Skip to main content

Infrastructure and ownership

Lunar Orbit · Know who can act at each boundary

“Managed” does not mean “someone else guarantees Apollo works.” It means a provider accepts responsibility for a defined part of the system. The team still needs to know where that responsibility ends.

Suppose the booking Pods are ready but passengers cannot reach the airline. The cause could be public DNS, a load balancer, a Gateway configuration, a Service with no endpoints, or the application itself. Ownership tells the incident team who can inspect and change each layer.

Separate five kinds of responsibility​

For each component, answer five questions:

  1. Who chooses its configuration?
  2. Who operates and monitors it day to day?
  3. Who repairs or replaces it after failure?
  4. Who controls access to it?
  5. Who sees and controls its cost?

A managed control plane may be patched by the provider while the platform team still chooses cluster access, node capacity, add-ons, and upgrade timing. The application team still owns booking behavior, probes, requests, rollout safety, and data correctness.

Diagram CLD-02 — provider, platform, and application responsibilities meet in one request path.

Build an ownership record, not just a diagram​

An architecture box labelled “managed database” is too vague during an incident. Its record should identify the service and region, the team allowed to change it, the provider health signal, Apollo's application signal, the backup owner, and the escalation route.

ComponentConfigured byOperated byUseful evidence
Managed control planePlatform and providerProviderAPI availability plus provider status
Worker nodesPlatformSharedReady nodes, capacity, instance health
Booking DeploymentApplicationApplicationrollout status and successful booking
Managed databaseApplication and data teamsProvider and data teamquery success and restore test

Shared responsibility must be written as two concrete responsibilities. “Shared” by itself tells nobody who acts next.

Ownership changes the debugging path​

Start with passenger impact, then find the first broken boundary. If Kubernetes reports a load-balancer address but the provider resource is still provisioning, the platform or provider path owns the next action. If traffic reaches a ready booking Pod and the database call fails, the application and data paths own it.

Evidence and limits​

Record owners before an incident, including after-hours responsibility and the evidence each owner accepts. Provider status alone does not prove Apollo meets its availability objective, and a healthy application dashboard does not prove the provider can restore its data.