SIE Alibaba Cloud ACK Terraform Module

This module creates a GPU-ready Alibaba Cloud Container Service for Kubernetes (ACK) cluster and the supporting infrastructure needed to run SIE. After Terraform creates the infrastructure, install the SIE runtime with the sie-cluster Helm chart.

By default, the module deploys in Frankfurt (eu-central-1) across zones eu-central-1a and eu-central-1c. It creates an ecs.g7.xlarge system pool and an on-demand ecs.gn7i-c8g1.2xlarge GPU pool with one NVIDIA A10 24 GiB GPU per node. The GPU pool scales from zero to ten nodes and uses a 500 GiB system disk. Both pools use the ACK-resolved AliyunLinux3ContainerOptimized image family.

What you get

• A VPC with separate system and GPU vSwitches in both configured zones.
• Private-node internet egress through an Enhanced Internet NAT gateway and a pay-as-you-go EIP.
• A private-API ACK Pro cluster using Flannel, with deletion protection enabled by default.
• Autoscaling system and GPU node pools with encrypted ESSD system disks, IMDSv2-only metadata access, and separate ECS-trusted RAM roles.
• ACK CSI components, the managed ack-nvidia-device-plugin, RRSA, and the managed ack-pod-identity-webhook.
• A private, encrypted, versioned OSS bucket for model weights under models/ and large request payloads under payloads/.
• A least-privilege RRSA workload role scoped to the configured SIE namespace and ServiceAccount.
• KMS envelope encryption for ACK Secrets, with automatic key rotation enabled by default.
• ACK API audit and control-plane logs in an SLS project.
• Optional private repositories in an existing ACR Enterprise Edition instance.
• An optional SSH bastion restricted to one operator /32; the ACK API remains private.

Not included: Kubernetes workloads, account-wide service activation and service roles, an ACR subscription, registry credentials, kubeconfig storage, or the purchase of a KMS Default Key Rotation entitlement.

Quick start

Use the module from the Terraform Registry

Configure OSS and SLS Signature V4 in the provider, then call the module with a unique cluster name and one of the supported KMS configurations. Pin version to the value shown in the Registry Provision Instructions for reproducible deployments.

terraform {
  required_version = ">= 1.14"

  required_providers {
    alicloud = {
      source  = "aliyun/alicloud"
      version = "~> 1.289.0"
    }
  }
}

provider "alicloud" {
  region = "eu-central-1"

  sign_version {
    oss = "v4"
    sls = "v4"
  }
}

module "sie_ack" {
  source = "superlinked/sie/alicloud"
  # version = "<version-from-registry-provision-instructions>"

  cluster_name               = "replace-with-a-unique-cluster-name"
  ack_secret_kms_instance_id = "replace-with-a-same-region-software-kms-instance-id"
}

Initialize Terraform, review a saved plan, and apply it when it matches your intended infrastructure:

terraform init
terraform plan -out=tfplan
terraform apply tfplan

Run the included example

The examples/dev-gn7i configuration deploys the default Frankfurt A10 topology. From a checkout of the module source:

cd examples/dev-gn7i
cp terraform.tfvars.example terraform.tfvars
# Set a unique cluster_name and select one KMS configuration below.
terraform init
terraform plan -out=tfplan
terraform apply tfplan

Terraform returns a command for obtaining a renewable, 15-minute private kubeconfig without writing credentials into state:

terraform output -raw kubeconfig_command

Run the printed command from a host with private connectivity to the ACK API, save its output in a mode-0600 temporary file, and use that file with kubectl. Continue with Connect to ACK and install SIE.

Prerequisites

Before applying the module:

1. Enable ACK, KMS, OSS, and SLS in the target account and region.
2. Create the Alibaba Cloud service roles required by ACK, NAT Gateway, Auto Scaling, and OOS. The Terraform principal must be able to create resources, create and pass RAM roles, and use the selected KMS resources.
3. Confirm that ACK Pro and the required VPC, vSwitch, NAT, EIP, ECS, OSS, KMS, and SLS quotas are available.
4. Confirm current ecs.g7.xlarge and ecs.gn7i-c8g1.2xlarge inventory in each configured zone. GPU and spot capacity can change independently of Terraform configuration.
5. Use ACK Kubernetes 1.32 or later. The managed NVIDIA device-plugin lifecycle requires 1.32 or later.
6. Confirm that ACK offers AliyunLinux3ContainerOptimized for the selected Kubernetes version, architecture, and region. For IMDSv2, the resolved Alibaba Cloud Linux 3 Container-Optimized image must be release 20241226 or later, and installed ACK components must support cgroup v2 and IMDSv2.
7. Install Terraform 1.14 or later, the Aliyun CLI, jq, kubectl, and Helm.
8. Configure credentials through the standard Alicloud provider credential chain. Do not put access keys in Terraform files or state.
9. Ensure the machine used for kubectl and Helm can reach the private ACK API, either through an existing private network path or the optional bastion.

