Detail I forgot to mention is that I am using EOS, so the streams application was consistently getting this error causing it to restart that task, which obviously would bottle neck everything. Once I increased the threads on the brokers the epoch error subsided until the volume increased.
Right now the streams application has 36 threads per box (5 boxes w/ 48 threads) and when running normally it is keeping up, but once this epoch error starts it causes a cascade of slowness. The CPU is not even being fully utilized due to this error happening every minute or so. Perhaps some other tuning is needed too, but I'm lost what to look at. I do have JMX connection to the broker and streams applications if there's any useful information I should be looking at. -----Original Message----- From: Sophie Blee-Goldman <sop...@confluent.io.INVALID> Sent: Thursday, October 27, 2022 11:22 PM To: users@kafka.apache.org Subject: Re: Kafka Streams - Producer attempted to produce with an old epoch. CAUTION: This email originated from outside of ShopHQ. Do not click links or open attachments unless you recognize the sender and know the content is safe! I'm not one of the real experts on the Producer and even further from one with broker performance, so someone else may need to chime in for that, but I did have a few questions: What specifically are you unsatisfied with w.r.t the performance? Are you hoping for a higher throughput of your Streams app's output, or is there something about the brokers? I'm curious why you started with increasing the broker threads, especially if the perf issue/bottleneck is with the Streams app's processing (but maybe it is not). I would imagine that throwing more and more threads at the machine could even make things worse, it definitely will if the thread count gets high enough though it's hard to say where/when it might start to decline. Point is, if the brokers are eating up all the cpu time with their own threads then the clients embedded in Streams may be getting starved out at times, causing that StreamThread/consumer to drop out of the group and resulting in the producer getting fenced. Or it could be blocking i/o for rocksdb and leading to write stalls, which could similarly get that StreamThread kicked from the consumer group (if the application has state, especially if quite a lot). How many StreamThreads did you give the app btw? On Thu, Oct 27, 2022 at 8:01 PM Andrew Muraco <amur...@imds.tv> wrote: > Hi, > I have a kafka streams application deployed on 5 nodes and with full > traffic I am getting the error message: > > org.apache.kafka.common.errors.InvalidProducerEpochException: Producer > attempted to produce with an old epoch. > Written offsets would not be recorded and no more records would be > sent since the producer is fenced, indicating the task may be migrated > out > > I have 5 x 24 CPU/48 core machines with 128gb of ram. These machines > are the kafka brokers with 2x1TB disks for kafka logs and also running > the kafka Streams application. > 2x replication factor on topic, topic is producing about 250k per second. > I have 2 aggregations in the topology to 2 output topics, the final > output topics are in the 10s of k range per second. > > I'm assuming I have a bottleneck somewhere, I increased the broker > thread counts and observed that this frequency of this error reduced, > but it's still happening. > Here's the broker configuration I'm using now, but I might be > overshooting some of these values. > > num.network.threads=48 > num.io.threads=48 > socket.send.buffer.bytes=512000 > socket.receive.buffer.bytes=512000 > replica.socket.receive.buffer.bytes=1024000 > socket.request.max.bytes=10485760 > num.replica.fetchers=48 > log.cleaner.threads=48 > queued.max.requests=48000 > > I can't find good documentation on the effect of broker configuration > on performance. >