Slow start and connection draining
termnetworkingcloud
Two behaviours for changing a pool safely: ramping a newly added member's share of traffic upward over time, and letting an outgoing member finish existing connections before removal.
A fresh instance has cold caches, an empty connection pool and an unwarmed runtime, so giving it a full share immediately makes it slow, fail its own health check, get ejected and oscillate. Draining has the mirror problem: too short cuts live requests, never-completing blocks maintenance. Both are usually discovered during a release rather than designed, and both matter more than the balancing algorithm - which is why a capacity plan for N servers should be tested at N-1, since that is the state a failure actually puts you in.
Also known as: ramp-up, warm-up, graceful shutdown, drain