A tool to automatically generate a NixOS config from a Docker Compose project.
Running a Docker Compose stack/project on NixOS is not well supported. One approach is to define a systemd service that runs docker-compose up
on start and docker compose down
on stop.
But with this approach, changes to individual services are not visible to NixOS, which means that NixOS will need to restart the systemd service on any change to the Compose file. This can be mitigated by defining a systemd reload handler, but it still is finicky to work with and will always remain opaque to NixOS.
To top it all off, using Docker Compose on NixOS is fairly redundant as the features you get with Compose are available natively on NixOS.
compose2nix
takes your existing Docker Compose file(s) and converts each YAML service definition into a oci-container
config. The tool also sets up systemd services to create all networks and volumes that are part of the Compose project. Since compose2nix
uses the same library that the Docker CLI relies on under the hood, you also get Compose file validation and syntax checking "for free".
- Supports both Docker and Podman out of the box.
- Each Compose service maps into a systemd service that is natively managed by NixOS.
- A change to one container service only impacts that container and any of its dependents.
- Generated systemd services can be extended from your NixOS config.
compose2nix
supports setting additional systemd service and unit options through Docker Compose labels (search for thecompose2nix.systemd.
label in the samples).
Install the compose2nix
CLI via one of the following methods:
-
Run using
nix run
(recommended):# Latest nix run github:aksiksi/compose2nix -- -h # Specific version nix run github:aksiksi/compose2nix/v0.3.0 -- -h # Specific commit nix run github:aksiksi/compose2nix/0c38d282d6662fc902fca7ef5b33e889f9e3e59a -- -h
-
Install from
nixpkgs
:# NixOS config environment.systemPackages = [ pkgs.compose2nix ];
-
Add the following to your
flake.nix
:compose2nix.url = "github:aksiksi/compose2nix"; compose2nix.inputs.nixpkgs.follows = "nixpkgs";
Optionally, you can pin to a specific version:
compose2nix.url = "github:aksiksi/compose2nix/v0.3.0";
You can then install the package by adding the following to your NixOS config:
environment.systemPackages = [ inputs.compose2nix.packages.x86_64-linux.default ];
Run compose2nix
. Note that project must either be passed in or set in the Compose file's top-level "name".
compose2nix -project=myproject
By default, the tool looks for docker-compose.yml
in the current directory and outputs the NixOS config to docker-compose.nix
.
- Basic implementation
- Support for most common Docker Compose features
- Support for using secret environment files
- Input: https://github.com/aksiksi/compose2nix/blob/main/testdata/compose.yml
- Output (Docker): https://github.com/aksiksi/compose2nix/blob/main/testdata/TestBasic.docker.nix
- Output (Podman): https://github.com/aksiksi/compose2nix/blob/main/testdata/TestBasic.podman.nix
agenix
works by decrypting secrets and placing them in /run/agenix/
. To feed this into your Nix config:
-
Place all secret env variables in the encrypted env file (e.g.,
my-env-file.env
). -
Mark the decrypted env file as readable by the user running
compose2nix
. -
Run
compose2nix
with the env file path(s) and set-include_env_files=true
:compose2nix --env_files=/run/agenix/my-env-file.env --include_env_files=true
Note
If you also want to ensure that you only include env files in the output Nix config, set -env_files_only=true
.
In this case, the project is called myproject
and the service name is myservice
. Replace podman
with docker
if using the Docker runtime.
sudo systemctl list-units podman-*
sudo systemctl list-units *myservice*
Note: if the Compose service has a container_name
set, then the systemd service will not include the project name.
sudo systemctl restart podman-myproject-myservice.service
- Pull the latest image for the container (requires
jq
):
sudo podman pull $(sudo podman inspect myproject-myservice | jq -r .[0].ImageName)
- Restart the service:
sudo systemctl restart podman-myproject-myservice.service
- Add a
io.containers.autoupdate=registry
label to each Compose service you want to have auto-updated.- Make sure to use a fully-qualified image path (e.g.,
docker.io/repo/image
). Otherwise, Podman will fail to start the container.
- Make sure to use a fully-qualified image path (e.g.,
- Run
sudo podman auto-update --dry-run
to see which containers would get updated. Omit--dry-run
to update & restart services.
You can optionally enable a Podman-provided timer that runs the command above once per day at midnight (by default):
# Enable the existing timer unit.
systemd.timers."podman-auto-update".wantedBy = [ "timers.target" ];
See this page for details: https://docs.podman.io/en/latest/markdown/podman-auto-update.1.html
By default, all generated services will be started by systemd on boot.
You can override this behavior in two different ways:
-
Disable auto-start for all services: Re-generate your config with
-auto_start=false
. -
Disable or enable auto-start for a single service: Add a Compose label to your service like this:
services: my-service: labels: # Enable - "compose2nix.settings.autoStart=true" # Disable - "compose2nix.settings.autoStart=false"
By default, this will only remove networks.
sudo systemctl stop podman-compose-myservice-root.target
You can do one of the following:
- Re-generate your NixOS config with:
-remove_volumes=true
- Run
sudo podman volume prune
to manually cleanup unreferenced volumes
sudo systemctl start podman-compose-myservice-root.target
compose2nix
has basic support for the Build spec. See [Supported Compose Features] below for details.
By default, a systemd service will be generated for each container build. This is a one-shot service that simply runs the build when started.
For example, if you have a service named my-service
with a build
set:
sudo systemctl start podman-build-my-service.service
Note that, until this is run, the container for my-service
will not be able to start due to the missing image.
