Merge pull request #331 from vshlapakov/feature-producer-retries

Async producer retries for failed messages

Author: dpkp

kafka/common.py

@@ -71,6 +71,11 @@
 KafkaMessage = namedtuple("KafkaMessage",
     ["topic", "partition", "offset", "key", "value"])
 
+# Define retry policy for async producer
+# Limit value: int >= 0, 0 means no retries
+RetryOptions = namedtuple("RetryOptions",
+    ["limit", "backoff_ms", "retry_on_timeouts"])
+
 
 #################
 #   Exceptions  #
@@ -205,6 +210,12 @@ class KafkaConfigurationError(KafkaError):
     pass
 
 
+class AsyncProducerQueueFull(KafkaError):
+    def __init__(self, failed_msgs, *args):
+        super(AsyncProducerQueueFull, self).__init__(*args)
+        self.failed_msgs = failed_msgs
+
+
 def _iter_broker_errors():
     for name, obj in inspect.getmembers(sys.modules[__name__]):
         if inspect.isclass(obj) and issubclass(obj, BrokerResponseError) and obj != BrokerResponseError:
@@ -218,3 +229,18 @@ def check_error(response):
     if response.error:
         error_class = kafka_errors.get(response.error, UnknownError)
         raise error_class(response)
+
+
+RETRY_BACKOFF_ERROR_TYPES = (
+    KafkaUnavailableError, LeaderNotAvailableError,
+    ConnectionError, FailedPayloadsError
+)
+
+
+RETRY_REFRESH_ERROR_TYPES = (
+    NotLeaderForPartitionError, UnknownTopicOrPartitionError,
+    LeaderNotAvailableError, ConnectionError
+)
+
+
+RETRY_ERROR_TYPES = RETRY_BACKOFF_ERROR_TYPES + RETRY_REFRESH_ERROR_TYPES

kafka/producer/base.py

@@ -5,18 +5,23 @@
 import time
 
 try:
-    from queue import Empty, Queue
+    from queue import Empty, Full, Queue
 except ImportError:
-    from Queue import Empty, Queue
+    from Queue import Empty, Full, Queue
 from collections import defaultdict
 
 from threading import Thread, Event
 
 import six
 
 from kafka.common import (
-    ProduceRequest, TopicAndPartition, UnsupportedCodecError
+    ProduceRequest, TopicAndPartition, RetryOptions,
+    kafka_errors, UnsupportedCodecError, FailedPayloadsError,
+    RequestTimedOutError, AsyncProducerQueueFull, UnknownError
 )
+from kafka.common import (
+    RETRY_ERROR_TYPES, RETRY_BACKOFF_ERROR_TYPES, RETRY_REFRESH_ERROR_TYPES)
+
 from kafka.protocol import CODEC_NONE, ALL_CODECS, create_message_set
 from kafka.util import kafka_bytestring
 
@@ -25,21 +30,33 @@
 BATCH_SEND_DEFAULT_INTERVAL = 20
 BATCH_SEND_MSG_COUNT = 20
 
+# unlimited
+ASYNC_QUEUE_MAXSIZE = 0
+ASYNC_QUEUE_PUT_TIMEOUT = 0
+# no retries by default
+ASYNC_RETRY_LIMIT = 0
+ASYNC_RETRY_BACKOFF_MS = 0
+ASYNC_RETRY_ON_TIMEOUTS = False
+
 STOP_ASYNC_PRODUCER = -1
 
 
 def _send_upstream(queue, client, codec, batch_time, batch_size,
-                   req_acks, ack_timeout, stop_event):
+                   req_acks, ack_timeout, retry_options, stop_event):
     """
     Listen on the queue for a specified number of messages or till
     a specified timeout and send them upstream to the brokers in one
     request
     """
-    stop = False
+    reqs = {}
+    client.reinit()
 
     while not stop_event.is_set():
         timeout = batch_time
-        count = batch_size
+
+        # it's a simplification: we're comparing message sets and
+        # messages: each set can contain [1..batch_size] messages
+        count = batch_size - len(reqs)
         send_at = time.time() + timeout
         msgset = defaultdict(list)
 
@@ -48,7 +65,6 @@ def _send_upstream(queue, client, codec, batch_time, batch_size,
         while count > 0 and timeout >= 0:
             try:
                 topic_partition, msg, key = queue.get(timeout=timeout)
-
             except Empty:
                 break
 
@@ -63,20 +79,60 @@ def _send_upstream(queue, client, codec, batch_time, batch_size,
             msgset[topic_partition].append((msg, key))
 
