kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Guozhang Wang <wangg...@gmail.com>
Subject Re: Repeating UNKNOWN_PRODUCER_ID errors for Kafka streams applications
Date Wed, 05 Jun 2019 23:38:43 GMT
Hello Pieter,

It's normal to see offset increasing by 2: that is because each message on
that partition is from a separate transaction which Streams EOS leverage
on, and whenever a a txn commits a commit marker, written as a special
message would be appended but not exposed to consumers. And hence that txn
marker will take one offset as well.

Your overridden configs look good to me, and it's indeed weird why it did
not mitigate your observed issue. Which Kafka version did you use?


Guozhang


On Wed, Jun 5, 2019 at 4:43 AM Pieter Hameete <pieter.hameete@blockbax.com>
wrote:

> Hi Guozhang,
>
> Some additional finding: it seems to only happen on Kakfa Streams
> repartition topics. We haven't seen this happening for any other topics so
> far.
>
> Best,
>
> Pieter
>
> -----Oorspronkelijk bericht-----
> Van: Pieter Hameete <pieter.hameete@blockbax.com>
> Verzonden: Wednesday, 5 June 2019 11:23
> Aan: users@kafka.apache.org
> Onderwerp: RE: Repeating UNKNOWN_PRODUCER_ID errors for Kafka streams
> applications
>
> Hi Guozhang,
>
> Thanks for your reply! I noticed my original mail went out twice by
> accident, sorry for that.
>
> We currently have a small variety of keys so not all partitions are
> 'actively used' indeed. The strange thing is though is that the errors
> occur for the partitions that actively receive records every few seconds. I
> have checked this using kafkacat to consume the specific partitions.
> Something I noticed was that for each received record the partition offset
> was 2 higher than the previous record, instead of the expected 1. Could
> that be due to the producers retrying (see warning logs in my original
> mail)?
>
> I've done the override for the configs in the repartition topics as
> follows, on one of the brokers:
>
> The values are taken from your KIP-443
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-443%3A+Return+to+default+segment.ms+and+segment.index.bytes+in+Streams+repartition+topics
>
> kafka-topics --zookeeper localhost:2181 --alter --topic
> event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
> --config segment.index.bytes=10485760 kafka-topics --zookeeper
> localhost:2181 --alter --topic
> event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
> --config segment.bytes= 52428800 kafka-topics --zookeeper localhost:2181
> --alter --topic
> event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
> --config segment.ms=604800000 kafka-topics --zookeeper localhost:2181
> --alter --topic
> event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
> --config retention.ms=-1
>
> Verifying afterwards:
>
> kafka-topics --zookeeper localhost:2181 --describe --topic
> event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
>
> Topic:event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition
>
>  PartitionCount:32       ReplicationFactor:3
>  Configs:segment.bytes=52428800,retention.ms
> =-1,segment.index.bytes=10485760,segment.ms
> =604800000,cleanup.policy=delete
>
> Is there anything that seems off to you? Or something else I can
> investigate further? We'd really like to nail this issue down. Especially
> because the cause seems different than the 'low traffic' cause in JIRA
> issue KAFKA-7190 as the partitions for which errors are thrown are
> receiving data.
>
> Best,
>
> Pieter
>
> -----Oorspronkelijk bericht-----
> Van: Guozhang Wang <wangguoz@gmail.com>
> Verzonden: Wednesday, 5 June 2019 02:23
> Aan: users@kafka.apache.org
> Onderwerp: Re: Repeating UNKNOWN_PRODUCER_ID errors for Kafka streams
> applications
>
> Hello Pieter,
>
> If you only have one record every few seconds that may be too small given
> you have at least 25 partitions (as I saw you have a xxx--repartition-24
> partition), which means that for a single partition, it may not see any
> records for a long time, and in this case you may need to override it to
> very large values. On the other hand, if you can reduce your num.partitions
> that may also help increasing the traffic per partition.
>
> Also could you show me how did you override the configs in the repartition
> topics?
>
>
> Guozhang
>
> On Tue, Jun 4, 2019 at 2:10 AM Pieter Hameete <pieter.hameete@blockbax.com
> >
> wrote:
>
> > Hello,
> >
> > Our Kafka streams applications are showing the following warning every
> > few seconds (on each of our 3 brokers, and on each of the 2 instances
> > of the streams application):
> >
> >
> > [Producer
> > clientId=event-rule-engine-dd71ae9b-523c-425d-a7c0-c62993315b30-Stream
> > Thread-1-1_24-producer, transactionalId=event-rule-engine-1_24]
> > Resetting sequence number of batch with current sequence 1 for
> > partition
> > event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition-24
> > to 0
> >
> >
> >
> > Followed by:
> >
> >
> >
> > [Producer
> > clientId=event-rule-engine-dd71ae9b-523c-425d-a7c0-c62993315b30-Stream
> > Thread-1-1_24-producer, transactionalId=event-rule-engine-1_24] Got
> > error produce response with correlation id 5902 on topic-partition
> > event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition-24
> > , retrying (2147483646 attempts left). Error: UNKNOWN_PRODUCER_ID
> >
> > The brokers are showing errors that look related:
> >
> >
> > Error processing append operation on partition
> > event-rule-engine-KSTREAM-REDUCE-STATE-STORE-0000000015-repartition-24
> > (kafka.server.ReplicaManager)
> >
> > org.apache.kafka.common.errors.UnknownProducerIdException: Found no
> > record of producerId=72 on the broker. It is possible that the last
> > message with the producerId=72 has been removed due to hitting the
> retention limit.
> >
> >
> >
> > We would expect the UNKNOWN_PRODUCER_ID error to occur once. After a
> > retry the record would be published on the partition and the
> > PRODUCER_ID would be known. However, this error keeps occurring every
> > few seconds. This is roughly at the same rate at which records are
> > produced on the input topics partitions, so it seems like it occurs for
> (nearly) every input record.
> >
> >
> >
> > The following JIRA issue:
> > https://issues.apache.org/jira/browse/KAFKA-7190
> > looks related. Except the Jira issue mentions ‘little traffic’, and I
> > am not sure if a message per every few seconds is regarded as little
> traffic.
> > Matthias mentions in the issue that a workaround seems to be to
> > increase topic configs `segment.bytes`, `segment.index.bytes`, and
> > `segment.ms` for the corresponding repartition topics. We’ve tried
> > manually overriding these configs for a relevant topic to the config
> > values in the linked pull request
> > (https://github.com/apache/kafka/pull/6511) but this did not result in
> the errors disappearing.
> >
> >
> >
> > Could anyone help us to figure out what is happening here, and why the
> > proposed fix for the above JIRA issue is not working in this case?
> >
> >
> >
> > Best,
> >
> >
> >
> > Pieter
> >
> >
>
> --
> -- Guozhang
>


-- 
-- Guozhang

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