If you run the CLI with -build=true
, the systemd service will be marked as a dependency for the service
container. This means that the build will be run before the container is started.
However, it is important to note that the build will be re-run on every restart of the root target or system. This will result in the build image being updated (potentially).
- Enable CDI support in your NixOS config:
{
hardware.nvidia-container-toolkit.enable = true;
}
Docker only:
Make sure you are running Docker 25+:
{
virtualisation.docker.package = pkgs.docker_25;
}
- Pass in CDI devices via either
devices
ordeploy
(both map to the same thing under the hood):
services:
myservice:
# ... other fields
# Option 1
devices:
- nvidia.com/gpu=all
# Option 2
deploy:
resources:
reservations:
devices:
# Driver must be set to "cdi" - all others are ignored.
- driver: cdi
device_ids:
- nvidia.com/gpu=all
# Required for Podman.
security_opt:
- label=disable
I always aim to support the latest stable version of NixOS (24.05 at the time of writing). As a result, some NixOS unstable options are not used.
If the option has a strong usecase, I am open to adding a CLI flag that can be deprecated once the option is stable.
If you are using the Docker runtime and a Compose service connects to multiple networks, you'll need to use v25+. Otherwise, the container service will fail to start.
You can pin the Docker version to v25 like so:
{
virtualisation.docker.package = pkgs.docker_25;
}
Discussion: #24
For some reason, when you run a rootful Podman container in a network that is marked as internal
, port forwarding to the host does not work. Podman seems to completely isolate the network from the external world - including the host network! Note that Docker claims to support this behavior out-of-the-box (ref).
There is a workaround: remove the internal
setting and set the network driver option no_default_route=1
(example).
networks:
my-network:
driver: bridge
driver_opts:
no_default_route: 1 # <<< This is what prevents external network access.
ipam:
config:
- subnet: 10.8.1.0/24
gateway: 10.8.1.0
This will allow you to connect from the host, while also preventing internet access from within the container.
This is where the check is done in Netavark: link
$ compose2nix -h
Usage of compose2nix:
-auto_format
if true, Nix output will be formatted using "nixfmt" (must be present in $PATH).
-auto_start
auto-start setting for generated service(s). this applies to all services, not just containers. (default true)
-check_systemd_mounts
if set, volume paths will be checked against systemd mount paths on the current machine and marked as container dependencies.
-create_root_target
if set, a root systemd target will be created, which when stopped tears down all resources. (default true)
-default_stop_timeout duration
default stop timeout for generated container services. (default 1m30s)
-env_files string
one or more comma-separated paths to .env file(s).
-env_files_only
only use env file(s) in the NixOS container definitions.
-generate_unused_resources
if set, unused resources (e.g., networks) will be generated even if no containers use them.
-ignore_missing_env_files
if set, missing env files will be ignored.
-include_env_files
include env files in the NixOS container definition.
-inputs string
one or more comma-separated path(s) to Compose file(s). (default "docker-compose.yml")
-output string
path to output Nix file. (default "docker-compose.nix")
-project string
project name used as a prefix for generated resources. this overrides any top-level "name" set in the Compose file(s).
-remove_volumes
if set, volumes will be removed on systemd service stop.
-root_path string
root path to use for any relative paths in the Compose file (e.g., volumes). if unset, the current working directory will be used.
-runtime string
one of: ["podman", "docker"]. (default "podman")
-service_include string
regex pattern for services to include.
-use_compose_log_driver
if set, always use the Docker Compose log driver.
-use_upheld_by
if set, upheldBy will be used for service dependencies (NixOS 24.05+).
-version
display version and exit
-write_nix_setup
if true, Nix setup code is written to output (runtime, DNS, autoprune, etc.) (default true)
If a feature is missing, please feel free to create an issue. In theory, any Compose feature can be supported because compose2nix
uses the same library as the Docker CLI under the hood.
Notes | ||
---|---|---|
image |
✅ | |
container_name |
✅ | |
environment |
✅ | |
env_file |
✅ | |
volumes |
✅ | |
labels |
✅ | |
ports |
✅ | |
dns |
✅ | |
cap_add/cap_drop |
✅ | |
logging |
✅ | |
depends_on |
Only short syntax is supported. | |
restart |
✅ | |
deploy.restart_policy |
✅ | |
deploy.resources.limits |
✅ | |
deploy.resources.reservations.cpus |
✅ | |
deploy.resources.reservations.memory |
✅ | |
deploy.resources.reservations.devices |
Only CDI driver is supported. | |
devices |
✅ | |
networks.aliases |
✅ | |
networks.ipv*_address |
✅ | |
network_mode |
✅ | |
privileged |
✅ | |
extra_hosts |
✅ | |
sysctls |
✅ | |
shm_size |
✅ | |
runtime |
✅ | |
security_opt |
✅ | |
command |
✅ | |
entrypoint |
✅ | |
healthcheck |
✅ | |
hostname |
✅ | |
mac_address |
✅ | |
user |
✅ |
labels |
✅ |
name |
✅ |
driver |
✅ |
driver_opts |
✅ |
enable_ipv6 |
✅ |
ipam |
✅ |
external |
✅ |
internal |
✅ |
driver |
✅ |
driver_opts |
✅ |
labels |
✅ |
name |
✅ |
external |
✅ |
Notes | ||
---|---|---|
args |
✅ | |
tags |
✅ | |
context |
Git repo is not supported | |
network |
❌ | |
image +build |
❌ |
name
- ✅
- Use in a Nix shell:
nix shell github:aksiksi/compose2nix
- Install the command using
go
:go install github.com/aksiksi/compose2nix
- Clone this repo and run
make build
.