The optional ACR integration also requires an existing ACR Enterprise Edition instance and working network and DNS connectivity from the cluster.

Configure KMS encryption

ACK Secret envelope encryption is enabled by default. Select one of these KMS configurations before planning.

Use an existing rotating key

create_ack_secret_kms_key = false
ack_secret_kms_key_id     = "replace-with-a-same-region-rotating-kms-key-id"

The key must resolve exactly once in the selected region and be Enabled, Aliyun_AES_256, authorized for ENCRYPT/DECRYPT, and configured for automatic rotation. The module uses the key but does not manage its lifecycle.

Create a key in an existing software KMS instance

ack_secret_kms_instance_id            = "replace-with-a-same-region-software-kms-instance-id"
ack_secret_kms_rotation_interval_days = 30

Leave create_ack_secret_kms_key=true. The module creates the cluster key in the supplied instance. The rotation interval must be a whole number from 7 through 365 days. The software KMS instance remains outside this module.

Use an active Default Key Rotation entitlement

ack_secret_kms_default_rotation_entitled = true

Leave create_ack_secret_kms_key=true. Set this value only after the paid, one-per-account-and-region entitlement is active. This option uses a fixed 365-day rotation interval. The module cannot purchase, renew, cancel, or import the entitlement.

Automatic rotation is recommended. Setting ack_secret_kms_automatic_rotation_enabled=false keeps KMS envelope encryption enabled but creates the key without automatic rotation. In that configuration, you are responsible for the key's rotation policy.

The ack_secret_encryption.rotation_interval output uses the provider's canonical seconds format, such as 31536000s for 365 days.

Connect to ACK and install SIE

The module never creates a public ACK API endpoint. Use an existing private network path or enable the optional bastion with an existing ECS key pair and one canonical, globally routable operator /32. Retrieve kubeconfig with the kubeconfig_command output and keep it in a protected temporary file outside the module directory.

Install SIE chart 0.8.2 with its ACK values file and the Terraform outputs for OSS and RRSA. Its v0.8.2 application version selects the matching SIE runtime images. The Terraform module version is independent of the chart version.

MODEL_CACHE_URL="$(terraform output -raw model_cache_bucket_url)"
PAYLOAD_STORE_URL="$(terraform output -raw payload_store_url)"
RRSA_ROLE_NAME="$(terraform output -raw rrsa_workload_role_name)"

helm pull oci://ghcr.io/superlinked/charts/sie-cluster --version 0.8.2 --untar
helm upgrade --install sie-cluster ./sie-cluster \
  --namespace sie \
  --create-namespace \
  --values ./sie-cluster/values-ack.yaml \
  --set workers.common.clusterCache.enabled=true \
  --set-string workers.common.clusterCache.url="${MODEL_CACHE_URL}" \
  --set-string payloadStore.url="${PAYLOAD_STORE_URL}" \
  --set-string serviceAccount.annotations."pod-identity\.alibabacloud\.com/role-name"="${RRSA_ROLE_NAME}"

unset MODEL_CACHE_URL PAYLOAD_STORE_URL RRSA_ROLE_NAME

When upgrading an existing installation with custom model profiles, review the SIE 0.8.0 breaking changes for adapter options and launch arguments before deploying SIE 0.8.2.

The ACK chart values leave Ingress disabled. Install and secure an ACK-compatible ingress controller before enabling Ingress.

Model cache, payload store, and workload identity

The shared OSS bucket exposes two separate paths:

• oss://<bucket>/models stores model weights. The workload role can list and read this prefix but cannot modify it.
• oss://<bucket>/payloads stores work items that are too large for the queue. The workload role can list, read, write, and delete this prefix. OSS lifecycle rules remove current and noncurrent payload objects after the configured expiration period and abort incomplete multipart uploads.

Payload storage is required for requests over 1 MiB, including many image requests. Keep create_model_cache=true unless you configure another supported payload store when installing SIE.

ACK projects a short-lived OIDC token into the sie-server ServiceAccount. The runtime exchanges that token for STS credentials and signs OSS requests with region-scoped Signature V4. Node and bastion roles have no OSS permission, metadata credential fallback is disabled, and no long-lived Alibaba Cloud AccessKey is stored in Terraform, Helm, or a Kubernetes Secret.

OSS is not suitable for the mutable sie-config store because OSS PutObject does not provide the required non-empty compare-and-swap operation. Keep sie-config on the chart's local or PVC-backed store.

Node pools and supported images

AliyunLinux3ContainerOptimized is the supported image_type for both system and GPU pools. ACK resolves the concrete regional image; do not replace it with a date-stamped image ID. This image family provides the cgroup v2 and IMDSv2 properties required by the module. Plain AliyunLinux3 may resolve without cgroup v2, and the provider's ContainerOS value does not select the required GPU image family.

