But then, yes, having a cleaner period that’s distinct from the idle timeout value is exactly the plan.
It doesn’t trust the client it just has a longer timeout than the client. The spike of load from Gatling is a factor of the number of concurrent users and the client timeout of 1 minute which is realistic (we typically timeout clients before servers - this is best practice to avoid stale connection errors / handling costs). So yes, with concurrent users exceeding 1M we do expect to have this number of idle connections. They will of course be distributed across a number of instances, in this case ELB instances. But in prod we typically offload client connections to CDN edge nodes.
I didn’t find a git issue relating to the cleaner period. Is there something I can track or do you have a release in mind for it?