What load balancers taught me about deploys
Drain, health checks, and why cutovers stay boring when traffic already knows where to go.
A load balancer is not just “spread requests.” The useful lesson is drain: take a node out of rotation, wait for in-flight work to finish, then ship.
Health checks that only ping /health lie when the process is up but the dependency path is dead. Prefer checks that exercise the critical hop you actually need for the deploy.
I keep this as a learn, not a playbook — the playbook lives in the runbook; the learn is why we bother writing one.