@rohit I had a look at onyx-kafka and it turns out that when we switched Kafka clients, we ended up adding poll timeout, while still using the backoff. This will mean we're idling unnecessarily, though I'm a bit surprised your throughout is dropping so much as a result

@lucasbradstreet: that idling behaviour looks like what I was seeing. I'll create a project tomorrow with minimal moving parts so as to get a deeper understanding of what's happening. Thanks a lot!