaultdict(lambda: collections.defaultdict(list))
for partition in self._fetchable_partitions():
node_id = self._client.cluster.leader_for_partition(partition)
position = self._subscriptions.assignment[partition].position
# fetch if there is a leader and no in-flight requests
if node_id is None or node_id == -1:
log.debug("No leader found for partition %s."
" Requesting metadata update", partition)
self._client.cluster.request_update()
elif self._client.in_flight_request_count(node_id) == 0:
partition_info = (
partition.partition,
position,
self.config['max_partition_fetch_bytes']
)
fetchable[node_id][partition.topic].append(partition_info)
log.debug("Adding fetch request for partition %s at offset %d",
partition, position)
if self.config['api_version'] >= (0, 10, 1):
version = 3
elif self.config['api_version'] >= (0, 10):
version = 2
elif self.config['api_version'] == (0, 9):
version = 1
else:
version = 0
requests = {}
for node_id, partition_data in six.iteritems(fetchable):
if version < 3:
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
partition_data.items())
else:
# As of version == 3 partitions will be returned in order as
# they are requested, so to avoid starvation with
# `fetch_max_bytes` option we need this shuffle
# NOTE: we do have partition_data in random order due to usage
# of unordered structures like dicts, but that does not
# guaranty equal distribution, and starting Python3.6
# dicts retain insert order.
partition_data = list(partition_data.items())
random.shuffle(partition_data)
requests[node_id] = FetchRequest[version](
-1, # replica_id
self.config['fetch_max_wait_ms'],
self.config['fetch_min_bytes'],
self.config['fetch_max_bytes'],
partition_data)
return requests
def _send_offset_commit_request(self, offsets):
"""Commit offsets for the specified list of topics and partitions.
This is a non-blocking call which returns a request future that can be
polled in the case of a synchronous commit or ignored in the
asynchronous case.
Arguments:
offsets (dict of {TopicPartition: OffsetAndMetadata}): what should
be committed
Returns:
Future: indicating whether the commit was successful or not
"""
assert self.config['api_version'] >= (0, 8, 1), 'Unsupported Broker API'
assert all(map(lambda k: isinstance(k, TopicPartition), offsets))
assert all(map(lambda v: isinstance(v, OffsetAndMetadata),
offsets.values()))
if not offsets:
log.debug('No offsets to commit')
return Future().success(True)
elif self.coordinator_unknown():
return Future().failure(Errors.GroupCoordinatorNotAvailableError)
node_id = self.coordinator_id
# create the offset commit request
offset_data = collections.defaultdict(dict)
for tp, offset in six.iteritems(offsets):
offset_data[tp.topic][tp.partition] = offset
if self.config['api_version'] >= (0, 9):
request = OffsetCommitRequest[2](
self.group_id,
self.generation,
self.member_id,
OffsetCommitRequest[2].DEFAULT_RETENTION_TIME,
[(
topic, [(
partition,
offset.offset,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
elif self.config['api_version'] >= (0, 8, 2):
request = OffsetCommitRequest[1](
self.group_id, -1, '',
[(
topic, [(
partition,
offset.offset,
-1,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
elif self.config['api_version'] >= (0, 8, 1):
request = OffsetCommitRequest[0](
self.group_id,
[(
topic, [(
partition,
offset.offset,
offset.metadata
) for partition, offset in six.iteritems(partitions)]
) for topic, partitions in six.iteritems(offset_data)]
)
log.debug("Sending offset-commit request with %s for group %s to %s",
offsets, self.group_id, node_id)
future = Future()
_f = self._client.send(node_id, request)
_f.add_callback(self._handle_offset_commit_response, offsets, future, time.time())
_f.add_errback(self._failed_request, node_id, request, future)
return future
def missing_fetch_positions(self):
missing = set()
for partition, state in six.iteritems(self.assignment):
if not state.has_valid_position:
missing.add(partition)
return missing
def _send_offset_commit_request(self, offsets):
"""Commit offsets for the specified list of topics and partitions.
This is a non-blocking call which returns a request future that can be
polled in the case of a synchronous commit or ignored in the
asynchronous case.
Arguments:
offsets (dict of {TopicPartition: OffsetAndMetadata}): what should
be committed
Returns:
Future: indicating whether the commit was successful or not
"""
assert self.config["api_version"] >= (0, 8,
1), "Unsupported Broker API"
assert all(map(lambda k: isinstance(k, TopicPartition), offsets))
assert all(
map(lambda v: isinstance(v, OffsetAndMetadata), offsets.values()))
if not offsets:
log.debug("No offsets to commit")
return Future().success(None)
node_id = self.coordinator()
if node_id is None:
return Future().failure(Errors.GroupCoordinatorNotAvailableError)
# create the offset commit request
offset_data = collections.defaultdict(dict)
for tp, offset in six.iteritems(offsets):
offset_data[tp.topic][tp.partition] = offset
if self._subscription.partitions_auto_assigned():
generation = self.generation()
else:
generation = Generation.NO_GENERATION
# if the generation is None, we are not part of an active group
# (and we expect to be). The only thing we can do is fail the commit
# and let the user rejoin the group in poll()
if self.config["api_version"] >= (0, 9) and generation is None:
return Future().failure(Errors.CommitFailedError())
if self.config["api_version"] >= (0, 9):
request = OffsetCommitRequest[2](
self.group_id,
generation.generation_id,
generation.member_id,
OffsetCommitRequest[2].DEFAULT_RETENTION_TIME,
[(
topic,
[(partition, offset.offset, offset.metadata)
for partition, offset in six.iteritems(partitions)],
) for topic, partitions in six.iteritems(offset_data)],
)
elif self.config["api_version"] >= (0, 8, 2):
request = OffsetCommitRequest[1](
self.group_id,
-1,
"",
[(
topic,
[(partition, offset.offset, -1, offset.metadata)
for partition, offset in six.iteritems(partitions)],
) for topic, partitions in six.iteritems(offset_data)],
)
elif self.config["api_version"] >= (0, 8, 1):
request = OffsetCommitRequest[0](
self.group_id,
[(
topic,
[(partition, offset.offset, offset.metadata)
for partition, offset in six.iteritems(partitions)],
) for topic, partitions in six.iteritems(offset_data)],
)
log.debug(
"Sending offset-commit request with %s for group %s to %s",
offsets,
self.group_id,
node_id,
)
future = Future()
_f = self._client.send(node_id, request)
_f.add_callback(self._handle_offset_commit_response, offsets, future,
time.time())
_f.add_errback(self._failed_request, node_id, request, future)
return future
