constellation/rfc/014-terraform-provider.md

5.2 KiB

RFC 014: Constellation Terraform Provider

The Constellation Terraform Provider allows its user to manage the full lifecycle of a Constellation cluster -- namely initialization and updates (constellation apply) -- via Terraform.

Design Goals

  • The User needs to be able to perform initialization and upgrades of a Cluster through the provider. Deletion is covered by deleting the underlying infrastructure through Terraform.
  • Infrastructure provisioning should explicitly not be performed by the Constellation Terraform provider.
  • The User needs to receive some sort of Access Token to the Cluster (e.g. a Kubeconfig, whether long- or short-lived) when applying a configuration containing the provider and the resources listed below.

Terraform Configuration Layout for the Constellation Resource

This resembles an examplary configuration of a Constellation cluster through Terraform. While the exact naming and struct layout of the individual components may change until the implementation, the structure of the overall configuration looks like this:

terraform {
  required_providers {
    constellation = {
      source  = "tbd/constellation"
      version = "2.13.0"
    }
  }
}

provider "constellation" { }

resource "constellation_cluster" "foo" {
    uid = "bar"
    name = "baz"
    image = data.constellation_image.ref # or provide manually crafted values
    kubernetes_version = "v1.27.6"
    debug = false
    init_endpoint = "10.10.10.10" # should use public ip of LB resource, ideally also provisioned through TF
    kubernetes_api_endpoint = "10.10.10.10" # should use public ip of LB resource, ideally also provisioned through TF
    constellation_microservice_version = "v2.13.0" # optional value, set to provider version by default.
    extra_microservices = {
        csi_driver = true
        # + more
        # possiblly also constellation microservices with version and maybe service options,
        # which would make constellation_microservice_version obsolete.
        # exact API TBD
    }
    master_secret = "foo" # updating this would force recreation of the cluster
    init_secret = "bar" # maybe derive from master_secret, updating this would force recreation of the cluster
    network_config = {
        # TBD
        # should contain CIDRs for pod network, service cidr, node network... for Cilium
        # the aforementioned values might be outputs of infrastructure that is also provisioned
        # through Terraform, such as a VPC.
        # and in-cluster Kubernetes API endpoint, e.g. for Kubelets
    }
    attestation = data.constellation_attestation.attestation # or provide manually crafted values
}

# constellation_cluster provides:
# constellation_cluster.foo.kubeconfig
# constellation_cluster.foo.owner_id
# constellation_cluster.foo.cluster_id


data "constellation_attestation" "foo" {
    attestation_variant = "GCP-SEV-ES"
    image_version = "v2.13.0" # or "lockstep"
    maa_url = "https://www.example.com" # optional, for Azure only
}

# constellation_attestation provides:
# data.constellation_attestation.foo.measurements
# data.constellation_attestation.foo.attestation


data "constellation_image" "foo" {
    image_version = "v2.21.0" # or "lockstep"
    attestation_variant = "GCP-SEV-ES"
    csp = "GCP"
    region = "us-central1" # optional, for AWS only
}

# constellation_image provides:
# constellation_image.foo.reference (CSP-specific image reference)

The Terraform state is to be considered sensitive, as it contains the Master- & Init-Secret, as well as the Kubeconfig.

The constellation_cluster resource is the main resource implemented by the provider. It declares a Constellation cluster with a specific configuration. Applying it will create the cluster if not existing, upgrade the cluster when the changes can be performed in place (e.g. K8s / node image / microservice update) or recreate the resource when the update can't be performed in-place (e.g. changing the master secret), update it with the according configuration if already existing, or deletes it ("it" as in the state, but not the underlying infrastructure) if not present in the configuration but in the state. If resource recreation is necessary (i.e. if the resources cannot be updated in-place), an error is thrown that indicates that content on previously created persistent volumes (encrypted with the old mastersecret) cannot be retrieved with the post-recreation / new mastersecret.

The "constellation_attestation" and "constellation_image" objects are data sources, which are objects that should be evaluated by the Provider each time the state is refreshed (i.e. each time any Terraform command that evaluates configuration against state), but have no observable side effects. For image and attestation, this is required as the provider need to evaluate latest values or map CSP-agnostic image references (e.g. v2.13.0) to CSP-specific image references (e.g. /CommunityGalleries/.../Image for Azure). This is implemented as an nilpotent API call and thus has no observable side-effects, but needs to be re-evaluated as the values returned by the API might change between evaluations.