RPL GPU Platform

RPL GPU Platform is a lab-focused control plane for granting GPU access, launching short-lived Jupyter workspaces on Docker Swarm workers, and managing runtime images without exposing Docker operations to end users.

Highlights

• Request/approve GPU workspace access with admin review flows.
• Launch isolated Jupyter sessions on labeled Docker Swarm GPU nodes.
• Route notebooks through Caddy with per-workspace paths and activity tracking.
• Manage runtime images and derived Python package environments from the UI.
• Enforce local resource requests and idle workspace shutdown.
• Ship releases through GitHub Actions, release-please, GHCR, and Renovate.

Stack

• NestJS modular monolith
• React server-rendered UI
• SCSS compiled to static CSS
• PostgreSQL
• Drizzle migrations
• Docker Swarm for notebook services
• Caddy for reverse proxy routes

Local Development

Local development uses single-node Docker Swarm. pnpm dev:setup initializes Swarm when Docker is inactive, labels the local node as a GPU worker, builds the local Jupyter image, and starts Postgres/Caddy.

If Docker is already joined to another Swarm as a worker, leave that Swarm before running local setup:

docker swarm leave

cp .env.example .env
pnpm install
pnpm dev:setup
pnpm db:generate
pnpm db:migrate
pnpm db:seed
pnpm dev

pnpm dev runs NestJS watch mode and Sass watch mode together. Edit styles in assets/styles/app.scss; Sass recompiles public/app.css automatically.

Open the app through Caddy:

http://localhost:18080

Direct NestJS debug endpoint:

http://localhost:3000

Caddy health check:

http://localhost:18080/healthz

pnpm dev loads .env. If you change PORT, update CADDY_APP_UPSTREAM in .env to match the NestJS port, for example CADDY_APP_UPSTREAM=host.docker.internal:3100, then restart Caddy with pnpm dev:setup.

Run the full local lifecycle check after Swarm, Caddy, Postgres, and the app are running:

RUN_LOCAL_E2E=1 pnpm test -- test/local-e2e.spec.ts

Local Cleanup

Stop dev Postgres/Caddy but keep database data and local Swarm state:

pnpm dev:down

Reset dev Postgres/Caddy data while keeping local Swarm state:

pnpm dev:reset
pnpm dev:setup
pnpm db:migrate
pnpm db:seed

Remove local workspace services/volumes, reset Compose data, and leave the local Swarm. Use this before turning the machine back into a Swarm worker or when you want a fully clean local dev environment:

pnpm dev:clean

To also remove the local Jupyter dev image:

REMOVE_DEV_IMAGE=1 pnpm dev:clean

Temporary bootstrap users:

admin / adminlabrpl
student01 / Student01Lab!

Production Build

pnpm build
pnpm start

Production Deployment

See DEPLOYMENT.md for the manual Docker Swarm deployment process using release images and migration orchestration.

See docs/PRODUCTION_SWARM.md for Swarm cluster setup and configuration.