When removing a server from a pool, you don't want to kill active requests. Connection draining (deregistration delay) handles this gracefully.
The load balancer stops sending new connections to the server. Existing connections continue until they complete or timeout.
Typical drain timeout: to seconds. After timeout, remaining connections are forcibly closed.
Connection draining is standard during deployments. Pull a server from the pool, drain it, deploy new code, health check, add it back. Users never see errors.