[
https://issues.apache.org/jira/browse/KAFKA-13310?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17417270#comment-17417270
]
RivenSun commented on KAFKA-13310:
----------------------------------
[~showuon]
After thinking about it, here is my latest points.
Hope you can reply me as soon as possible, thank you.
1. Because the poll method of KafkaConsumer promises to return within the
time specified by the customer, unless the customer sets the
ConsumerRebalanceListener to perform a time-consuming operation. See the
comment of poll(final Duration timeout) method as below:
{code:java}
This method returns immediately if there are records available. Otherwise, it
will await the passed timeout. If the timeout expires, an empty record set will
be returned. Note that this method may block beyond the timeout in order to
execute custom ConsumerRebalanceListener callbacks.{code}
In order to keep this promise, I also suggest passing the *Timer set by the
customer for poll()* to the maybeAutoCommitOffsetsSync(*Timer timer*) method of
ConsumerCoordinator
2. Modify ConsumerCoordinator's maybeAutoCommitOffsetsSync(Timer timer)
method
Below are my preliminary code changes：
{code:java}
private void maybeAutoCommitOffsetsSync(Timer timer) {
if (autoCommitEnabled) {
Map<TopicPartition, OffsetAndMetadata> allConsumedOffsets =
subscriptions.allConsumed();
cleanUpConsumedOffsets(allConsumedOffsets);
try {
log.debug("Sending synchronous auto-commit of offsets {}",
allConsumedOffsets);
if (!commitOffsetsSync(allConsumedOffsets, timer))
log.debug("Auto-commit of offsets {} timed out before
completion", allConsumedOffsets);
} catch (WakeupException | InterruptException e) {
log.debug("Auto-commit of offsets {} was interrupted before
completion", allConsumedOffsets);
// rethrow wakeups since they are triggered by the user
throw e;
} catch (Exception e) {
// consistent with async auto-commit failures, we do not propagate
the exception
log.warn("Synchronous auto-commit of offsets {} failed: {}",
allConsumedOffsets, e.getMessage());
}
}
}{code}
{code:java}
private void cleanUpConsumedOffsets(Map<TopicPartition, OffsetAndMetadata>
willCommitOffsets) {
if (willCommitOffsets.isEmpty())
return;
Set<String> subscription = subscriptions.subscription();
Set<TopicPartition> toGiveUpTopicPartitions = new HashSet<>();
Iterator<Map.Entry<TopicPartition, OffsetAndMetadata>> iterator =
willCommitOffsets.entrySet().iterator();
while (iterator.hasNext()) {
Map.Entry<TopicPartition, OffsetAndMetadata> entry = iterator.next();
if (!subscription.contains(entry.getKey().topic())) {
toGiveUpTopicPartitions.add(entry.getKey());
iterator.remove();
}
}
if (toGiveUpTopicPartitions.size() > 0) {
//Because toGiveUpTopicPartitions may receive
`UnknownTopicOrPartitionException` when submitting their offsets.
//We are prepared to abandon them. The worst effect is that these
partitions may repeatedly consume some messages
log.warn("Synchronous auto-commit of offsets {} will be abandoned",
toGiveUpTopicPartitions);
}
}
{code}
> KafkaConsumer cannot jump out of the poll method, and the consumer is blocked
> in the ConsumerCoordinator method maybeAutoCommitOffsetsSync(Timer timer).
