Skip to content

Google Cloud image reference (archived)

Release binding: v0.1.0-enterprise-ready · 2026-05-19 · Marketplace-owned VM image, Terraform deployment

Archived technical reference — no current marketplace purchase path

HX-SDP is currently available only through a Private Appliance or scoped pilot. This page retains the dated Google Cloud packaging and runtime reference; it does not establish a currently available Google Cloud Marketplace listing, image access, or entitlement. Use these procedures only for a package and configuration explicitly supplied within your agreed deployment scope.

Contact HolonomiX to agree the scope and obtain an authorized release bundle, or see Deploy: Private Appliance. The former marketplace purchase instructions have been withdrawn.

This reference describes an architecture alongside Vertex AI and Google Cloud managed data services, including a sovereign and air-gapped target through Google Distributed Cloud. Any such target requires agreement and validation for your pilot. HX-SDP is not a replacement for any Google managed service.

1. Architecture

A single Compute Engine VM running three services:

  • hx-engine: the GPU-native engine, loopback only
  • hx-gate: the multi-tenant API gateway, the only externally reachable service, TLS 1.3 on 0.0.0.0:8443
  • redis-server: local state, loopback only

The archived packaging reference describes a hardened, signed VM image with an attached Marketplace license. Its Marketplace-owned image model is intended to retain customer access independently of producer-side IAM changes. This design description does not establish current image availability or a licensed entitlement.

2. Prerequisites

  • A GCP project with the Compute Engine API enabled.
  • IAM: roles/compute.admin on the deployment project.
  • GPU quota for at least one accelerator in the chosen zone.
  • Optional: an OIDC issuer for production auth.

3. Reference deployment resources

The archived marketplace packaging reference describes a Terraform module supplied in the deployment package. Its reference resource plan includes:

  • a Compute Engine VM with the Marketplace-owned image attached
  • a pd-ssd boot disk (default 100 GB)
  • a firewall rule hx-sdp-allow-gate permitting tcp/8443 from the configured source ranges
  • usage labels goog-product=hx-sdp on every resource

4. Custom deployment (Terraform CLI)

For a pilot explicitly using this cloud configuration, use the Terraform module and source reference supplied in your authorized deployment package. The archived package layout places the module at deploy/gcp/terraform. The example below is a technical reference, not a self-service deployment offer.

module "hx_sdp" {
  source                = "<module source from your deployment package>"
  project_id            = "my-prod-project"
  region                = "us-central1"
  zone                  = "us-central1-a"
  machine_type          = "a2-highgpu-1g"
  gpu_type              = "nvidia-tesla-a100"
  gpu_count             = 1
  network               = "projects/my-prod-project/global/networks/secured-vpc"
  subnet                = "projects/my-prod-project/regions/us-central1/subnetworks/secured-subnet"
  allowed_source_ranges = ["10.0.0.0/8"]
  enable_oslogin        = true
  enable_shielded_vm    = true
}

5. Machine types

Machine type GPU GPU memory Use case
g2-standard-8 nvidia-l4 x 1 24 GB small and medium production
a2-highgpu-1g nvidia-tesla-a100 x 1 40 GB production, around 1B vectors
a2-ultragpu-1g nvidia-a100-80gb x 1 80 GB large production, around 2B vectors
n2-standard-8 none (CPU only) n/a evaluation only

The image is GPU-ready out of the box: the NVIDIA driver and CUDA runtime are installed at bake time, and a GPU preflight confirms GPU, driver, and CUDA at every boot, writing /var/log/hx-sdp/gpu-preflight.json.

Constraints:

  • GCP requires on_host_maintenance=TERMINATE for any GPU-attached VM. The Terraform module enforces this when gpu_count > 0.
  • Multi-GPU configurations are not supported in v0.1.x; ship one GPU per VM.
  • GPU live-migration is unsupported on GCP; maintenance windows must accept VM restart.
  • CPU-only n2-standard-8 falls back to a substantially slower CPU code path and is not intended for production traffic.

6. First-boot lifecycle

On every fresh boot, cloud-init:

  1. verifies /opt/hx-sdp/MANIFEST.sha256 against the signed release manifest
  2. runs the GPU, CUDA, and driver preflight
  3. generates the bootstrap admin credential
  4. provisions a bootstrap TLS pair if no customer TLS exists
  5. starts redis-server, hx-engine, hx-gate, in that order
  6. writes /var/log/hx-sdp/first-boot-report.json

Manifest verification failure causes services to fail closed and the boot report to record decision=FAIL.

7. Validation

sudo /opt/hx-sdp/scripts/validation/run_first_boot_validation.sh
sudo /opt/hx-sdp/scripts/validation/gpu_preflight.sh

Then run the smoke test from the Quickstart: health check, tenant creation, PUT /v1/put, and POST /v1/query/vector.

8. Upgrade and rollback

The pattern is blue/green VM replacement: Terraform points at a new image family revision, plans, applies, and traffic cuts over. Roll back by pointing Terraform at the previous revision.

9. Uninstall

terraform destroy

10. Where to file issues