diff options
Diffstat (limited to 'kafka')
| -rw-r--r-- | kafka/client.py | 31 | ||||
| -rw-r--r-- | kafka/conn.py | 76 | ||||
| -rw-r--r-- | kafka/consumer.py | 269 | ||||
| -rw-r--r-- | kafka/producer.py | 6 | ||||
| -rw-r--r-- | kafka/protocol.py | 24 |
5 files changed, 216 insertions, 190 deletions
diff --git a/kafka/client.py b/kafka/client.py index 71ededa..33c6d77 100644 --- a/kafka/client.py +++ b/kafka/client.py @@ -3,11 +3,12 @@ from collections import defaultdict from functools import partial from itertools import count import logging -import socket import time -from kafka.common import ErrorMapping, TopicAndPartition -from kafka.common import ConnectionError, FailedPayloadsException +from kafka.common import ( + ErrorMapping, TopicAndPartition, ConnectionError, + FailedPayloadsException +) from kafka.conn import KafkaConnection from kafka.protocol import KafkaProtocol @@ -19,12 +20,12 @@ class KafkaClient(object): CLIENT_ID = "kafka-python" ID_GEN = count() - def __init__(self, host, port, bufsize=4096, client_id=CLIENT_ID): + def __init__(self, host, port, client_id=CLIENT_ID, timeout=10): # We need one connection to bootstrap - self.bufsize = bufsize self.client_id = client_id + self.timeout = timeout self.conns = { # (host, port) -> KafkaConnection - (host, port): KafkaConnection(host, port, bufsize) + (host, port): KafkaConnection(host, port, timeout=timeout) } self.brokers = {} # broker_id -> BrokerMetadata self.topics_to_brokers = {} # topic_id -> broker_id @@ -41,7 +42,7 @@ class KafkaClient(object): """ if (broker.host, broker.port) not in self.conns: self.conns[(broker.host, broker.port)] = \ - KafkaConnection(broker.host, broker.port, self.bufsize) + KafkaConnection(broker.host, broker.port, timeout=self.timeout) return self.conns[(broker.host, broker.port)] @@ -165,14 +166,24 @@ class KafkaClient(object): request = encoder_fn(client_id=self.client_id, correlation_id=requestId, payloads=payloads) + failed = False # Send the request, recv the response try: conn.send(requestId, request) if decoder_fn is None: continue - response = conn.recv(requestId) - except ConnectionError, e: # ignore BufferUnderflow for now - log.warning("Could not send request [%s] to server %s: %s" % (request, conn, e)) + try: + response = conn.recv(requestId) + except ConnectionError, e: + log.warning("Could not receive response to request [%s] " + "from server %s: %s", request, conn, e) + failed = True + except ConnectionError, e: + log.warning("Could not send request [%s] to server %s: %s", + request, conn, e) + failed = True + + if failed: failed_payloads += payloads self.topics_to_brokers = {} # reset metadata continue diff --git a/kafka/conn.py b/kafka/conn.py index 1a3e260..c80f428 100644 --- a/kafka/conn.py +++ b/kafka/conn.py @@ -4,7 +4,6 @@ import socket import struct from threading import local -from kafka.common import BufferUnderflowError from kafka.common import ConnectionError log = logging.getLogger("kafka") @@ -19,14 +18,14 @@ class KafkaConnection(local): we can do something in here to facilitate multiplexed requests/responses since the Kafka API includes a correlation id. """ - def __init__(self, host, port, bufsize=4096): + def __init__(self, host, port, timeout=10): super(KafkaConnection, self).__init__() self.host = host self.port = port - self.bufsize = bufsize self._sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) self._sock.connect((host, port)) - self._sock.settimeout(10) + self.timeout = timeout + self._sock.settimeout(self.timeout) self._dirty = False def __str__(self): @@ -36,44 +35,31 @@ class KafkaConnection(local): # Private API # ################### - def _consume_response(self): - """ - Fully consume the response iterator - """ - return "".join(self._consume_response_iter()) - - def _consume_response_iter(self): - """ - This method handles the response header and error messages. It - then returns an iterator for the chunks of the response - """ - log.debug("Handling response from Kafka") - - # Read the size off of the header - resp = self._sock.recv(4) - if resp == "": - self._raise_connection_error() - (size,) = struct.unpack('>i', resp) - - messagesize = size - 4 - log.debug("About to read %d bytes from Kafka", messagesize) - - # Read the remainder of the response - total = 0 - while total < messagesize: - resp = self._sock.recv(self.bufsize) - log.debug("Read %d bytes from Kafka", len(resp)) - if