Skip to main content
Version: Next

Docker

The repository ships a multi-stage docker/Dockerfile that builds the release binary with cargo-chef layer caching and runs it on debian:bookworm-slim. Release-tagged builds are published by CI as container images; the sections below cover running the engine, persisting state, and building your own image.

Run the engine​

The image entrypoint is arkflow --config /app/etc/config.yaml, so mounting your configuration at that path is all a single-node deployment needs:

docker run -d --name arkflow \
  -p 8080:8080 \
  -v $PWD/config.yaml:/app/etc/config.yaml:ro \
  -v arkflow-data:/app/data \
  ghcr.io/arkflow-rs/arkflow:0.6.0   # pin a released tag (or digest); do not float :latest in production
  • -p 8080:8080 exposes the HTTP health server (/health, /readiness, /liveness — see Operate overview) for container health checks.
  • -v arkflow-data:/app/data persists everything the default relative paths write under the working directory (/app): WAL directories, embedded state roots, and file:// checkpoint stores. Point durability.path, state.root, and checkpoint.object_store_uri at paths under this volume so a recreated container recovers instead of starting cold.

Distributed jobs​

A node participating in a split placement must publish a routable data plane: set health_check.data_port and health_check.data_host (the LAN address peers dial, not 0.0.0.0), and publish the port with -p <data_port>:<data_port>. Nodes without data_port keep the co-located placement behavior and never advertise the network_shuffle capability — see Distributed Jobs.

Validate before containerizing​

The image carries the full binary, so a misconfigured deployment can be caught before it enters orchestration:

docker run --rm \
  -v $PWD/config.yaml:/app/etc/config.yaml:ro \
  ghcr.io/arkflow-rs/arkflow:0.6.0 \
  /app/arkflow --config /app/etc/config.yaml --validate

Build your own image​

The Dockerfile needs the full toolchain headers (clang, protobuf-compiler, OpenSSL, libcurl, python3-dev for the Python UDF processor); the cargo-chef planner/builder split keeps dependency rebuilds out of the per-change path:

git clone https://github.com/arkflow-rs/arkflow.git
cd arkflow
docker build -f docker/Dockerfile -t arkflow:local .
note

The release profile uses fat LTO; the build needs ample disk and memory (BuildKit cache plus link-time optimization). CI prunes the local image cache and adds swap before building for the same reason.

Control plane and console​

The Hub and its console ship separately from the engine container — their deployment model (binary flags, SQLite/PostgreSQL storage, static console assets) is covered in Control plane deployment.