> Cpu and traffic of Broker‘s side increase sharply
> -----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: KAFKA-13310
> URL: https://issues.apache.org/jira/browse/KAFKA-13310
> Project: Kafka
> Issue Type: Bug
> Components: consumer
> Affects Versions: 2.8.1
> Environment: prod
> Reporter: RivenSun
> Priority: Major
> Attachments: SecondDeleteConsumerLog.png, SecondDeleteDebugLog.png,
> brokerCpu.png, brokerNetBytes.png, kafkaConsumerLog.png
>
>
> h2. Foreword
> Because our consumers' consumption logic is sometimes heavier, we refer
> to the configuration of Kafka stream
> [https://kafka.apache.org/documentation/#upgrade_10201_notable]
> Set max.poll.interval.ms to Integer.MAX_VALUE
> Our consumers have adopted method :
> consumer.subscribe(Pattern.compile(".*riven.*"));
>
> h2. Recurrence of the problem scene
> operate steps are
> (1) Test environment Kafka cluster: three brokers
> (2) Topics conforming to regular expressions include rivenTest1, rivenTest2,
> and rivenTest88
> (3) Only one consumer is needed, group.id is "rivenReassign",
> consumer.subscribe(Pattern.compile(".*riven.*"));
> (4) At the beginning, the group status is stable, and everything is normal
> for consumers, then I delete topic: rivenTest88
>
> h2. Phenomenon
> Problem phenomenon
> (1) The consumer is blocked in the poll method, no longer consume any
> messages, and the consumer log is always printing
> [main] WARN
> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator-[Consumer
> clientId=consumer-rivenReassign-1, groupId=rivenReassign] Offset commit
> failed on partition rivenTest88-1 at offset 0: This server does not host this
> topic-partition.
> (2) The describe consumerGroup interface of Adminclient has always timed
> out, and the group status is no longer stable
> (3) The cpu and traffic of the broker are *significantly increased*
>
>
> h2. Problem tracking
> By analyzing the kafkaConsumer code, the version is 2.8.1.
> I found that you introduced the waitForJoinGroup variable in the
> updateAssignmentMetadataIfNeeded method. For the reason, I attached the
> comment on the method: "try to update assignment metadata BUT do not need to
> block on the timer for join group". See as below:
>
> {code:java}
> if (includeMetadataInTimeout) {
> // try to update assignment metadata BUT do not need to block on the
> timer for join group
> updateAssignmentMetadataIfNeeded(timer, false);
> } else {
> while (!updateAssignmentMetadataIfNeeded(time.timer(Long.MAX_VALUE),
> true)) {
> log.warn("Still waiting for metadata");
> }
> }{code}
>
>
> By tracing the code back layer by layer, it is found that the function of
> this variable is to construct a time.timer(0L) and pass it back to the method
> joinGroupIfNeeded (final Timer timer) in AbstractCoordinator. See as below:
> {code:java}
> // if not wait for join group, we would just use a timer of 0
> if (!ensureActiveGroup(waitForJoinGroup ? timer : time.timer(0L))) {
> // since we may use a different timer in the callee, we'd still need
> // to update the original timer's current time after the call
> timer.update(time.milliseconds());
> return false;
> }
> {code}
> But you will find that there is a submethod onJoinPrepare in the method
> stack of joinGroupIfNeeded, and then there is a line of code in the
> onJoinPrepare method
> maybeAutoCommitOffsetsSync(time.timer(rebalanceConfig.rebalanceTimeoutMs)),
> the value of rebalanceConfig.rebalanceTimeoutMs is actually
> max.poll.interval.ms.
> Finally, I tracked down ConsumerCoordinator's method
> commitOffsetsSync(Map<TopicPartition, OffsetAndMetadata> offsets, Timer timer)
> The input parameter offsets is subscriptions.allConsumed(), when I delete
> the topic: rivenTest88, commitOffsetsSync(Map<TopicPartition,
> OffsetAndMetadata> offsets, Timer timer) method will *fall into an infinite
> loop! !*
> {code:java}
> public boolean commitOffsetsSync(Map<TopicPartition, OffsetAndMetadata>
> offsets, Timer timer) {
> invokeCompletedOffsetCommitCallbacks();
> if (offsets.isEmpty())
> return true;
> do {
> if (coordinatorUnknown() && !ensureCoordinatorReady(timer)) {
> return false;
> }
> RequestFuture<Void> future = sendOffsetCommitRequest(offsets);
> client.poll(future, timer);
> // We may have had in-flight offset commits when the synchronous commit
> began. If so, ensure that
> // the corresponding callbacks are invoked prior to returning in order to
> preserve the order that
> // the offset commits were applied.