resp == "": - raise BufferUnderflowError( - "Not enough data to read this response") - - total += len(resp) - yield resp - def _raise_connection_error(self): self._dirty = True raise ConnectionError("Kafka @ {}:{} went away".format(self.host, self.port)) + def _read_bytes(self, num_bytes): + bytes_left = num_bytes + resp = '' + log.debug("About to read %d bytes from Kafka", num_bytes) + if self._dirty: + self.reinit() + while bytes_left: + try: + data = self._sock.recv(bytes_left) + except socket.error: + log.exception('Unable to receive data from Kafka') + self._raise_connection_error() + if data == '': + log.error("Not enough data to read this response") + self._raise_connection_error() + bytes_left -= len(data) + log.debug("Read %d/%d bytes from Kafka", num_bytes - bytes_left, num_bytes) + resp += data + + return resp + ################## # Public API # ################## @@ -89,7 +75,7 @@ class KafkaConnection(local): sent = self._sock.sendall(payload) if sent is not None: self._raise_connection_error() - except socket.error: + except socket.error, e: log.exception('Unable to send payload to Kafka') self._raise_connection_error() @@ -98,8 +84,14 @@ class KafkaConnection(local): Get a response from Kafka """ log.debug("Reading response %d from Kafka" % request_id) - self.data = self._consume_response() - return self.data + # Read the size off of the header + resp = self._read_bytes(4) + + (size,) = struct.unpack('>i', resp) + + # Read the remainder of the response + resp = self._read_bytes(size) + return str(resp) def copy(self): """ @@ -124,5 +116,5 @@ class KafkaConnection(local): self.close() self._sock = socket.socket(socket.AF_INET, socket.SOCK_STREAM) self._sock.connect((self.host, self.port)) - self._sock.settimeout(10) + self._sock.settimeout(self.timeout) self._dirty = False diff --git a/kafka/consumer.py b/kafka/consumer.py index 226700e..eba2912 100644 --- a/kafka/consumer.py +++ b/kafka/consumer.py @@ -5,8 +5,8 @@ from itertools import izip_longest, repeat import logging import time from threading import Lock -from multiprocessing import Process, Queue, Event, Value -from Queue import Empty +from multiprocessing import Process, Queue as MPQueue, Event, Value +from Queue import Empty, Queue from kafka.common import ( ErrorMapping, FetchRequest, @@ -24,6 +24,11 @@ AUTO_COMMIT_INTERVAL = 5000 FETCH_DEFAULT_BLOCK_TIMEOUT = 1 FETCH_MAX_WAIT_TIME = 100 FETCH_MIN_BYTES = 4096 +FETCH_BUFFER_SIZE_BYTES = 4096 +MAX_FETCH_BUFFER_SIZE_BYTES = FETCH_BUFFER_SIZE_BYTES * 8 + +ITER_TIMEOUT_SECONDS = 60 +NO_MESSAGES_WAIT_TIME_SECONDS = 0.1 class FetchContext(object): @@ -34,13 +39,15 @@ class FetchContext(object): self.consumer = consumer self.block = block - if block and not timeout: - timeout = FETCH_DEFAULT_BLOCK_TIMEOUT - - self.timeout = timeout * 1000 + if block: + if not timeout: + timeout = FETCH_DEFAULT_BLOCK_TIMEOUT + self.timeout = timeout * 1000 def __enter__(self): """Set fetch values based on blocking status""" + self.orig_fetch_max_wait_time = self.consumer.fetch_max_wait_time + self.orig_fetch_min_bytes = self.consumer.fetch_min_bytes if self.block: self.consumer.fetch_max_wait_time = self.timeout self.consumer.fetch_min_bytes = 1 @@ -48,9 +55,9 @@ class FetchContext(object): self.consumer.fetch_min_bytes = 0 def __exit__(self, type, value, traceback): - """Reset values to default""" - self.consumer.fetch_max_wait_time = FETCH_MAX_WAIT_TIME - self.consumer.fetch_min_bytes = FETCH_MIN_BYTES + """Reset values""" + self.consumer.fetch_max_wait_time = self.orig_fetch_max_wait_time + self.consumer.fetch_min_bytes = self.orig_fetch_min_bytes class Consumer(object): @@ -206,8 +213,14 @@ class SimpleConsumer(Consumer): before a commit auto_commit_every_t: default 5000. How much time (in milliseconds) to wait before commit - fetch_size_bytes: number of bytes to request in a FetchRequest + buffer_size: default 4K. Initial number of bytes to tell kafka we + have available. This will double as needed. + max_buffer_size: default 16K. Max number of bytes to tell kafka we have + available. None means no limit. + iter_timeout: default None. How much time (in seconds) to wait for a + message in the iterator before exiting. None means no + timeout, so it will wait forever. Auto commit details: If both auto_commit_every_n and auto_commit_every_t are set, they will @@ -218,12 +231,23 @@ class SimpleConsumer(Consumer): def __init__(self, client, group, topic, auto_commit=True, partitions=None, auto_commit_every_n=AUTO_COMMIT_MSG_COUNT, auto_commit_every_t=AUTO_COMMIT_INTERVAL, - fetch_size_bytes=FETCH_MIN_BYTES): - + fetch_size_bytes=FETCH_MIN_BYTES, + buffer_size=FETCH_BUFFER_SIZE_BYTES, + max_buffer_size=MAX_FETCH_BUFFER_SIZE_BYTES, + iter_timeout=None): + + if max_buffer_size is not None and buffer_size > max_buffer_size: + raise ValueError("buffer_size (%d) is greater than " + "max_buffer_size (%d)" % + (buffer_size, max_buffer_size)) + self.buffer_size = buffer_size + self.max_buffer_size = max_buffer_size self.partition_info = False # Do not return partition info in msgs self.fetch_max_wait_time = FETCH_MAX_WAIT_TIME self.fetch_min_bytes = fetch_size_bytes self.fetch_started = defaultdict(bool) # defaults to false + self.iter_timeout = iter_timeout + self.queue = Queue() super(SimpleConsumer, self).__init__( client, group, topic, @@ -267,12 +291,6 @@ class SimpleConsumer(Consumer): reqs.append(OffsetRequest(self.topic, partition, -2, 1)) elif whence == 2: reqs.append(OffsetRequest(self.topic, partition, -1, 1)) - - # The API returns back the next available offset - # For eg: if the current offset is 18, the API will return - # back 19. So, if we have to seek 5 points before, we will - # end up going back to 14, instead of 13. Adjust this - deltas[partition] -= 1 else: pass @@ -289,123 +307,111 @@ class SimpleConsumer(Consumer): count: Indicates the maximum number of messages to be fetched block: If True, the API will block till some messages are fetched. - timeout: If None, and block=True, the API will block infinitely. - If >0, API will block for specified time (in seconds) + timeout: If block is True, the function will block for the specified + time (in seconds) until count messages is fetched. If None, + it will block forever. """ messages = [] - iterator = self.__iter__() - - # HACK: This splits the timeout between available partitions if timeout: - timeout = timeout * 1.0 / len(self.offsets) + max_time = time.time() + timeout - with FetchContext(self, block, timeout): - while count > 0: - try: - messages.append(next(iterator)) - except StopIteration: - break + while count > 0 and (timeout is None or timeout > 0): + message = self.get_message(block, timeout) + if message: + messages.append(message) count -= 1 + else: + # Ran out of messages for the last request. + if not block: + # If we're not blocking, break. + break + if timeout: + # If we're blocking and have a timeout, reduce it to the + # appropriate value + timeout = max_time - time.time() return messages - def __iter__(self): - """ - Create an iterate per partition. Iterate through them calling next() - until they are all exhausted. - """ - iters = {} - for partition, offset in self.offsets.items(): - iters[partition] = self.__iter_partition__(partition, offset) + def get_message(self, block=True, timeout=0.1): + if self.queue.empty(): + # We're out of messages, go grab some more. + with FetchContext(self, block, timeout): + self._fetch() + try: + return self.queue.get_nowait() + except Empty: + return None - if len(iters) == 0: - return + def __iter__(self): + if self.iter_timeout is None: + timeout = ITER_TIMEOUT_SECONDS + else: + timeout = self.iter_timeout while True: - if len(iters) == 0: + message = self.get_message(True, timeout) + if message: + yield message + elif self.iter_timeout is None: + # We did not receive any message yet but we don't have a + # timeout, so give up the CPU for a while before trying again + time.sleep(NO_MESSAGES_WAIT_TIME_SECONDS) + else: + # Timed out waiting for a message break - for partition, it in iters.items(): + def _fetch(self): + # Create fetch request payloads for all the partitions + requests = [] + partitions = self.offsets.keys() + while partitions: + for partition in partitions: + requests.append(FetchRequest(self.topic, partition, + self.offsets[partition], + self.