         # Send collected requests upstream
-        reqs = []
         for topic_partition, msg in msgset.items():
             messages = create_message_set(msg, codec, key)
             req = ProduceRequest(topic_partition.topic,
                                  topic_partition.partition,
-                                 messages)
-            reqs.append(req)
+                                 tuple(messages))
+            reqs[req] = 0
+
+        if not reqs:
+            continue
+
+        reqs_to_retry, error_cls = [], None
+        do_backoff, do_refresh = False, False
+
+        def _handle_error(error_cls, reqs, all_retries):
+            if ((error_cls == RequestTimedOutError and
+                 retry_options.retry_on_timeouts) or
+                    error_cls in RETRY_ERROR_TYPES):
+                all_retries += reqs
+            if error_cls in RETRY_BACKOFF_ERROR_TYPES:
+                do_backoff = True
+            if error_cls in RETRY_REFRESH_ERROR_TYPES:
+                do_refresh = True
 
         try:
-            client.send_produce_request(reqs,
-                                        acks=req_acks,
-                                        timeout=ack_timeout)
-        except Exception:
-            log.exception("Unable to send message")
+            reply = client.send_produce_request(reqs.keys(),
+                                                acks=req_acks,
+                                                timeout=ack_timeout,
+                                                fail_on_error=False)
+            for i, response in enumerate(reply):
+                if isinstance(response, FailedPayloadsError):
+                    _handle_error(FailedPayloadsError, response.failed_payloads, reqs_to_retry)
+                elif isinstance(response, ProduceResponse) and response.error:
+                    error_cls = kafka_errors.get(response.error, UnknownError)
+                    _handle_error(error_cls, [reqs.keys()[i]], reqs_to_retry)
+
+        except Exception as ex:
+            error_cls = kafka_errors.get(type(ex), UnknownError)
+            _handle_error(error_cls, reqs.keys(), reqs_to_retry)
+
+        if not reqs_to_retry:
+            reqs = {}
+            continue
+
+        # doing backoff before next retry
+        if do_backoff and retry_options.backoff_ms:
+            log.info("Doing backoff for %s(ms)." % retry_options.backoff_ms)
+            time.sleep(float(retry_options.backoff_ms) / 1000)
+
+        # refresh topic metadata before next retry
+        if do_refresh:
+            client.load_metadata_for_topics()
+
+        reqs = dict((key, count + 1) for (key, count) in reqs.items()
+                if key in reqs_to_retry and count < retry_options.limit)
 
 
 class Producer(object):
@@ -111,12 +167,18 @@ def __init__(self, client, async=False,
                  codec=None,
                  batch_send=False,
                  batch_send_every_n=BATCH_SEND_MSG_COUNT,
-                 batch_send_every_t=BATCH_SEND_DEFAULT_INTERVAL):
+                 batch_send_every_t=BATCH_SEND_DEFAULT_INTERVAL,
+                 async_retry_limit=ASYNC_RETRY_LIMIT,
+                 async_retry_backoff_ms=ASYNC_RETRY_BACKOFF_MS,
+                 async_retry_on_timeouts=ASYNC_RETRY_ON_TIMEOUTS,
+                 async_queue_maxsize=ASYNC_QUEUE_MAXSIZE,
+                 async_queue_put_timeout=ASYNC_QUEUE_PUT_TIMEOUT):
 
         if batch_send:
             async = True
             assert batch_send_every_n > 0
             assert batch_send_every_t > 0
+            assert async_queue_maxsize >= 0
         else:
             batch_send_every_n = 1
             batch_send_every_t = 3600
@@ -135,10 +197,13 @@ def __init__(self, client, async=False,
         self.codec = codec
 
         if self.async:
-            log.warning("async producer does not guarantee message delivery!")
-            log.warning("Current implementation does not retry Failed messages")
-            log.warning("Use at your own risk! (or help improve with a PR!)")
-            self.queue = Queue()  # Messages are sent through this queue
+            # Messages are sent through this queue
+            self.queue = Queue(async_queue_maxsize)
+            self.async_queue_put_timeout = async_queue_put_timeout
+            async_retry_options = RetryOptions(
+                limit=async_retry_limit,
+                backoff_ms=async_retry_backoff_ms,
+                retry_on_timeouts=async_retry_on_timeouts)
             self.thread_stop_event = Event()
             self.thread = Thread(target=_send_upstream,
                                  args=(self.queue,
@@ -148,6 +213,7 @@ def __init__(self, client, async=False,
                                        batch_send_every_n,
                                        self.req_acks,
                                        self.ack_timeout,
+                                       async_retry_options,
                                        self.thread_stop_event))
 
