Design a health-check aggregator that continuously probes a large set of services and dependencies from several vantage points, combines the raw probe results into a per-service and per-dependency health verdict, and serves that verdict to dashboards and to automation that gates traffic. The hard part is telling a real outage apart from a single flaky prober, and doing it within seconds.

Operating context. The system probes about 30,000 target endpoints (HTTP, gRPC, TCP) plus dependency edges, at per-target intervals of 5-30 s, from 5 or more regions. A service usually has many instances, and a target is probed from multiple vantage points, so the verdict is a rollup with quorum. Automation queries the current verdict synchronously to decide whether to route to a target, so false positives are expensive.

Out of scope. Remediation and auto-healing actions, the target's own health-endpoint implementation, alert notification routing, and load-balancing or capacity decisions. Assume those consume your verdict or live elsewhere.

What to produce. A high-level architecture covering: how probers are distributed across regions and assigned targets; how many-probers-per-target and many-instances-per-service results are aggregated with quorum; the health state model with hysteresis; storing and serving the current verdict plus a short history; and how you distinguish a prober-side network partition from a true target failure. Sketch the components and the flow; we will probe specifics in checkpoints.