kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Phillip Walker <pwal...@bandwidth.com>
Subject Re: Consumer group disappears and consumers loops
Date Thu, 31 Dec 2015 17:41:25 GMT
I've run into exactly the same problem, and a similar workaround, along
with a config change, seems to be working for now, but I can probably add a
little detail.

Scenario: The code reads from a very high-volume topic on a remote server
(with 120 partitions and no SSL or security). I am currently debugging, so
I am using only a single thread. Several calls to consumer.poll(100) were
always required to start retrieving any records at all, and only 40
partitions were returned on each call that returns records. At first, I was
getting about 24,000 records per call and was not finished with processing
them (running through an IDE) within the session timeout period, getting
the same messages as Martin. That part, I would expect, since I wasn't
calling consumer.poll() often enough to prevent the session from timing out.

I initially tried a workaround where I paused all partitions I owned,
called consumer.poll(0) well within the heartbeat time, and then resumed
the partitions as a sort of keep-alive mechanism. I stilled timed out every
30 seconds, so apparently I wasn't heartbeating that way.

I dramatically decreased max.partitions.fetch.bytes so that I would get far
fewer records that I could easily process within the heartbeat time. With
just that change, after a call to consumer.poll(100) that returned records,
I would still have a number of subsequent consumer.poll(100) calls return 0
records (despite being nowhere near the end of any partition). In fact,
those calls that returned 0 records did not appear to be counting as a
heartbeat, because I was still getting the "Group testGroup generation 1 is
dead and removed" message even as the poll method was being called
repeatedly and simultaneously and returning nothing.

By lowering max.partitions.fetch.bytes and boosting the consumer.poll()
wait time to a number greater than the heartbeat time, I started receiving
records on every call and am having no timeout issues. I still get records
from only 40 partitions per call, though, despite consumer.assignment()
returning the full 120.



[image: email-signature-logo.jpg]

Phillip Walker

Software Developer

Bandwidth <http://www.bandwidth.com/>

o 919.238.1452

m 919.578.1964

e email@bandwidth.com

LinkedIn <https://www.linkedin.com/in/phillipwalker>  |  Twitter
<https://twitter.com/bandwidth>

On Tue, Dec 22, 2015 at 11:31 AM, Rune Sørensen <rune@falconsocial.com>
wrote:

