ADR 0003: Deployment Profiles

Status

Accepted for 0.1.

Decision

Graviton documents three distinct profiles instead of presenting one vague production topology:

• embedded runtime inside one application
• single-node filesystem service with backend-wide file-lock coordination and a conservative Recreate rollout
• shared S3 plus PostgreSQL service with namespaced advisory-lock coordination, still requiring operator qualification for concurrent processes and rolling upgrades

ReplicatedBlockStore is an application-level reliability primitive. It is not automatically enabled by the packaged server. The Kubernetes example therefore uses one replica with a ReadWriteOnce volume and does not imply high availability.

Consequences

• built-in filesystem clients coordinate ordinary operations and maintenance across processes that share working file-lock semantics
• a local filesystem deployment still cannot scale by increasing replicas against a node-local volume
• an S3 plus PostgreSQL deployment has shared state but must still prove failover and rollout behavior in its real environment
• readiness checks verify required backends, while liveness only verifies the process
• deployment manifests remain examples whose ingress, egress, resources, storage class, secrets, and retention must be reviewed by the operator