The recurring theme of failures where something simply stops happening maps well onto absence-based alerting. For the four days with no bids, a rate alert on the expected event (for example, bid transactions per hour dropping below a rolling baseline, or in Prometheus terms absent()/increase() over a window) would have caught it without needing any error to fire. The same pattern works as a heartbeat or dead-man's-switch check on the last successful operation, such as the last successful chart upgrade, which would have exposed the three silently failed upgrades. For the Ready-but-no-data-plane node, a synthetic pod-to-pod and pod-to-storage probe per node would turn the cross-reference you describe into a continuous signal. Did you consider adding a canary like that, or is the scheduled cross-check enough?