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) + + retry_partitions = set() + for resp in responses: + partition = resp.partition try: - if self.partition_info: - yield (partition, it.next()) + for message in resp.messages: + # Update partition offset + self.offsets[partition] = message.offset + 1 + + # Count, check and commit messages if necessary + self.count_since_commit += 1 + self._auto_commit() + + # Put the message in our queue + if self.partition_info: + self.queue.put((partition, message)) + else: + self.queue.put(message) + except ConsumerFetchSizeTooSmall, e: + if (self.max_buffer_size is not None and + self.buffer_size == self.max_buffer_size): + log.error("Max fetch size %d too small", + self.max_buffer_size) + raise e + if self.max_buffer_size is None: + self.buffer_size *= 2 else: - yield it.next() + self.buffer_size = max(self.buffer_size * 2, + self.max_buffer_size) + log.warn("Fetch size too small, increase to %d (2x) " + "and retry", self.buffer_size) + retry_partitions.add(partition) + except ConsumerNoMoreData, e: + log.debug("Iteration was ended by %r", e) except StopIteration: + # Stop iterating through this partition log.debug("Done iterating over partition %s" % partition) - del iters[partition] - - # skip auto-commit since we didn't yield anything - continue - - # Count, check and commit messages if necessary - self.count_since_commit += 1 - self._auto_commit() - - def __iter_partition__(self, partition, offset): - """ - Iterate over the messages in a partition. Create a FetchRequest - to get back a batch of messages, yield them one at a time. - After a batch is exhausted, start a new batch unless we've reached - the end of this partition. - """ - - # The offset that is stored in the consumer is the offset that - # we have consumed. In subsequent iterations, we are supposed to - # fetch the next message (that is from the next offset) - # However, for the 0th message, the offset should be as-is. - # An OffsetFetchRequest to Kafka gives 0 for a new queue. This is - # problematic, since 0 is offset of a message which we have not yet - # consumed. - if self.fetch_started[partition]: - offset += 1 - - fetch_size = self.fetch_min_bytes - - while True: - # use MaxBytes = client's bufsize since we're only - # fetching one topic + partition - req = FetchRequest( - self.topic, partition, offset, self.client.bufsize) - - (resp,) = self.client.send_fetch_request( - [req], - max_wait_time=self.fetch_max_wait_time, - min_bytes=fetch_size) - - assert resp.topic == self.topic - assert resp.partition == partition - - next_offset = None - try: - for message in resp.messages: - next_offset = message.offset - - # update the offset before the message is yielded. This - # is so that the consumer state is not lost in certain - # cases. - # - # For eg: the message is yielded and consumed by the - # caller, but the caller does not come back into the - # generator again. The message will be consumed but the - # status will not be updated in the consumer - self.fetch_started[partition] = True - self.offsets[partition] = message.offset - yield message - except ConsumerFetchSizeTooSmall, e: - fetch_size *= 1.5 - log.warn( - "Fetch size too small, increasing to %d (1.5x) and retry", - fetch_size) - continue - except ConsumerNoMoreData, e: - log.debug("Iteration was ended by %r", e) - - if next_offset is None: - break - else: - offset = next_offset + 1 - + partitions = retry_partitions def _mp_consume(client, group, topic, chunk, queue, start, exit, pause, size): """ @@ -443,8 +449,9 @@ def _mp_consume(client, group, topic, chunk, queue, start, exit, pause, size): # indicates a specific number of messages, follow that advice count = 0 - for partition, message in consumer: - queue.put((partition, message)) + message = consumer.get_message() + if message: + queue.put(message) count += 1 # We have reached the required size. The controller might have @@ -454,12 +461,11 @@ def _mp_consume(client, group, topic, chunk, queue, start, exit, pause, size): # can reset the 'start' event if count == size.value: pause.wait() - break - # In case we did not receive any message, give up the CPU for - # a while before we try again - if count == 0: - time.sleep(0.1) + else: + # In case we did not receive any message, give up the CPU for + # a while before we try again + time.sleep(NO_MESSAGES_WAIT_TIME_SECONDS) consumer.stop() @@ -504,7 +510,7 @@ class MultiProcessConsumer(Consumer): # Variables for managing and controlling the data flow from # consumer child process to master - self.queue = Queue(1024) # Child consumers dump messages into this + self.queue = MPQueue(1024) # Child consumers dump messages into this self.start = Event() # Indicates the consumers to start fetch self.exit = Event() # Requests the consumers to shutdown self.pause = Event() # Requests the consumers to pause fetch @@ -586,8 +592,9 @@ class