def _partition(self):
"""Consume messages from kafka
Consume messages from kafka using the Kazoo SetPartitioner to
allow multiple consumer processes to negotiate access to the kafka
partitions
"""
# KazooClient and SetPartitioner objects need to be instantiated after
# the consumer process has forked. Instantiating prior to forking
# gives the appearance that things are working but after forking the
# connection to zookeeper is lost and no state changes are visible
if not self._kazoo_client:
self._kazoo_client = KazooClient(hosts=self._zookeeper_url)
self._kazoo_client.start()
state_change_event = threading.Event()
self._set_partitioner = (SetPartitioner(
self._kazoo_client,
path=self._zookeeper_path,
set=self._consumer.fetch_offsets.keys(),
state_change_event=state_change_event,
identifier=str(datetime.datetime.now())))
try:
while 1:
if self._set_partitioner.failed:
raise Exception("Failed to acquire partition")
elif self._set_partitioner.release:
log.info("Releasing locks on partition set {} "
"for topic {}".format(self._partitions,
self._kafka_topic))
self._set_partitioner.release_set()
self._partitions = []
elif self._set_partitioner.acquired:
if not self._partitions:
self._partitions = [p for p in self._set_partitioner]
if not self._partitions:
log.info("Not assigned any partitions on topic {},"
" waiting for a Partitioner state change".
format(self._kafka_topic))
state_change_event.wait()
state_change_event.clear()
continue
log.info("Acquired locks on partition set {} "
"for topic {}".format(self._partitions,
self._kafka_topic))
# Reconstruct the kafka consumer object because the
# consumer has no API that allows the set of partitons
# to be updated outside of construction.
self._consumer.stop()
self._consumer = self._create_kafka_consumer(
self._partitions)
return
elif self._set_partitioner.allocating:
log.info("Waiting to acquire locks on partition set")
self._set_partitioner.wait_for_acquire()
except Exception:
log.exception('KafkaConsumer encountered fatal exception '
'processing messages.')
raise
def __iter__(self):
"""Consume messages from kafka using the Kazoo SetPartitioner to
allow multiple consumer processes to negotiate access to the kafka
partitions
"""
# KazooClient and SetPartitioner objects need to be instantiated after
# the consumer process has forked. Instantiating prior to forking
# gives the appearance that things are working but after forking the
# connection to zookeeper is lost and no state changes are visible
kazoo_client = KazooClient(hosts=self._zookeeper_url)
kazoo_client.start()
set_partitioner = (SetPartitioner(
kazoo_client,
path=self._zookeeper_path,
set=self._consumer.fetch_offsets.keys()))
consumed_from_kafka = self._statsd.get_counter(
name='consumed_from_kafka')
try:
partitions = []
while 1:
if set_partitioner.failed:
raise Exception("Failed to acquire partition")
elif set_partitioner.release:
log.info("Releasing locks on partition set {} "
"for topic {}".format(partitions,
self._kafka_topic))
set_partitioner.release_set()
partitions = []
elif set_partitioner.acquired:
if not partitions:
partitions = [p for p in set_partitioner]
log.info("Acquired locks on partition set {} "
"for topic {}".format(partitions,
self._kafka_topic))
# Refresh the last known offsets again to make sure
# that they are the latest after having acquired the
# lock. Updates self._consumer.fetch_offsets.
self._consumer.fetch_last_known_offsets()
# Modify self._consumer.fetch_offsets to hold only the
# offsets for the set of Kafka partitions acquired
# by this instance of the persister.
partitioned_fetch_offsets = {}
for p in partitions:
partitioned_fetch_offsets[p] = (
self._consumer.fetch_offsets[p])
self._consumer.fetch_offsets = partitioned_fetch_offsets
# When Kafka resizes the partitions it's possible that it
# will remove data at our current offset. When this
# happens the next attempt to read from Kafka will generate
# an OffsetOutOfRangeError. We trap this error and seek to
# the head of the current Kafka data. Because this error
# only happens when Kafka removes data we're currently
# pointing at we're gauranteed that we won't read any
# duplicate data however we will lose any information
# between our current offset and the new Kafka head.
try:
for message in self._consumer:
if not set_partitioner.acquired:
break
consumed_from_kafka += 1
log.debug("Consuming message from kafka, "
"partition {}, offset {}".format(
message[0], message[1].offset))
yield message
except kafka.common.OffsetOutOfRangeError:
log.error("Kafka OffsetOutOfRange. Jumping to head.")
self._consumer.seek(0, 0)
elif set_partitioner.allocating:
log.info("Waiting to acquire locks on partition set")
set_partitioner.wait_for_acquire()
except:
log.exception('KafkaConsumer encountered fatal exception '
'processing messages.')
raise
