kafka-users mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Craig Swift <craig.sw...@returnpath.com.INVALID>
Subject MirrorMaker - Not consuming from all partitions
Date Wed, 09 Sep 2015 15:29:00 GMT
Hello,

Hope everyone is doing well. I was hoping to get some assistance with a
strange issue we're experiencing while using the MirrorMaker to pull data
down from an 8 node Kafka cluster in AWS into our data center. Both Kafka
clusters and the mirror are using version 0.8.1.1 with dedicated Zookeeper
clusters for each cluster respectively (running 3.4.5).

The problem we're seeing is that the mirror starts up and begins consuming
from the cluster on a specific topic. It correctly attaches to all 24
partitions for that topic - but inevitably there are a series of partitions
that either don't get read or are read at a very slow rate. Those
partitions are always associated with the same brokers. For example, all
partitions on broker 2 won't be read or all partitions on broker 2 and 4
won't be read. On restarting the mirror, these 'stuck' partitions may stay
the same or move. If they move the backlog is drained very quickly. If we
add more mirrors for additional capacity the same situation happens except
that each mirror has it's own set of stuck partitions. I've included the
mirror's configurations below along with samples from the logs.

1) The partition issue seems to happen when the mirror first starts up.
Once in a blue moon it reads from everything normally, but on restart it
can easily get back into this state.

2) We're fairly sure it isn't a processing/throughput issue. We can turn
the mirror off for a while, incur a large backlog of data, and when it is
enabled it chews through the data very quickly minus the handful of stuck
partitions.

3) We've looked at both the zookeeper and broker logs and there doesn't
seem to be anything out of the normal. We see the mirror connecting, there
are a few info messages about zookeeper nodes already existing, etc. No
specific errors.

4) We've enabled debugging on the mirror and we've noticed that during the
zk heartbeat/updates we're missing these messages for the 'stuck'
partitions:

[2015-09-08 18:38:12,157] DEBUG Reading reply sessionid:0x14f956bd57d21ee,
packet:: clientPath:null serverPath:null finished:false header:: 357,5
 replyHeader:: 357,8597251893,0  request::
'/consumers/mirror-kafkablk-kafka-gold-east-to-kafkablk-den/offsets/MessageHeadersBody/5,#34303537353838,-1
 response::
s{4295371756,8597251893,1439969185754,1441759092134,19500,0,0,0,7,0,4295371756}
 (org.apache.zookeeper.ClientCnxn)

i.e. we see this message for all the processing partitions, but never for
the stuck ones. There are no errors in the log prior to this though, and
once in a great while we might see a log entry for one of the stuck
partitions.

5) We've checked latency/response time with zookeeper from the brokers and
the mirror and it appears fine.

Mirror consumer config:
group.id=mirror-kafkablk-kafka-gold-east-to-kafkablk-den
consumer.id=mirror-kafkablk-mirror00-den-kafka-gold-east-to-kafkablk-den
zookeeper.connect=zk.strange.dev.net:2181
fetch.message.max.bytes=15728640
socket.receive.buffer.bytes=64000000
socket.timeout.ms=60000
zookeeper.connection.timeout.ms=60000
zookeeper.session.timeout.ms=30000
zookeeper.sync.time.ms=4000
auto.offset.reset=smallest
auto.commit.interval.ms=20000

Mirror producer config:
client.id=mirror-kafkablk-mirror00-den-kafka-gold-east-to-kafkablk-den
metadata.broker.list=kafka00.lan.strange.dev.net:9092,
kafka01.lan.strange.dev.net:9092,kafka02.lan.strange.dev.net:9092,
kafka03.lan.strange.dev.net:9092,kafka04.lan.strange.dev.net:9092
request.required.acks=1
producer.type=async
request.timeout.ms=20000
retry.backoff.ms=1000
message.send.max.retries=6
serializer.class=kafka.serializer.DefaultEncoder
send.buffer.bytes=134217728
compression.codec=gzip

Mirror startup settings:
--num.streams 2 --num.producers 4

Any thoughts/suggestions would be very helpful. At this point we're running
out of things to try.


Craig J. Swift
Software Engineer

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