Question

Your service processes long-poll/streaming requests that routinely stay open 90-120 seconds. You set a generous 180-second connection-drain (graceful termination) period so scale-in wouldn't drop work. Yet after every evening scale-in you still see a burst of mid-stream disconnects and a small number of partial writes to your ledger that later need reconciliation. The drain window is clearly longer than a request, the pods get SIGTERM and the LB stops sending new traffic, and CPU is fine. Why is in-flight work still being dropped, and how do you fix it?