When both sites fail, Global Accelerator randomly forwards requests to each site. This may be causing problems since lb-check is not part of Kubernetes probes and therefore requests can still reach Keycloak and since Keycloak itself may not be completely down, the requests can still be handled in some undefined way.