You are viewing a plain text version of this content. The canonical link for it is here.
Posted to users@kafka.apache.org by "kafka@skianis.com" <ka...@skianis.com> on 2017/11/27 01:48:44 UTC

Consumer slow to start up when many partitions exist and the brokers are under load.

Hi,
I assume there might be some delay from the moment a consumer request metadata for the first time to the time the server replies which can get worse with the increase of topics/partitions. This delay gets even worse if the cluster is under decent load.
So say i have 150 topics and 16 partitions each. With the cluster under load, the consumer takes a reasonable amount to connect.Increase the topics to approx 400 (still 16 partitions each) and with the same load the consumer may take 6-7 minutes to connect.In both scenarios once connected, both consumers consume without major issues.
Nothing obvious in standard metrics such as CPU/Memory/Heap etc.

I am wondering 1) Is this expected, if yes, I can't think of why. The brokers are a bit busy but not that much2) has anybody experienced this3) Is there a recommendation on the number of topics/partition a consumer can handle? Does it matter if there is a group of consumers and not just one.
The above questions are mainly about the initial call to pull metadata. If the first call is slow then I am wondering if subsequent calls would be slow.The only other concern is that when there are many topics/partitions, it looks as if consumption rate actually decreases. Is this something expected or did someone observe similar behaviour.
Version 10.2.0,, cluster of 4 Kafkas, 3 Zookeepers
Any suggestions will be greatly appreciated. I will try to provide further data if needed.
Thanks,Panos