Every endpoint returns 200. The sequence is broken

A health endpoint tells you the process is running. This is about asserting that the endpoints doing real work still do it: that a POST creates the record, that the response body contains what a client depends on, and that a sequence of dependent calls still completes. Health checks are worth keeping; they are just the floor rather than the coverage.

Yes, including flows where the token has to be fetched fresh. A setup script runs before the request and can perform the OAuth exchange, sign a JWT, or pull a short-lived credential, storing the result for the check to use. Secrets live in encrypted environment variables rather than in the repo.

Write them as a single check that walks the chain: create, poll until the resource is ready, read it back the way a client would, then clean up. Each step is asserted and timed separately, so a chain that breaks at step three reports step three rather than a generic failure.

Give the chain a teardown step that deletes what it created, and where the API makes that awkward, use a dedicated tenant or account that your data pipeline filters out. Both are common; the important part is deciding which one before you schedule the check every minute.

Yes, with private locations. You run a lightweight agent inside your VPC, cluster, or data center, and checks assigned to it execute from within your network so traffic never leaves it. Internal APIs get the same assertions, thresholds, and history as public ones.

Generate it. Keep one config listing your services, their hosts, and their latency budgets, then loop over it to create the checks. Adding a service becomes a line in that file and a pull request, which is also how the coverage gap stops reappearing every quarter.

Separate degradation from failure and retry before alerting. Each check carries both a degraded threshold and a maximum response time, so an endpoint that answers correctly but slowly raises a degradation alert rather than a page, and the two can route to entirely different channels. Failed runs are retried with fixed, linear, or exponential backoff, from the same region or a different one, so a cold start or a network blip resolves on its own.

Give them a status page rather than a support queue. Status pages run on your own domain, public or private, and the services on them are wired to the endpoint checks you already have. An incident opens when a check fails, notifies subscribers, and resolves when it recovers, so consumers of your API get the same signal your on-call does without anyone writing an update.

Read the trace and let Rocky AI go first. Checks emit OpenTelemetry spans, so a failing request links to the backend span and the database call behind it, and those spans export onward to Grafana Tempo, Honeycomb, or Datadog. Rocky AI reads the response, the timing, and that trace together and names the service or third-party dependency at fault. On a forty-service estate, that first correct guess is most of the debugging.