> Hi,
>
> Sorry for the long delay in replying.
>
> As for your questions:
> No we are not using SSL.
> The problem went away for Martin when he was running against his kafka
> instance locally on his development machine, but we are still seeing the
> issue when we run it in our testing environment, where the broker is on a
> remote machine.
> The network in our testing environment seems stable, from the measurements
> I have made so far.
>
> *Rune Tor*
> Sørensen
> +45 3172 2097 <+4531722097>
> LinkedIn <https://www.linkedin.com/in/runets> Twitter
> <https://twitter.com/Areian>
> *Copenhagen*
> Falcon Social
> H.C. Andersens Blvd. 27
> 1553 Copenhagen
> *Budapest*
> Falcon Social
> Colabs Startup Center Zrt
> 1016 Budapest, Krisztina krt. 99
> [image: Falcon Social]
> <
> https://www.falconsocial.com/?utm_source=Employee%20emails&utm_medium=email&utm_content=Rune%20Tor&utm_campaign=Mail%20signature
> >
> Social Media Management for Enterprise
>
> On Tue, Dec 1, 2015 at 11:56 PM, Jason Gustafson <jason@confluent.io>
> wrote:
>
> > I've been unable to reproduce this issue running locally. Even with a
> poll
> > timeout of 1 millisecond, it seems to work as expected. It would be
> helpful
> > to know a little more about your setup. Are you using SSL? Are the
> brokers
> > remote? Is the network stable?
> >
> > Thanks,
> > Jason
> >
> > On Tue, Dec 1, 2015 at 10:06 AM, Jason Gustafson <jason@confluent.io>
> > wrote:
> >
> > > Hi Martin,
> > >
> > > I'm also not sure why the poll timeout would affect this. Perhaps the
> > > handler is still doing work (e.g. sending requests) when the record set
> > is
> > > empty?
> > >
> > > As a general rule, I would recommend longer poll timeouts. I've
> actually
> > > tended to use Long.MAX_VALUE myself. I'll have a look just to make sure
> > > everything still works with smaller values though.
> > >
> > > -Jason
> > >
> > >
> > >
> > > On Tue, Dec 1, 2015 at 2:35 AM, Martin Skøtt <
> > > martin.skoett@falconsocial.com> wrote:
> > >
> > >> Hi Jason,
> > >>
> > >> That actually sounds like a very plausible explanation. My current
> > >> consumer
> > >> is using the default settings, but I have previously used these (taken
> > >> from
> > >> the sample in the Javadoc for the new KafkaConsumer):
> > >>  "auto.commit.interval.ms", "1000"
> > >>  "session.timeout.ms", "30000"
> > >>
> > >> My consumer loop is quite simple as it just calls a domain specific
> > >> service:
> > >>
> > >> while (true) {
> > >>     ConsumerRecords<String, Object> records = consumer.poll(10000);
> > >>     for (ConsumerRecord<String, Object> record : records) {
> > >>         serve.handle(record.topic(), record.value());
> > >>     }
> > >> }
> > >>
> > >> The domain service does a number of things (including lookups in a
> RDBMS
> > >> and saving to ElasticSearch). In my local test setup a poll will often
> > >> result between 5.000 and 10.000 records and I can easily see the
> > >> processing
> > >> of those taking more than 30 seconds.
> > >>
> > >> I'll probably take a look at adding some threading to my consumer and
> > add
> > >> more partitions to my topics.
> > >>
> > >> That is all fine, but it doesn't really explain why increasing poll
> > >> timeout
> > >> made the problem go away :-/
> > >>
> > >> Martin
> > >>
> > >> On 30 November 2015 at 19:30, Jason Gustafson <jason@confluent.io>
> > wrote:
> > >>
> > >> > Hey Martin,
> > >> >
> > >> > At a glance, it looks like your consumer's session timeout is
> > expiring.
> > >> > This shouldn't happen unless there is a delay between successive
> calls
> > >> to
> > >> > poll which is longer than the session timeout. It might help if you
> > >> include
> > >> > a snippet of your poll loop and your configuration (i.e. any
> > overridden
> > >> > settings).
> > >> >
> > >> > -Jason
> > >> >
> > >> > On Mon, Nov 30, 2015 at 8:12 AM, Martin Skøtt <
> > >> > martin.skoett@falconsocial.com> wrote:
> > >> >
> > >> > > Well, I made the problem go away, but I'm not sure why it works
> :-/
> > >> > >
> > >> > > Previously I used a time out value of 100 for Consumer.poll().
> > >> Increasing
> > >> > > it to 10.000 makes the problem go away completely?! I tried other
> > >> values
> > >> > as
> > >> > > well:
> > >> > >    - 0 problem remained
> > >> > >    - 3000, same as heartbeat.interval, problem remained, but
less
> > >> > frequent
> > >> > >
> > >> > > Not really sure what is going on, but happy that the problem
went
> > away
> > >> > :-)
> > >> > >
> > >> > > Martin
> > >> > >
> > >> > > On 30 November 2015 at 15:33, Martin Skøtt <
> > >> > martin.skoett@falconsocial.com
> > >> > > >
> > >> > > wrote:
> > >> > >
> > >> > > > Hi Guozhang,
> > >> > > >
> > >> > > > I have done some testing with various values of
> > >> heartbeat.interval.ms
> > >> > > and
> > >> > > > they don't seem to have any influence on the error messages.
> > Running
> > >> > > > kafka-consumer-groups also continues to return that the