> invokeCompletedOffsetCommitCallbacks();
> if (future.succeeded()) {
> if (interceptors != null)
> interceptors.onCommit(offsets);
> return true;
> }
> if (future.failed() && !future.isRetriable())
> throw future.exception();
> timer.sleep(rebalanceConfig.retryBackoffMs);
> } while (timer.notExpired());
> return false;
> }{code}
>
>
> *The reason for the endless loop is:*
> (1) The expiration time of the timer is too long, which is
> max.poll.interval.ms
> (2) The offsets to be submitted contain dirty data and TopicPartition that
> no longer exists
> (3) The response future of sendOffsetCommitRequest(final Map<TopicPartition,
> OffsetAndMetadata> offsets) has always failed, and the exception in the
> future is UnknownTopicOrPartitionException. This exception is allowed to be
> retried.
> Then since the infinite loop interval above is 100ms by default,
> timer.sleep(rebalanceConfig.retryBackoffMs);
> If a large number of consumers have this problem at the same time, a large
> number of network requests will be generated to the Kafka broker, *resulting
> in a sharp increase in the cpu and traffic of the broker machine!*
>
>
> h2. Suggest
> 1.maybeAutoCommitOffsetsSync(time.timer(rebalanceConfig.rebalanceTimeoutMs)),
> the time of this method is recommended not to use max.poll.interval.ms,
> This parameter is open to users to configure. Through the explanation of
> this parameter on the official website, I would never think that this
> parameter will be used in this place. At the same time, it will block
> KafkaConsumer's poll (final Duration timeout), even if I set consumer.poll
> (Duration.ofMillis(1000)).
> 2. In fact, in the poll (Timer timer, boolean waitForJoinGroup) method of
> ConsumerCoordinatord, before calling the ensureActiveGroup method, the
> consumer ensures that the local metadata is up to date, see the code
>
> {code:java}
> if (rejoinNeededOrPending()) {
> // due to a race condition between the initial metadata fetch and the
> initial rebalance,
> // we need to ensure that the metadata is fresh before joining initially.
> This ensures
> // that we have matched the pattern against the cluster's topics at least
> once before joining.
> if (subscriptions.hasPatternSubscription()) {
> // For consumer group that uses pattern-based subscription, after a
> topic is created,
> // any consumer that discovers the topic after metadata refresh can
> trigger rebalance
> // across the entire consumer group. Multiple rebalances can be
> triggered after one topic
> // creation if consumers refresh metadata at vastly different times.
> We can significantly
> // reduce the number of rebalances caused by single topic creation by
> asking consumer to
> // refresh metadata before re-joining the group as long as the
> refresh backoff time has
> // passed.
> if (this.metadata.timeToAllowUpdate(time.milliseconds()) == 0) {
> this.metadata.requestUpdate();
> }
> if (!client.ensureFreshMetadata(timer)) {
> return false;
> }
> }
> if (!ensureActiveGroup(timer)) {
> return false;
> }
> }
> {code}
>
> That is to say, the consumer knows which topic/topicPartition is legal before
> onJoinPrepare. In this case, why didn't you find the
> UnknownTopicOrPartitionException in the commitOffsetsSync method mentioned
> above,do not put the submitted offsets and the latest local metadata together
> for analysis, remove the non-existent topicpartitions, and then try to submit
> the offsets again. I think I can break out of the infinite loop by doing this
> 3. Why must the offset be submitted synchronously in the onJoinPrepare
> method? Can't the offset be submitted asynchronously? Or provide a parameter
> for the user to choose whether to submit synchronously or asynchronously. Or
> provide a new parameter to control the maximum number of retries for
> synchronous submission here, instead of using the Timer constructed by
> max.poll.interval.ms.
> And if you don’t really submit the offset here, it will not have much
> impact. It may cause repeated consumption of some messages. I still suggest
> to provide a new parameter to control whether you need to submit the offset.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
