e73ffe648e
tried to overcome issues with slowdown around the 4500 connection mark by using the supervisor pattern to terminate the threads on the backend. this seems to get us further, but we still hit a slowdown and the allocations die out before we hit any serious usage