Deployment model

dyson-deployment is the GitOps source of truth for running Dyson. Application code, Dockerfiles, tests, and image publishing stay in dyson; Kubernetes, platform services, and environment desired state stay in dyson-deployment.

Release flow

Dyson commit
  -> test and build component images
  -> publish signed immutable ECR digests
  -> update an environment overlay
  -> Argo CD reconciles the selected cluster

Components are promoted independently. An API change does not require changing the strategy, market-data, or control-panel digest. Environments select exact digests rather than rebuilding or consuming latest.

Configuration ownership

Path	Owns
apps/	Reusable workload, service, storage, and service-account bases
environments/	Image digests, replicas, deployment IDs, namespaces, and secret references
clusters/	Argo CD child applications, destinations, sync waves, and sync policy
bootstrap/	Argo CD project and root applications
platform/	NATS, observability, alerting, and single-node data services

Secrets are supplied through Doppler-backed or deliberately created Kubernetes Secrets. Secret values do not belong in Git.

Deployment shapes

Single-node staging

The staging-only K3s topology defines Argo CD, NATS, a Mongo-compatible store, InfluxDB, Grafana, and Dyson workloads on one host. It is intentionally non-HA, uses local persistent volumes, and is suitable for integration testing only.

The committed staging overlay declares one replica each for market data, the Strategy API, and the control panel. Order-executing strategy workloads remain scaled to zero until explicitly activated.

Cluster environments

Dev and staging target the non-production cluster; production targets its own cluster. Shared platform definitions provide NATS and the monitoring stack. Most non-production applications auto-sync, while secrets require deliberate initial synchronization. Production applications require manual sync.

Current workload coverage

Reusable manifests exist for market data, the Strategy API, the control panel, and toxic-taker. Dyson also builds kuru-quoter, but this deployment repository does not yet contain its workload manifest. No committed environment currently enables an order-executing strategy.

Activation boundary

Before increasing an execution-capable replica count, confirm that the overlay pins a real digest, required secrets and data services are reachable, MongoDB contains the selected deployment/version documents, NATS subjects align with the publishers and strategy, and the strategy remains intentionally stopped until the operator starts it.