GSLB at the edge is a DNS steering contract first, not a per-packet dataplane. Pin health-check paths that reflect real readiness (app plus dependency), keep pool membership changes slower than probe intervals, and decide failover versus weighted steering before you ship — sudden full cutovers hide capacity cliffs that weighted shifts would have shown. Terminate TLS once at the edge or once at origin; dual termination without a clear SNI story produces intermittent handshake failures that look like random upstream errors.
Ops prove-out on a canary pool: steer a small traffic share, then compare status codes, RTT, and origin error rates against the control pool under the same synthetic mix. Fail one origin in the canary and confirm DNS answers stop advertising it within the expected TTL window without cascading the healthy pool. If only one region flaps, check probe egress and geo answer scope before blaming the load-balancer algorithm itself.