[ https://issues.apache.org/jira/browse/KAFKA-6718?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16999090#comment-16999090 ]
Levani Kokhreidze commented on KAFKA-6718: ------------------------------------------ Thanks for the ideas [~vvcephei] will definitely look into it. My initial idea was around how Kubernetes [affinity rules|[https://kubernetes.io/docs/concepts/configuration/assign-pod-node/#affinity-and-anti-affinity]] work. It's similar to the elasticsearch mechanisms you have provided (tbh, I think most of the distributed systems share same idea behind rack awareness and standby tasks). I think with this task we have a possibility to let users specify something similar to affinity/anti-affinity rules based on some interface implementation which can have some default implementation out of the box (rack awareness maybe by default) but maybe can be extended to some other metrics? Like most probably you don't want to have standby task on the Kafka streams process were disk space is running low or CPU usage is high? Wdyt about this? From my experience working with Kafka streams and standby tasks, this is something i would really happy to have :) > Rack Aware Stand-by Task Assignment for Kafka Streams > ----------------------------------------------------- > > Key: KAFKA-6718 > URL: https://issues.apache.org/jira/browse/KAFKA-6718 > Project: Kafka > Issue Type: New Feature > Components: streams > Reporter: Deepak Goyal > Priority: Major > Labels: needs-kip > > |Machines in data centre are sometimes grouped in racks. Racks provide > isolation as each rack may be in a different physical location and has its > own power source. When tasks are properly replicated across racks, it > provides fault tolerance in that if a rack goes down, the remaining racks can > continue to serve traffic. > > This feature is already implemented at Kafka > [KIP-36|https://cwiki.apache.org/confluence/display/KAFKA/KIP-36+Rack+aware+replica+assignment] > but we needed similar for task assignments at Kafka Streams Application > layer. > > This features enables replica tasks to be assigned on different racks for > fault-tolerance. > NUM_STANDBY_REPLICAS = x > totalTasks = x+1 (replica + active) > # If there are no rackID provided: Cluster will behave rack-unaware > # If same rackId is given to all the nodes: Cluster will behave rack-unaware > # If (totalTasks <= number of racks), then Cluster will be rack aware i.e. > each replica task is each assigned to a different rack. > # Id (totalTasks > number of racks), then it will first assign tasks on > different racks, further tasks will be assigned to least loaded node, cluster > wide.| > We have added another config in StreamsConfig called "RACK_ID_CONFIG" which > helps StickyPartitionAssignor to assign tasks in such a way that no two > replica tasks are on same rack if possible. > Post that it also helps to maintain stickyness with-in the rack.| -- This message was sent by Atlassian Jira (v8.3.4#803005)