Hi Gunnar, As far as the Connect behavior goes, sounds like https://issues.apache.org/jira/browse/KAFKA-12252. To verify, you can either set the 'connect.protocol' property of your Connect cluster to 'compatible' (easy mode) or cherry-pick the recently-merged fix for it ( https://github.com/apache/kafka/pull/10014) onto the version of Connect you're running, rebuild, and try again. Either step should prevent the symptoms you describe, and if not, a new Jira ticket is likely warranted.
Cheers, Chris On Mon, May 17, 2021 at 1:50 AM Gunnar Morling <gunnar.morl...@googlemail.com.invalid> wrote: > Hi, > > I was testing the early access preview of the new ZooKeeper-less mode and > noticed two things I wanted to bring up here. My testing scenario was a > cluster of three Kafka nodes in combined mode and a single Kafka Connect > node, all running via Docker Compose. > > * I stopped two of the Kafka nodes; the remaining node then was marked as > "leader" as per its metrics and in the new metadata shell; is this > expected? As one out of three nodes doesn't have a quorum, I'd rather have > expected to see that there isn't a leader in this situation. Indeed trying > to create a topic in this situation times out, so I reckon that the > remaining node doesn't actually act as the quorum leader. > * After restarting the two nodes, in most cases they, as well as the > Connect node, each would use up 100% CPU, like threads stuck in a spinning > loop. Is this issue already known, or should I log one in Jira? > > Thanks a lot for your help (and this great feature in general, really > exciting to see this), > > --Gunnar >
