I can reproduce this on 0.9.0.0. The error log is [2016-01-28 16:12:32,840] ERROR [Replica Manager on Broker 1]: Error processing fetch operation on partition [ad-events,1] offset 75510318 (kafka.server.ReplicaManager)
I also print the sent offset from producer time partition offset 16:12:32.840 1 75510318 It seems the offset is produced and consumed at the same time. On Wed, Jan 27, 2016 at 10:18 PM, Robert Metzger <rmetz...@apache.org> wrote: > Yes, I've asked the user to test with the 0.9.0.0 release (I saw Gwen's > comment in KAFKA-725). > > I have a potentially related question: Is it an issue that both Flink and > Gearpump* are not committing their offsets through the SimpleConsumer API? > Flink is directly committing the offsets into ZK (and maintaining them > internally). > I'm asking because I see some log statements like this: > WARN Partition [WordCount,22] on broker 0: No checkpointed highwatermark is > found for partition [WordCount,22] (kafka.cluster.Partition) > > Reading from the Kafka server's code, it doesn't seem that the > commitOffsets() request updates any internal state (at least for > version=0). > > *I didn't find any call to commitOffsets() in their code. > > > On Wed, Jan 27, 2016 at 2:32 PM, Ismael Juma <mli...@juma.me.uk> wrote: > > > Hi Manu and Robert, > > > > It would help to know if this still happens in trunk or the 0.9.0 branch. > > > > Ismael > > On 27 Jan 2016 13:05, "Robert Metzger" <rmetz...@apache.org> wrote: > > > > > Hi Manu, > > > > > > in the streaming-benchmark, are seeing the issue only when reading with > > > Gearpump, or is it triggered by a different processing framework as > well? > > > > > > I'm asking because there is a Flink user who is using Kafka 0.8.2.1 as > > well > > > who's reporting a very similar issue on SO: > > > > > > > > > http://stackoverflow.com/questions/34982483/flink-streaming-job-switched-to-failed-status/34987963 > > > . > > > His issue is also only present under load. > > > > > > > > > > > > > > > On Thu, Jan 21, 2016 at 2:28 AM, Manu Zhang <owenzhang1...@gmail.com> > > > wrote: > > > > > > > Hi, > > > > > > > > Any suggestions for this issue or do I need to provide more > > information ? > > > > Any links I can refer to would be also very helpful. > > > > > > > > Thanks, > > > > Manu Zhang > > > > > > > > > > > > On Tue, Jan 19, 2016 at 8:41 PM, Manu Zhang <owenzhang1...@gmail.com > > > > > > wrote: > > > > > > > > > Hi all, > > > > > > > > > > Is KAFKA-725 Broker Exception: Attempt to read with a maximum > offset > > > less > > > > > than start offset <https://issues.apache.org/jira/browse/KAFKA-725 > > > > > > still > > > > > valid ? We are seeing a similar issue when we are carrying out the > > > > yahoo's > > > > > streaming-benchmarks < > https://github.com/yahoo/streaming-benchmarks> > > > on > > > > a > > > > > 4-node cluster. Our issue id is > > > > > https://github.com/gearpump/gearpump/issues/1872. > > > > > > > > > > We are using Kafka scala-2.10-0.8.2.1. 4 brokers are installed on 4 > > > nodes > > > > > with Zookeeper on 3 of them. On each node, 4 producers produce data > > to > > > a > > > > > Kafka topic with 4 partitions and 1 replica. Each producer has a > > > > throughput > > > > > of 17K messages/s. 4 consumers are distributed (not necessarily > > evenly) > > > > > across the cluster and consume from Kafka as fast as possible. > > > > > > > > > > I tried logging the produced offsets (with callback in send) and > > found > > > > > that the "start offset" already existed when the consumer failed > with > > > the > > > > > fetch exception. > > > > > > > > > > This happened only when producers are producing at high throughput. > > > > > > > > > > Any ideas would be much appreciated. > > > > > > > > > > Thanks, > > > > > Manu Zhang > > > > > > > > > > > > > > >
