Hi Zk Users, We're seeing a moderate decrease (~20% drop) in write-throughput when we upgrade from a Java 8 JVM to a Java 17 JVM. I'm wondering if anyone else has seen this, and if you know any JVM options that might mitigate it?
Background: I'm responsible for the Zookeeper deployments here at Pinterest. We are rolling out a JVM upgrade and it's running smoothly for most clusters, but a few of our clusters are run at the limits of the write throughput ZK can achieve and are seeing a regression. Specifically, where we had seen up to 21k transactions per second before we are now seeing 17k. That's in a cluster with 5 participants, and no observers. We're running Zookeeper 3.9.3. We're using G1 GC on both versions. CPU usage on the leader went down slightly so I suspect something is causing the critical path to spend more time blocked. I figured before we spent too much time debugging I'd ask around and see if anyone has solved this already. Thank you! Brett Folkins
