Saltar al contenido principal

ADR 004 — Abstract InfrastructureProvider ABC for workload runtime ops

Context​

AlphaSwarm's existing IaC story is Terraform-first (AGENTS hard rule 42): every state-mutating cluster operation goes through alphaswarm/terraform/runtime.py::TerraformRuntime. That guarantee is great for provisioning (create cluster, create namespace, apply RBAC, register Auth0 tenant) but it's an awkward fit for live workload operations — restarting a pod, scaling a Deployment, exec-ing a shell, tailing logs — which today incur a full terraform plan + apply round trip and write to terraform_runs even though no IaC actually changed.

The refactor introduces the alphaswarm_controller micro-project that needs to support five backends (docker_compose, kubernetes, AWS, Azure, GCP). Two paths were considered:

  1. Translate every workload op into Terraform — every restart becomes a Terraform null_resource + provisioner. Preserves the rule 42 ledger as a single source of truth, but turns Terraform into a glorified kubectl wrapper.
  2. Introduce a sibling abstraction — InfrastructureProvider ABC with five implementations, each calling its backend's native SDK (kubernetes-client, docker SDK, boto3, azure-mgmt, google-cloud-run). Terraform stays for provisioning only.

Decision​

Adopt path 2: an abstract InfrastructureProvider ABC for runtime workload operations. Specifically:

class InfrastructureProvider(ABC):
    @abstractmethod
    async def start(self, spec: DeploymentSpec) -> DeploymentStatus: ...

    @abstractmethod
    async def stop(self, service_id: str) -> DeploymentStatus: ...

    @abstractmethod
    async def scale(self, service_id: str, replicas: int) -> DeploymentStatus: ...

    @abstractmethod
    async def status(self, service_id: str) -> DeploymentStatus: ...

    @abstractmethod
    async def apply_config(self, service_id: str, config: dict) -> bool: ...

    @abstractmethod
    async def stream_metrics(self, service_id: str): ...  # async generator

Five concrete providers live under alphaswarm_controller/src/alphaswarm_controller/providers/:

  • docker_compose.py — docker Python SDK + docker compose subprocess for multi-container profiles
  • kubernetes.py — kubernetes-client/python (in-cluster + kubeconfig); Deployment apply, scale-to-0, ConfigMap patch, Metrics Server query
  • aws.py — boto3; EKS delegates to kubernetes.py; ECS/Fargate via update_service; config sync via SSM Parameter Store
  • azure.py — azure-mgmt; AKS delegates to kubernetes.py; ACI via container groups; config sync via App Configuration / Key Vault
  • gcp.py — google-cloud SDKs; GKE delegates to kubernetes.py; Cloud Run via revision updates; config sync via Secret Manager

Each provider:

  • Reads credentials from env vars only (alphaswarm_core.credentials.CredentialResolver).
  • Translates DeploymentSpec to its backend's native API.
  • Returns a normalised DeploymentStatus.
  • Maps backend-specific exceptions to structured {status, data, error} envelopes.

Amendment to AGENTS hard rule 42 (this PR)​

Rule 42 changes from "all Terraform IaC lifecycle actions go through TerraformRuntime" to:

  1. All Terraform IaC PROVISIONING actions go through alphaswarm/terraform/runtime.py::TerraformRuntime. Cluster bootstrap, IAM, Auth0 tenant, namespaces, secrets, network policies, and Ingress class registration are all "provisioning". The terraform_runs ledger, the terraform_stack_spec_versions hash-lock, the kill-switch hook (/terraform/halt), and OPA policy enforcement all depend on it.

A new rule 45 covers the workload ops side:

  1. All runtime workload operations go through alphaswarm_controller.InfrastructureProvider (via WorkloadRuntime). Start, stop, scale, restart, exec, log-tail, and apply_config are workload ops. They never reach for Terraform. A new workload_runs ledger row is created per mutating action with full audit context (user_id, action, target, provider, timestamp) BEFORE the provider call executes.

Consequences​

Positive

  • Restart latency drops from ~30 s (Terraform plan + apply) to ~200 ms (kubectl scale).
  • The five providers are fully independent — each can be implemented + tested in parallel by an orchestrate fan-out (see plan §8.2).
  • Terraform stays clean for IaC provisioning and immutable audit trails. The terraform_runs ledger remains the source of truth for "what infrastructure exists".
  • The alphaswarm_controller micro-project becomes a thin, testable layer with mocked SDKs in CI.
  • Hard rule 27 (IdentityProvider), 28 (KubernetesAdapter), and the new ABC all follow the same self-registering metaclass pattern — consistent across the codebase.

Negative

  • Two separate audit ledgers (terraform_runs + workload_runs) instead of one. Documented in alphaswarm_docs/docs/how-to/operations/incident-response.md.
  • The five providers each take their own credential chain. Mitigated by CredentialResolver so service code never sees raw env vars.
  • Provisioning vs runtime boundary is a soft line — adding a new namespace is provisioning, but auto-creating a per-tenant namespace at user signup is workload-ish. Each new operation requires an explicit choice; ADR 005 includes a decision tree.

Alternatives considered​

  • Translate every op into Terraform — rejected. Operational cost of running terraform apply on every pod restart is prohibitive (~30 s p99), and Terraform's lock semantics serialise unrelated ops on the same workspace.
  • Use Crossplane — investigated; rejected for now. Crossplane is excellent for declarative cloud APIs but adds a CRD layer and operator dependency for marginal value over the five-provider Python ABC. Revisit when AlphaSwarm exceeds five backends.
  • Use Pulumi instead of Terraform — out of scope. The existing TerraformRuntime works and is hash-locked; replacing it is a separate ADR.

Implementation references​

  • ABC: alphaswarm_controller/src/alphaswarm_controller/providers/base.py
  • Five providers: alphaswarm_controller/src/alphaswarm_controller/providers/{docker_compose,kubernetes,aws,azure,gcp}.py
  • Workload ledger model: alphaswarm/persistence/models_workload.py (new in this PR)
  • Telemetry streaming: alphaswarm_controller/src/alphaswarm_controller/services/telemetry.py
  • AGENTS rule 45: AGENTS.md (this PR)