MultiProcessConsumer(Consumer): count: Indicates the maximum number of messages to be fetched block: If True, the API will block till some messages are fetched. - timeout: If None, and block=True, the API will block infinitely. - If >0, API will block for specified time (in seconds) + timeout: If block is True, the function will block for the specified + time (in seconds) until count messages is fetched. If None, + it will block forever. """ messages = [] @@ -598,7 +605,10 @@ class MultiProcessConsumer(Consumer): self.size.value = count self.pause.clear() - while count > 0: + if timeout: + max_time = time.time() + timeout + + while count > 0 and (timeout is None or timeout > 0): # Trigger consumption only if the queue is empty # By doing this, we will ensure that consumers do not # go into overdrive and keep consuming thousands of @@ -618,6 +628,7 @@ class MultiProcessConsumer(Consumer): self.count_since_commit += 1 self._auto_commit() count -= 1 + timeout = max_time - time.time() self.size.value = 0 self.start.clear() diff --git a/kafka/producer.py b/kafka/producer.py index a82d99b..5aead43 100644 --- a/kafka/producer.py +++ b/kafka/producer.py @@ -69,7 +69,7 @@ def _send_upstream(topic, queue, client, batch_time, batch_size, client.send_produce_request(reqs, acks=req_acks, timeout=ack_timeout) - except Exception as exp: + except Exception: log.exception("Unable to send message") @@ -146,9 +146,9 @@ class Producer(object): try: resp = self.client.send_produce_request([req], acks=self.req_acks, timeout=self.ack_timeout) - except Exception as e: + except Exception: log.exception("Unable to send messages") - raise e + raise return resp def stop(self, timeout=1): diff --git a/kafka/protocol.py b/kafka/protocol.py index 612acf6..54b8eee 100644 --- a/kafka/protocol.py +++ b/kafka/protocol.py @@ -119,9 +119,17 @@ class KafkaProtocol(object): read_message = True yield OffsetAndMessage(offset, message) except BufferUnderflowError: + # NOTE: Not sure this is correct error handling: + # Is it possible to get a BUE if the message set is somewhere + # in the middle of the fetch response? If so, we probably have + # an issue that's not fetch size too small. + # Aren't we ignoring errors if we fail to unpack data by + # raising StopIteration()? + # If _decode_message() raises a ChecksumError, couldn't that + # also be due to the fetch size being too small? if read_message is False: # If we get a partial read of a message, but haven't - # yielded anyhting there's a problem + # yielded anything there's a problem raise ConsumerFetchSizeTooSmall() else: raise StopIteration() @@ -171,7 +179,7 @@ class KafkaProtocol(object): Params ====== client_id: string - correlation_id: string + correlation_id: int payloads: list of ProduceRequest acks: How "acky" you want the request to be 0: immediate response @@ -231,7 +239,7 @@ class KafkaProtocol(object): Params ====== client_id: string - correlation_id: string + correlation_id: int payloads: list of FetchRequest max_wait_time: int, how long to block waiting on min_bytes of data min_bytes: int, the minimum number of bytes to accumulate before @@ -338,7 +346,7 @@ class KafkaProtocol(object): Params ====== client_id: string - correlation_id: string + correlation_id: int topics: list of strings """ topics = [] if topics is None else topics @@ -376,12 +384,16 @@ class KafkaProtocol(object): topic_metadata = {} for i in range(num_topics): + # NOTE: topic_error is discarded. Should probably be returned with + # the topic metadata. ((topic_error,), cur) = relative_unpack('>h', data, cur) (topic_name, cur) = read_short_string(data, cur) ((num_partitions,), cur) = relative_unpack('>i', data, cur) partition_metadata = {} for j in range(num_partitions): + # NOTE: partition_error_code is discarded. Should probably be + # returned with the partition metadata. ((partition_error_code, partition, leader, numReplicas), cur) = \ relative_unpack('>hiii', data, cur) @@ -408,7 +420,7 @@ class KafkaProtocol(object): Params ====== client_id: string - correlation_id: string + correlation_id: int group: string, the consumer group you are committing offsets for payloads: list of OffsetCommitRequest """ @@ -459,7 +471,7 @@ class KafkaProtocol(object): Params ====== client_id: string - correlation_id: string + correlation_id: int group: string, the consumer group you are fetching offsets for payloads: list of OffsetFetchRequest """ |