Deployment¶

Environments¶

Each environment has its own Terraform state, AlloyDB instances, GCS buckets, and KMS keys. Environments are symmetric — same Helm chart, different values.

Compute¶

• GKE Autopilot for all services.
• Dedicated node pools for AI workers (high memory) and BIM workers (high CPU + optional GPU).
• GPU node pool is provisioned on-demand; default is managed Vertex AI calls, not self-hosted inference.

Release pipeline¶

flowchart LR
  PR[PR merged] --> CI[CI<br/>build · test · scan]
  CI --> IMG[Push image<br/>Artifact Registry]
  IMG --> DEV[ArgoCD<br/>dev rollout]
  DEV --> STG[Promote<br/>staging]
  STG --> UAT[Promote<br/>uat]
  UAT --> PROD[Canary<br/>prod]
  PROD --> WATCH[SLO burn-rate alert]
  WATCH -->|breach| RB[Automated rollback]
  WATCH -->|healthy| FULL[Full rollout]

Infrastructure as code¶

Every service ships:

• A Helm chart with values per environment.
• A Terraform module defining its AlloyDB instance, GCS buckets, KMS keys, and IAM.
• A .argocd/ manifest wiring it into the rollout.

Progressive delivery¶

• Canary rollout in prod — new version serves a small percentage of traffic first.
• SLO burn-rate alerts wired to the rollout controller — a breach triggers automated rollback before it reaches the full fleet.
• Manual approval gate for any schema migration that isn't additive.

Secrets¶

• Google Secret Manager only. No secrets in Kustomize, Helm values, or env files.
• CMEK on AlloyDB and GCS.
• mTLS between services via Anthos Service Mesh (or Istio).

What each environment is for¶

• dev — engineers push freely; cluster is cheap and can be torn down.
• staging — integration target; contract tests run here continuously.
• uat — customer sign-off per milestone; stable for the duration of a UAT cycle.
• prod — canary-first, SLO-gated.