Self-service on a catalog-driven developer platform
Client context
A state government-owned workers’ compensation insurer. The engineering estate spans a large end-of-life Java and Oracle claims monolith and a modern catalog-driven internal developer platform on AWS EKS, deployed through a distinctive rainbow release model.
The problem
The platform’s core pipeline existed (service catalog YAML rendered by KubeVela CUE templates, synced by Argo CD to EKS, with a Kafka-driven GitOps choreographer auto-bumping image versions), but self-service had gaps: creating a GitHub repository, getting a development database, or seeing what a catalog change would actually do to infrastructure still required platform-team involvement. CI also ran on aging self-hosted runners that needed migration and security hardening.
What was built
- Self-service repository provisioning: a catalog entry plus PR-approval flow through the developer portal that creates and manages GitHub repositories, with Slack notifications and PR commentary.
- Self-service developer databases: engineers raise a PR to get an Oracle RDS dev database provisioned from nightly snapshots, with automated nightly shutdown for cost saving and ephemeral databases for CI.
- Catalog assembly processor: a toolkit that turns catalog YAML into Terraform HCL and KubeVela-rendered Kubernetes manifests, and posts terraform plan and kubectl diff output as PR commentary, so reviewers see the real infrastructure change before merge.
- GitHub runner migration: self-hosted runners rebuilt on Actions Runner Controller scale-sets on EKS, with GitHub OIDC mapped to scoped AWS IAM roles, scale-to-zero, and observability.
- Platform plumbing: GitOps choreographer enhancements (Micronaut, Kafka, DynamoDB), KubeVela plugin versioning for safe catalog schema evolution, and a CLI (SSO auth, kubeconfig, registry login, dashboards) delivered as a documented handover tool on the final day.
Outcomes
- Self service provisioning was not possible before the catalog existed. It became the access path that let developers build against the uplifted legacy platform at all, and the wider uplift program depended on it.
- The CI pipeline for the legacy claims and policy system took developers from zero feedback on changes to fast feedback against a production like database, reducing failed releases.
- The runner migration to Actions Runner Controller replaced webhook based scaling with long poll scale sets, adding scale to zero and observability the old model could not offer.