Golden paths and self-service
Golden paths
A golden path (Spotify's term; Netflix says paved road) is the supported, opinionated way to do a common task: create a service, add a database, set up alerts. It's the easiest path and the one with the best support, but teams may leave it when they have a good reason, accepting more responsibility.
A good "create a new service" golden path, in one click or command:
- Creates a repository from a template with the app skeleton, Dockerfile, tests and README.
- Adds a CI pipeline that builds, tests, scans, signs and pushes an image.
- Creates the deployment manifests and registers the service with GitOps for dev, staging and production.
- Sets up the namespace, service account and IAM role, and a basic dashboard and alerts.
- Registers the service in the catalogue with its owner.
A developer goes from idea to a running "hello world" in production-like environments in under an hour instead of weeks, and every service starts compliant.
Golden paths need maintenance: when the template improves, existing services should benefit too. Use shared reusable workflows, Helm library charts, base images and versioned modules so updates flow out, rather than one-off copies that drift.
Self-service infrastructure
Developers need databases, buckets and queues. Options, from simplest to most abstract:
| Approach | How a developer asks | Trade-off |
|---|---|---|
| Terraform modules + pull requests | Add a module block (module "orders_db" { source = "platform/rds-postgres" … }) to a repo; CI plans, the platform team or policy approves |
Familiar and transparent; developers must know some Terraform |
| Templates in the portal | Fill in a form that opens that pull request for them | Friendlier; same machinery underneath |
| Crossplane | Apply a Kubernetes object such as kind: PostgresInstance; a Composition creates the RDS instance, subnet group, secret and alarms |
Kubernetes-native and GitOps-friendly; another control plane to run |
| AWS Controllers for Kubernetes (ACK) | Apply objects mapping directly to AWS resources | Less abstraction, AWS-only |
| AWS Service Catalog | Launch an approved CloudFormation product | AWS-native; less developer-friendly |
Whatever the mechanism, the platform encodes guardrails in the module or composition: encryption on, backups configured, tags applied, sizes limited by environment, network private. Developers choose a few meaningful options (size, engine version, environment); the platform decides the rest.
A Crossplane-style claim might look like:
apiVersion: platform.example.com/v1alpha1
kind: PostgresInstance
metadata:
name: orders-db
namespace: orders
spec:
size: small # maps to an instance class per environment
version: "16"
backupRetentionDays: 14
writeConnectionSecretToRef:
name: orders-db-conn
Abstractions: how much to hide
Too little abstraction leaves developers with all the complexity; too much hides things they need when debugging, and creates a platform only the platform team understands. Good rules:
- Expose intent (size, environment, criticality) rather than low-level knobs, but let people see what was created underneath.
- Provide escape hatches: a way to add a custom annotation or raw resource when the abstraction doesn't fit.
- Build the thinnest viable platform first — perhaps just documentation and a few templates — and add layers only when users need them.