             # Thread will die if main thread exits
@@ -199,8 +265,18 @@ def _send_messages(self, topic, partition, *msg, **kwargs):
             raise TypeError("the key must be type bytes")
 
         if self.async:
-            for m in msg:
-                self.queue.put((TopicAndPartition(topic, partition), m, key))
+            for idx, m in enumerate(msg):
+                try:
+                    item = (TopicAndPartition(topic, partition), m, key)
+                    if self.async_queue_put_timeout == 0:
+                        self.queue.put_nowait(item)
+                    else:
+                        self.queue.put(item, True, self.async_queue_put_timeout)
+                except Full:
+                    raise AsyncProducerQueueFull(
+                        msg[idx:],
+                        'Producer async queue overfilled. '
+                        'Current queue size %d.' % self.queue.qsize())
             resp = []
         else:
             messages = create_message_set([(m, key) for m in msg], self.codec, key)

kafka/producer/keyed.py

@@ -7,7 +7,8 @@
 
 from .base import (
     Producer, BATCH_SEND_DEFAULT_INTERVAL,
-    BATCH_SEND_MSG_COUNT
+    BATCH_SEND_MSG_COUNT, ASYNC_QUEUE_MAXSIZE, ASYNC_QUEUE_PUT_TIMEOUT,
+    ASYNC_RETRY_LIMIT, ASYNC_RETRY_BACKOFF_MS, ASYNC_RETRY_ON_TIMEOUTS
 )
 
 log = logging.getLogger("kafka")
@@ -37,7 +38,12 @@ def __init__(self, client, partitioner=None, async=False,
                  codec=None,
                  batch_send=False,
                  batch_send_every_n=BATCH_SEND_MSG_COUNT,
-                 batch_send_every_t=BATCH_SEND_DEFAULT_INTERVAL):
+                 batch_send_every_t=BATCH_SEND_DEFAULT_INTERVAL,
+                 async_retry_limit=ASYNC_RETRY_LIMIT,
+                 async_retry_backoff_ms=ASYNC_RETRY_BACKOFF_MS,
+                 async_retry_on_timeouts=ASYNC_RETRY_ON_TIMEOUTS,
+                 async_queue_maxsize=ASYNC_QUEUE_MAXSIZE,
+                 async_queue_put_timeout=ASYNC_QUEUE_PUT_TIMEOUT):
         if not partitioner:
             partitioner = HashedPartitioner
         self.partitioner_class = partitioner
@@ -46,7 +52,12 @@ def __init__(self, client, partitioner=None, async=False,
         super(KeyedProducer, self).__init__(client, async, req_acks,
                                             ack_timeout, codec, batch_send,
                                             batch_send_every_n,
-                                            batch_send_every_t)
+                                            batch_send_every_t,
+                                            async_retry_limit,
+                                            async_retry_backoff_ms,
+                                            async_retry_on_timeouts,
+                                            async_queue_maxsize,
+                                            async_queue_put_timeout)
 
     def _next_partition(self, topic, key):
         if topic not in self.partitioners:

kafka/producer/simple.py

@@ -10,7 +10,8 @@
 
 from .base import (
     Producer, BATCH_SEND_DEFAULT_INTERVAL,
-    BATCH_SEND_MSG_COUNT
+    BATCH_SEND_MSG_COUNT, ASYNC_QUEUE_MAXSIZE, ASYNC_QUEUE_PUT_TIMEOUT,
+    ASYNC_RETRY_LIMIT, ASYNC_RETRY_BACKOFF_MS, ASYNC_RETRY_ON_TIMEOUTS
 )
 
 log = logging.getLogger("kafka")
@@ -45,13 +46,23 @@ def __init__(self, client, async=False,
                  batch_send=False,
                  batch_send_every_n=BATCH_SEND_MSG_COUNT,
                  batch_send_every_t=BATCH_SEND_DEFAULT_INTERVAL,
-                 random_start=True):
+                 random_start=True,
+                 async_retry_limit=ASYNC_RETRY_LIMIT,
+                 async_retry_backoff_ms=ASYNC_RETRY_BACKOFF_MS,
+                 async_retry_on_timeouts=ASYNC_RETRY_ON_TIMEOUTS,
+                 async_queue_maxsize=ASYNC_QUEUE_MAXSIZE,
+                 async_queue_put_timeout=ASYNC_QUEUE_PUT_TIMEOUT):
         self.partition_cycles = {}
         self.random_start = random_start
         super(SimpleProducer, self).__init__(client, async, req_acks,
                                              ack_timeout, codec, batch_send,
                                              batch_send_every_n,
-                                             batch_send_every_t)
+                                             batch_send_every_t,
+                                             async_retry_limit,
+                                             async_retry_backoff_ms,
+                                             async_retry_on_timeouts,
+                                             async_queue_maxsize,
+                                             async_queue_put_timeout)
 
     def _next_partition(self, topic):
         if topic not in self.partition_cycles: