Local Clinic Deployment (Docker-in-Docker)

This page explains how to deploy a Federated Learning Net Client locally using Docker-in-Docker (DinD). This approach packages the entire clinic stack — all services, configuration, and pre-pulled Docker images — into a single self-contained Docker container. It is designed for clinic environments where:

A simple, portable deployment is required (one container, one command).
FL-Net first needs to be tested in an isolated environment before a production deployment.

How DinD Works

The clinic stack runs inside a Docker container that itself runs a Docker daemon (Docker-in-Docker). The inner daemon starts a docker compose stack with all the clinic services. This means:

All clinic service images are bundled into the DinD image at build time — no internet needed at runtime.
The inner compose stack creates its own isolated Docker network inside the DinD container.
Only port 80 is exposed from the DinD container to the host.

Host machine
└── DinD container (clinic-dind)
    ├── dockerd (inner Docker daemon)
    └── docker compose stack (inner)
        ├── local-learning-api
        ├── local-learning-api-db  (PostgreSQL)
        ├── orch-api
        ├── controller  ──────────────────────────▶ Global Relay Server (TCP 9150)
        ├── instance-manager-frontend
        ├── keycloak
        ├── keycloak-postgres
        └── reverse-proxy (Nginx :80)  ◀──── Host port

Services at a Glance

Service	Image	Purpose
`local-learning-api`	`local-learning-api`	REST API for the clinic node — manages local data, connects to global server
`local-learning-api-db`	`postgres:17.5`	PostgreSQL database for the local learning API
`orch-api`	`orch-api`	Orchestration API — spawns federated learning tool containers
`controller`	`controller`	FLNet controller — connects to the global relay server and manages federated runs
`instance-manager-frontend`	`local-fl-net`	Web frontend for the clinic node
`keycloak`	`quay.io/keycloak/keycloak:26.5`	Identity provider (authentication)
`keycloak-postgres`	`postgres:17.5`	PostgreSQL database for Keycloak
`reverse-proxy-unencrypted`	`nginx`	Nginx reverse proxy, binds on `:80`

Prerequisites

On the build machine (internet access required):

Docker installed and running
Access to the GitLab container registry (gitlab.cosy.bio:5050) — you need valid credentials
The clinic-dind/ directory from the deployment repository

On the clinic machine (target deployment):

Docker installed
No internet access required at runtime
Port 80 (or a custom port) available for the clinic frontend

Step 1: Build the DinD Image

The build process pulls all required images, saves them into a .tar archive, and bundles everything into a single Docker image.

From the clinic-dind/ directory, run:

./build-local-dind.sh

This script does the following:

Reads all image: entries from docker-compose.yml
Pulls each image (docker pull)
Saves them all into images.tar (docker save)
Builds the DinD Docker image tagged feddb-dind
Removes the temporary images.tar

The resulting image feddb-dind contains everything needed to run the clinic stack offline.

Tip: If images.tar already exists, the pull and save steps are skipped. Delete it to force a fresh image pull.

To only prepare the image archive (without building the DinD image):

./prepare-local-dind.sh

Step 2: Prepare Clinic Data Files

The clinic node requires two data files that are mounted into the DinD container at runtime:

File	Purpose
`clinic.csv`	The clinic's patient/measurement dataset in the expected CSV format
`us-130-clinics.json`	Connector configuration describing how the ETL process should read the CSV

These are mounted read-only into /connectors/ inside the DinD container. The inner local-learning-api service reads them from there.

Step 3: Run the DinD Container

Use the provided run script:

./run-local-dind.sh

By default this exposes the clinic frontend on port 4601 of the host.

Configurable environment variables

You can override any of these before running:

HOST_PORT=4601 \
DEPLOYED_ON_ADDRESS=http://localhost:4601 \
COMPOSE_PROJECT_NAME=fl-net-clinic1 \
CLINIC_DATASET_PATH=/path/to/your/clinic.csv \
CLINIC_CONNECTOR_PATH=/path/to/us-130-clinics.json \
./run-local-dind.sh

Variable	Default	Description
`HOST_PORT`	`4601`	The port on the host machine that maps to the clinic frontend (`:80` inside DinD).
`DEPLOYED_ON_ADDRESS`	`http://localhost:{HOST_PORT}`	The full URL the clinic frontend is reachable at. Used by Keycloak for redirect URLs.
`COMPOSE_PROJECT_NAME`	`fl-net-clinic1`	Docker Compose project name for the inner stack. Change this when running multiple clinics on the same host.
`CLINIC_DATASET_PATH`	`{repo}/deployment/example-data/us-130/clinics/clinic_001.csv`	Absolute path to the clinic CSV dataset on the host.
`CLINIC_CONNECTOR_PATH`	`{script_dir}/us-130-clinics.json`	Absolute path to the connector config JSON on the host.

What the run script does

The run-local-dind.sh script runs the feddb-dind image with:

--privileged — required for the inner Docker daemon
--cgroupns host — required for cgroup management inside DinD
-v /sys/fs/cgroup:/sys/fs/cgroup:rw — cgroup access for the inner daemon
-v clinic-dind-docker:/var/lib/docker — persistent Docker layer cache for the inner daemon (so images survive restarts)
-v $CLINIC_DATASET_PATH:/connectors/clinic.csv:ro — mounts clinic data read-only
-v $CLINIC_CONNECTOR_PATH:/connectors/us-130-clinics:ro — mounts connector config read-only
-p $HOST_PORT:80 — exposes the Nginx reverse proxy

The inner Docker daemon, clinic stack, and all networking happen entirely inside this one container.

Step 4: What Happens at Startup

When the DinD container starts, entrypoint.sh runs automatically:

Starts the inner Docker daemon (dockerd-entrypoint.sh)
Waits up to 60 seconds for the daemon to be ready
Loads pre-saved images from /compose/images.tar into the inner Docker daemon (docker load)
Logs into the registry using the CI_REGISTRY_USER and CI_REGISTRY_PASSWORD environment variables — required so orch-api can later pull federated learning tool images on demand
Resolves host.docker.internal — auto-detects the host IP so the inner compose services can reach the host machine if needed
Writes a runtime .env — merges the baked-in .env with any environment variables passed to the DinD container
Runs docker compose up with the inner compose file, attaching logs from controller, orch-api, and local-learning-api

Step 5: Configure Global Server Connection

The inner controller service connects to the global relay server. These URLs are injected at runtime via environment variables:

Variable	Description
`GLOBAL_LEARNING_API_URL`	HTTP URL of the global learning API (e.g. `https://dev.federated-learning.net/api`)
`GLOBAL_LEARNING_API_WEBSOCKET_URL`	WebSocket URL of the global learning API (e.g. `wss://dev.federated-learning.net/api`)
`GLOBAL_SCHEMA_API_URL`	URL of the data modeler API on the global server
`GLOBAL_RELAY_HTTP_URL`	HTTP URL of the relay server on the global server
`GLOBAL_RELAY_TCP_ADDRESS`	TCP address of the relay server (e.g. `dev.federated-learning.net:9150`)

These are pre-set in .env to point at the staging global server. Pass them as environment variables to docker run to override:

docker run -it \
  --name clinic-dind \
  --privileged \
  --cgroupns host \
  --rm \
  -e COMPOSE_PROJECT_NAME=fl-net-clinic1 \
  -e DEPLOYED_ON_ADDRESS=http://localhost:4601 \
  -e GLOBAL_LEARNING_API_URL=https://your-global-server.com/api \
  -e GLOBAL_LEARNING_API_WEBSOCKET_URL=wss://your-global-server.com/api \
  -e GLOBAL_SCHEMA_API_URL=https://your-global-server.com/data-modeler \
  -e GLOBAL_RELAY_HTTP_URL=https://your-global-server.com/relay \
  -e GLOBAL_RELAY_TCP_ADDRESS=your-global-server.com:9150 \
  -e CI_REGISTRY_USER=your-gitlab-username \
  -e CI_REGISTRY_PASSWORD=your-gitlab-pat \
  -v clinic-dind-docker:/var/lib/docker \
  -v /sys/fs/cgroup:/sys/fs/cgroup:rw \
  -v /path/to/clinic.csv:/connectors/clinic.csv:ro \
  -v /path/to/us-130-clinics.json:/connectors/us-130-clinics:ro \
  -p 4601:80 \
  feddb-dind

Step 6: Nginx Routing (Clinic Node)

The inner Nginx reverse proxy routes all traffic inside the DinD container:

Path prefix	Upstream service	Notes
`/local-learning-api/`	`local-learning-api:8080`	Local learning REST API
`/auth/`	`keycloak:8080`	Keycloak authentication
`/`	`instance-manager-frontend:80`	Clinic web frontend (catch-all)

Step 7: Initial Setup

Once the stack is running, access the clinic node at http://localhost:4601 (or your configured host port).

Navigate to http://localhost:4601/auth/
Log in with:
- Username: admin (or the value of KC_BOOTSTRAP_ADMIN_USERNAME in env/keycloak-secrets.env)
- Password: password (or the value of KC_BOOTSTRAP_ADMIN_PASSWORD)
Change the admin password immediately.

Create clinic users

In Keycloak, switch to the FLNet-Client realm (left sidebar → Manage Realms).
Go to Users → Add User.
Make sure to tick Email verified (or configure SMTP, or disable email verification requirement).
Assign the user to a group:

Group	Access level
Admin	Full access — see and edit everything
Data-Access-Manager	Read all patient data, control query and federated learning access, view all logs
Data-Admin	Read, edit, and add patient data; view data-related logs

Running Multiple Clinics on the Same Host

To simulate multiple clinic nodes on a single host, run the DinD container multiple times with different project names and ports:

HOST_PORT=4601 COMPOSE_PROJECT_NAME=fl-net-clinic1 ./run-local-dind.sh

# In a second terminal:
HOST_PORT=4602 COMPOSE_PROJECT_NAME=fl-net-clinic2 \
  CLINIC_DATASET_PATH=/path/to/clinic2.csv \
  ./run-local-dind.sh

Each DinD container runs a completely isolated inner Docker stack. The COMPOSE_PROJECT_NAME ensures Docker volume names do not collide between instances.

Persisting the Inner Docker Layer Cache

The run script mounts a named volume clinic-dind-docker at /var/lib/docker inside the DinD container:

-v clinic-dind-docker:/var/lib/docker

This volume stores the inner daemon's image and container layers. On subsequent starts, images do not need to be re-loaded from the .tar file, making restarts significantly faster.

To start fresh (e.g. after a DinD image update):

docker volume rm clinic-dind-docker

Secrets Reference

The secrets files inside the DinD image contain default development passwords. For production deployments, rebuild the DinD image with updated secret files before distributing it.

File	Key variables
`env/keycloak-secrets.env`	`POSTGRES_PASSWORD`, `KC_DB_PASSWORD`, `KC_BOOTSTRAP_ADMIN_PASSWORD`, `LOCAL_LEARNING_SECRET`, `DATA_IMPORTER_SECRET`
`env/local-learning-secrets.env`	`POSTGRES_PASSWORD`, `QUARKUS_DATASOURCE_PASSWORD`, `QUARKUS_OIDC_CREDENTIALS_SECRET`
`env/orch-secrets.env`	`POSTGRES_PASSWORD`, `QUARKUS_DATASOURCE_PASSWORD`, `ORCH_DOCKER__GITLAB__REGISTRY_PASSWORD`, `PIPELINE_DOCKER_PASSWORD`, `PIPELINE_REPO_TOKEN`

Troubleshooting

DinD container starts but frontend is not reachable

Check that the host port is not already in use and that the inner Nginx started successfully:

docker exec clinic-dind docker compose logs reverse-proxy-unencrypted

Inner Docker daemon fails to start

The host kernel must support the cgroup configuration needed by DinD:

Make sure you pass --cgroupns host and -v /sys/fs/cgroup:/sys/fs/cgroup:rw
Make sure /lib/modules is readable (-v /lib/modules:/lib/modules:ro)

Clinic does not appear in the global server

Check that the controller service inside DinD can reach the global relay TCP address:

docker exec clinic-dind docker compose logs controller

Verify the GLOBAL_RELAY_TCP_ADDRESS is set correctly and that the relay TCP port is reachable from the clinic host.

Registry login fails / tool images can't be pulled

The CI_REGISTRY_USER and CI_REGISTRY_PASSWORD environment variables must be passed to the DinD container. These are used at startup to authenticate the inner Docker daemon with the GitLab registry (so orch-api can later pull federated learning tool images).

Multiple clinics: port or volume conflicts

Ensure each clinic run uses a unique HOST_PORT and COMPOSE_PROJECT_NAME. The COMPOSE_PROJECT_NAME prefixes all volume names inside Docker to avoid collisions.

How DinD Works​

Services at a Glance​

Prerequisites​

Step 1: Build the DinD Image​

Step 2: Prepare Clinic Data Files​

Step 3: Run the DinD Container​

Configurable environment variables​

What the run script does​

Step 4: What Happens at Startup​

Step 5: Configure Global Server Connection​

Step 6: Nginx Routing (Clinic Node)​

Step 7: Initial Setup​

First login (Keycloak admin)​

Create clinic users​

Running Multiple Clinics on the Same Host​

Persisting the Inner Docker Layer Cache​

Secrets Reference​

Troubleshooting​