Changing a node pool's image_type can roll its nodes. Check replacement capacity and workload disruption before applying the change. If node-pool creation fails because inventory or quota is unavailable and no pool was created, resolve the capacity issue and retry with the same Terraform state.

The default GPU pool uses on-demand capacity because Frankfurt spot inventory can be intermittent. Spot pools are supported, but must keep compensate_with_on_demand=true. Compensation can improve capacity but does not guarantee it and may increase cost.

The default GPU system disk is 500 GiB. The SIE chart's model cache uses a 300 GiB emptyDir size limit on that root disk, leaving nominal space for container images, logs, the operating system, and kubelet eviction thresholds. The size limit does not reserve disk space.

System and GPU pools use distinct ECS-trusted RAM roles. You can supply an existing system role with system_node_ram_role_name and per-GPU-pool roles with ram_role_name; otherwise, the module creates dedicated roles. Node roles do not receive OSS permissions because SIE accesses OSS through RRSA.

The Frankfurt A10 topology described above is the module's supported default. Always recheck image metadata, quota, and instance inventory before deploying it in another account or at a later date.

Optional ACR Enterprise repositories

ACR repository creation is disabled by default and never creates a paid ACR instance. To create private repositories in an existing Enterprise Edition instance:

enable_acr_repositories     = true
acr_enterprise_instance_id = "replace-with-an-existing-acr-ee-instance-id"
acr_registry_domain        = "replace-with-an-existing-acr-registry-hostname"

The module creates only the configured namespace and repositories. It does not create or subscribe to an ACR instance, retrieve registry credentials, or configure Kubernetes image-pull secrets. Configure VPC and DNS connectivity and image-pull authentication separately.

Key variables

Variable	Default	Purpose
alicloud_region	eu-central-1	Alibaba Cloud region for all resources
zones	eu-central-1a, eu-central-1c	ACK and node-pool zones
kubernetes_version	ACK default	Kubernetes 1.32 or later
deletion_protection	true	Protect the ACK cluster from deletion
system_node_pool	ecs.g7.xlarge, 1-5 nodes	System-pool image, capacity, and disk
gpu_node_pools	A10, on-demand, 0-10 nodes	GPU-pool image, capacity, disk, and RAM role
create_model_cache	true	Create the OSS model-cache and payload-store bucket
model_cache_force_destroy	false	Delete OSS object versions during bucket destruction
ack_secret_encryption_enabled	true	Encrypt ACK Secrets with KMS
ack_secret_kms_default_rotation_entitled	false	Confirm an active Default Key Rotation entitlement
ack_secret_kms_instance_id	null	Existing software KMS instance for a new key
ack_secret_kms_rotation_interval_days	365	Rotation interval for a software KMS instance key
audit_logging_enabled	true	Send ACK audit and control-plane logs to SLS
bastion_enabled	false	Create a restricted SSH path to the private API
enable_acr_repositories	false	Create repositories in an existing ACR EE instance

See the Registry Inputs tab or variables.tf for the complete typed interface and validation rules.

Outputs

Output	Description
cluster_id	ACK cluster ID
kubeconfig_command	Aliyun CLI command for a renewable private kubeconfig
model_cache_bucket_url	oss:// URL to pass to workers.common.clusterCache.url
payload_store_url	oss:// URL to pass to payloadStore.url
rrsa_workload_role_name	RAM role for the pod-identity ServiceAccount annotation
gpu_node_pools	GPU pool IDs, instance types, capacity, disk, and purchase mode
ack_secret_encryption	KMS key ID, creation mode, deletion protection, and rotation settings
audit_logging	SLS project, retention, and enabled control-plane components
bastion_connection	Sensitive connection details for the optional bastion
acr_repositories	Optional ACR repository IDs and endpoints

Additional outputs expose the cluster name and private endpoint, networking IDs, node roles, OSS endpoint, and RRSA identity details.

Costs and cleanup

ACK Pro, the system node, NAT gateway, EIP, OSS, SLS, and KMS can incur charges even when the GPU pool has scaled to zero. The optional bastion and an existing ACR Enterprise Edition instance can add cost. Check current Alibaba Cloud pricing for the selected region before applying.

The defaults protect the ACK cluster and a module-created KMS key, and retain OSS object versions. Before destroying a cluster, update the configuration as needed:

deletion_protection                 = false
ack_secret_kms_deletion_protection = false
# Set true only when all versions in the module-created bucket should be deleted.
model_cache_force_destroy = true

Apply those changes first, then run terraform destroy. A module-created KMS key enters PendingDeletion for its configured 7-30 day window instead of disappearing immediately. Existing KMS instances, entitlements, ACR instances, and other external prerequisites are not removed by this module. After destroy, check the selected account and region for retained resources and pending KMS key deletion.