When a health check fails, the load balancer must respond quickly. Failover is the process of routing traffic away from failed servers.
Immediate failover removes the server from the pool instantly. Requests in flight may fail, but new requests go elsewhere.
Graceful failover drains existing connections. No new requests go to the server, but current ones complete. This works for planned maintenance.
After failover, the remaining servers handle increased load. Size your cluster so you can lose servers without overloading survivors.