Deployment Paths#

POC Factory is delivered as a container application. Choose Docker Compose for a trusted single host or Kubernetes for a managed shared service. Both paths require an effective inference profile and persistent application state.

Workflow Overview#

The four checks below apply to either deployment path.

  1. Obtain the release bundle or Helm chart and confirm the image set.

  2. Prepare registry access, inference credentials, encryption key, and storage.

  3. Deploy the services.

  4. Validate health, identity, provider roles, one POC, and persistence.

Key Concepts#

These boundaries shape where the application can run and how it is recovered.

Deployment boundary#

The Docker release bundle uses published NGC images and mounts the host Docker socket for generated-candidate build and runtime validation. Treat that socket as privileged host access. The Kubernetes chart disables Docker-dependent validation by default and records those gates as skipped.

Shared state#

POC records and credentials live in PostgreSQL; archives and job state use mounted storage. A successful container restart is not a restore test. Plan backups for the related stores together.

Common Methods for Completing the Procedure#

Choose the deployment surface that matches your operating model.

Path

Use when

Required inputs

NGC and Docker Compose

Local evaluation, demos, or a trusted single-host team service

NGC release bundle and access, Docker Compose, inference profile, encryption key, host capacity

Kubernetes and Helm

Shared HTTPS service with managed identity, storage, and operations

Cluster, Helm chart, image pull secret or mirror, Kubernetes Secrets, ingress and TLS, persistent volumes, PostgreSQL

The release guide’s starting capacity for Compose is at least 8 GB RAM and 20 GB free disk; generated workloads can require more. The Helm production values request 4 GiB memory for the backend and limit it to 16 GiB. Size the actual environment for expected concurrency and generated artifacts.

Decision Factors#

Review these factors before selecting the release bundle or Helm chart.

  • Security: Use HTTPS, secure cookies, exact origin and host allowlists, and managed secrets for shared deployments.

  • Validation: Compose can run Docker build and runtime gates through its socket. Kubernetes defaults skip those gates; add a separate candidate validation environment when full container proof is required.

  • Storage: Compose uses host directories and named volumes. Kubernetes uses PVCs and can use managed PostgreSQL.

  • Availability: The production example keeps one backend replica because active progress and fine-tuning state are process-local and volumes commonly use ReadWriteOnce access.