Skip to content

Platform engineering · GitOps · CI/CD · 7 min read ·

Golden paths that engineers actually take

A platform team's output is not the platform; it is the paths other teams take through it. How to build paved roads that are faster than the alternative, keep them honest with the people who use them, and avoid turning them into mandates.

Platform teams fail in one of two ways. Either they build infrastructure nobody asked for and wonder why the product teams keep rolling their own, or they build a mandate, enforce it, and become the department everyone routes around. The thing that works sits between: a small number of paved roads that are so much easier than the alternative that teams take them by choice, and that are maintained as products with users.

A golden path is a product

The path for "I have a new service and I want it running in production" has a user, a job to be done and a time to complete. Measure it. If a new service takes a week to reach production on the paved road, the road is the problem, and no amount of advocacy will change that. The platform team's backlog is prioritised by what the users of the paths are waiting for, and the team talks to those users every week, the way a product team would.

Start with one path, the one most teams need most often, and make it excellent. A second path is justified when the first is quiet.

What the first path usually contains

For a team shipping a service, the road from a repository to a running, observed, secured workload:

  • A template that creates the repository with the pipeline, the deployment manifests, the dashboard and the alerts already wired. The team edits code, not plumbing.
  • A pipeline that builds, scans, signs and publishes without configuration, and that the team cannot (and does not want to) bypass, because it is the only route to the registry.
  • Promotion by Git, so moving a build from staging to production is a merge the team controls, recorded, reversible, and gated by a verification the platform runs.
  • Defaults that are safe: resource requests, pod security, network policy, a service account with no privileges, structured logs with the standard labels. All overridable, none forgotten.
  • Observability for free: the dashboard and the burn-rate alerts exist the moment the service does, keyed on labels the template set.
  • A place to ask, with a response time the platform team commits to.

Everything on the list is something a product team could build themselves; the point is that none of them should have to, and that when every service is built the same way the platform can change all of them at once.

Keep it honest

A paved road drifts from reality the moment the platform team stops walking it. Three habits prevent that:

Dogfood. The platform's own services ship through the same path. If it is painful, the platform team feels it first.

Make the escape hatch visible, not forbidden. A team with a real need the path does not meet should be able to step off it, say why, and be heard. The reasons are the roadmap. A path that cannot be left is a mandate, and mandates get routed around in ways you will not see until the audit.

Publish the service level. How long a new repository takes to reach production, how long a question waits for an answer, how often the path breaks. If the platform team would not accept those numbers from a vendor, the users will not accept them from the platform team.

Guardrails, not gates

The controls on the path (scanning, signing, admission, policy) should be invisible when the team is doing the right thing and immediate when it is not. A gate that stops a merge for a human review of something a policy could decide is a tax on every team to catch a mistake a few would make. Write the policy, enforce it in the pipeline and the cluster, and spend the human review on design.

Scaling the team that builds it

Platform engineering is a product organisation, and it is staffed like one: an owner who talks to users, engineers who build, and a rotation that supports. It is not the infrastructure team renamed. The most common failure is a platform team that is also the on-call for every workload, which makes the paved road the thing nobody has time to pave.

How you know it is working

  • Most new services are on the path within a day, without a ticket.
  • The teams that stepped off it can say why, and the reasons are shrinking.
  • An estate-wide change (a base image, a logging format, a policy) ships through one pull request to the template and lands everywhere on the next build.
  • The platform team's backlog is full of user requests rather than infrastructure ideas.

The delivery path on our own platform is built this way: one template, one pipeline, promotion by Git through verified rings, and the policies enforced where they can be. A platform engineering engagement typically starts by measuring the current path end to end, because the first week of the engagement is usually spent finding out where the week goes.