consumer
> > >> groups
> > >> > > > does not exists or is rebalancing. Do you have any suggestions
> to
> > >> how I
> > >> > > > could debug this further?
> > >> > > >
> > >> > > > Regards,
> > >> > > > Martin
> > >> > > >
> > >> > > >
> > >> > > > On 25 November 2015 at 18:37, Guozhang Wang <wangguoz@gmail.com
> >
> > >> > wrote:
> > >> > > >
> > >> > > >> Hello Martin,
> > >> > > >>
> > >> > > >> It seems your consumer's heartbeat.interval.ms config
value is
> > too
> > >> > > small
> > >> > > >> (default is 3 seconds) for your environment, consider
> increasing
> > it
> > >> > and
> > >> > > >> see
> > >> > > >> if this issue goes away.
> > >> > > >>
> > >> > > >> At the same time, we have some better error handling
fixes in
> > trunk
> > >> > > which
> > >> > > >> will be included in the next point release.
> > >> > > >>
> > >> > > >> https://issues.apache.org/jira/browse/KAFKA-2860
> > >> > > >>
> > >> > > >> Guozhang
> > >> > > >>
> > >> > > >>
> > >> > > >>
> > >> > > >> On Wed, Nov 25, 2015 at 6:54 AM, Martin Skøtt <
> > >> > > >> martin.skoett@falconsocial.com> wrote:
> > >> > > >>
> > >> > > >> > Hi,
> > >> > > >> >
> > >> > > >> > I'm experiencing some very strange issues with
0.9. I get
> these
> > >> log
> > >> > > >> > messages from the new consumer:
> > >> > > >> >
> > >> > > >> > [main] ERROR
> > >> > > >> >
> > org.apache.kafka.clients.consumer.internals.ConsumerCoordinator -
> > >> > > Error
> > >> > > >> > ILLEGAL_GENERATION occurred while committing offsets
for
> group
> > >> > > >> > aaa-bbb-reader
> > >> > > >> > [main] WARN
> > >> > > >> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator
> > >> > > >> > - Auto offset commit failed: Commit cannot be completed
due
> to
> > >> group
> > >> > > >> > rebalance
> > >> > > >> > [main] ERROR
> > >> > > >> >
> > org.apache.kafka.clients.consumer.internals.ConsumerCoordinator -
> > >> > > Error
> > >> > > >> > ILLEGAL_GENERATION occurred while committing offsets
for
> group
> > >> > > >> > aaa-bbb-reader
> > >> > > >> > [main] WARN
> > >> > > >> org.apache.kafka.clients.consumer.internals.ConsumerCoordinator
> > >> > > >> > - Auto offset commit failed:
> > >> > > >> > [main] INFO
> > >> > > >> org.apache.kafka.clients.consumer.internals.AbstractCoordinator
> > >> > > >> > - Attempt to join group aaa-bbb-reader failed due
to unknown
> > >> member
> > >> > > id,
> > >> > > >> > resetting and retrying.
> > >> > > >> >
> > >> > > >> > And this in the broker log:
> > >> > > >> > [2015-11-25 15:41:01,542] INFO [GroupCoordinator
0]:
> Preparing
> > to
> > >> > > >> > restabilize group aaa-bbb-reader with old generation
1
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:01,544] INFO [GroupCoordinator
0]:
> > >> > > >> > Group aaa-bbb-reader generation 1 is dead and removed
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:13,474] INFO [GroupCoordinator
0]:
> Preparing
> > to
> > >> > > >> > restabilize group aaa-bbb-reader with old generation
0
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:13,475] INFO [GroupCoordinator
0]:
> Stabilized
> > >> > > >> > group aaa-bbb-reader generation 1
> > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:13,477] INFO [GroupCoordinator
0]:
> Assignment
> > >> > > received
> > >> > > >> > from leader for group aaa-bbb-reader for generation
1
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:43,478] INFO [GroupCoordinator
0]:
> Preparing
> > to
> > >> > > >> > restabilize group aaa-bbb-reader with old generation
1
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> > [2015-11-25 15:41:43,478] INFO [GroupCoordinator
0]:
> > >> > > >> > Group aaa-bbb-reader generation 1 is dead and removed
> > >> > > >> > (kafka.coordinator.GroupCoordinator)
> > >> > > >> >
> > >> > > >> > When this happens the kafka-consumer-groups describe
command
> > >> keeps
> > >> > > >> saying
> > >> > > >> > that the group no longer exists or is rebalancing.
What is
> > >> probably
> > >> > > even
> > >> > > >> > worse is that my consumers appears to be looping
constantly
> > >> through
> > >> > > >> > everything written to the topics!?
> > >> > > >> >
> > >> > > >> > Does anyone have any input on what might be happening?
> > >> > > >> >
> > >> > > >> > I'm running 0.9 locally on my laptop using one
Zookeeper and
> > one
> > >> > > broker,
> > >> > > >> > both using the configuration provided in the distribution.
I
> > >> have 13
> > >> > > >> topics
> > >> > > >> > with two partitions each and a replication factor
of 1. I run
> > one
> > >> > > >> producer
> > >> > > >> > and once consumer also on the same machine.
> > >> > > >> >
> > >> > > >> > --
> > >> > > >> > Martin Skøtt
> > >> > > >> >
> > >> > > >>
> > >> > > >>
> > >> > > >>
> > >> > > >> --
> > >> > > >> -- Guozhang
> > >> > > >>
> > >> > > >
> > >> > > >
> > >> > >
> > >> >
> > >>
> > >
> > >
> >
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message