conda-lock

Conda lock is a lightweight library that can be used to generate fully reproducible lock files for conda environments.

It does this by performing a conda solve for each platform you desire a lockfile for.

This also has the added benefit of acting as an external pre-solve for conda as the lockfiles it generates results in the conda solver not being invoked when installing the packages from the generated lockfile.

Why?

Conda environment.yml files are very useful for defining desired environments but there are times when we want to be able to EXACTLY reproduce an environment by just installing and downloading the packages needed.

This is particularly handy in the context of a gitops style setup where you use conda to provision environments in various places.

Installation

pip install conda-lock
conda install -c conda-forge conda-lock

Basic usage

# generate a multi-platform lockfile
conda-lock -f environment.yml -p osx-64 -p linux-64

# optionally, update the previous solution, using the latest version of
# pydantic that is compatible with the source specification
conda-lock --update pydantic

# create an environment from the lockfile
conda-lock install [-p {prefix}|-n {name}]

# alternatively, render a single-platform lockfile and use conda command directly
conda-lock render -p linux-64
conda create -n my-locked-env --file conda-linux-64.lock

Pre 1.0 compatible usage (explicit per platform locks)

If you were making use of conda-lock before the 1.0 release that added unified lockfiles you can still get that behaviour by making use of the explicit output kind.

conda-lock --kind explicit -f environment.yml

Advanced usage

File naming

By default, conda-lock store its output in conda-lock.yml in the current working directory. This file will also be used by default for render, install, and update operations. You can supply a different filename with e.g.

conda-lock --lockfile superspecial.conda-lock.yml

The extension .conda-lock.yml will be added if not present. Rendered environment files (env or explicit) will be named as as "conda-{platform}.lock".

If you want to override that call conda-lock as follows.

conda-lock -k explicit --filename-template "specific-{platform}.conda.lock"

Compound specification

Conda-lock will build a spec list from several files if requested.

conda-lock -f base.yml -f specific.yml -p linux-64 --filename-template "specific-{platform}.lock"

In this case all dependencies are combined, and the ordered union of all channels is used as the final specification.

This works for all supported file types.

channel overrides

You can override the channels that are used by conda-lock in case you need to override the ones specified in an environment.yml

conda-lock -c conda-forge -p linux-64

platform specification

You may specify the platforms you wish to target by default directly in an environment.yml using the (nonstandard) platforms key:

# environment.yml
channels:
  - conda-forge
dependencies:
  - python=3.9
  - pandas
platforms:
  - osx-arm64
  - linux-64

If you specify target platforms on the command line with -p, these will override the values in the environment specification. If neither platforms nor -p are provided, conda-lock will fall back to a default set of platforms.

default category

You can may wish to split your dependencies into separate files for better organization, e.g. a environment.yml for production dependencies and a dev-environment.yml for development dependencies. You can assign all the dependencies parsed from a single file to a category using the (nonstandard) category key.

# dev-environment.yml
channels:
  - conda-forge
dependencies:
  - pytest
  - mypy=0.910
category: dev

The default category is main.

pip support

conda-lock can also lock the dependencies.pip section of environment.yml, using Poetry's dependency solver, if installed with the pip_support extra.

private pip repositories

Right now conda-lock only supports legacy pypi repos with basic auth. Most self-hosted repositories like Nexus, Artifactory etc. use this. To use this feature, add your private repo into Poetry's config including the basic auth in the url:

poetry config repositories.foo https://username:[email protected]/simple/

--dev-dependencies/--no-dev-dependencies

By default conda-lock will include dev dependencies in the specification of the lock (if the files that the lock is being built from support them). This can be disabled easily

conda-lock --no-dev-dependencies -f ./recipe/meta.yaml

--check-input-hash

Under some situation you may want to run conda lock in some kind of automated way (eg as a precommit) and want to not need to regenerate the lockfiles if the underlying input specification for that particular lock as not changed.

conda-lock --check-input-hash -p linux-64

When the input_hash of the input files, channels match those present in a given lockfile, that lockfile will not be regenerated.

--strip-auth, --auth and --auth-file

By default conda-lock will leave basic auth credentials for private conda channels in the lock file. If you wish to strip authentication from the file, provide the --strip-auth argument.

conda-lock --strip-auth -f environment.yml

In order to conda-lock install a lock file with its basic auth credentials stripped, you will need to create an authentication file in .json format like this:

{
  "domain": "username:password"
}

If you have multiple channels that require different authentication within the same domain, you can additionally specify the channel like this:

{
  "domain.org/channel1": "username1:password1",
  "domain.org/channel2": "username2:password2"
}

You can provide the authentication either as string through --auth or as a filepath through --auth-file.

conda-lock install --auth-file auth.json conda-linux-64.lock

--virtual-package-spec

Conda makes use of virtual packages that are available at runtime to gate dependency on system features. Due to these not generally existing on your local execution platform conda-lock will inject them into the solution environment with a reasonable guess at what a default system configuration should be.

If you want to override which virtual packages are injected you can create a file like

# virtual-packages.yml
subdirs:
  linux-64:
    packages:
      __glibc: 2.17
      __cuda: 11.4
  win-64:
    packages:
      __cuda: 11.4

conda-lock will automatically use a virtual-packages.yml it finds in the the current working directory. Alternatively one can be specified explicitly via the flag.

conda lock --virtual-package-spec virtual-packages-cuda.yml -p linux-64

Input hash stability

Virtual packages take part in the input hash so if you build an environment with a different set of virtual packages the input hash will change. Additionally the default set of virtual packages may be augmented in future versions of conda-lock. If you desire very stable input hashes we recommend creating a virtual-packages.yml file to lock down the virtual packages considered.

⚠️ in conjunction with micromamba

Micromamba does not presently support some of the overrides to remove all discovered virtual packages, consequently the set of virtual packages available at solve time may be larger than those specified in your specification.

Supported file sources

Conda lock supports more than just environment.yml specifications!

Additionally conda-lock supports meta.yaml (conda-build) and pyproject.toml ( flit, pdm and poetry based). These do come with some gotchas but are generally good enough for the 90% use-case.

meta.yaml

Conda-lock will attempt to make an educated guess at the desired environment spec in a meta.yaml. This is not guaranteed to work for complex recipes with many selectors and outputs. For multi-output recipes, conda-lock will fuse all the dependencies together. If that doesn't work for your case fall back to specifying the specification as an environment.yml

Since a meta.yaml doesn't contain channel information we make use of the following extra key to specify channels

# meta.yaml

extra:
  channels:
    - conda-forge
    - defaults

pyproject.toml

Since pyproject.toml files are commonly used by python packages it can be desirable to create a lock file directly from those dependencies to single-source a package's dependencies. This makes use of some conda-forge infrastructure (pypi-mapping) to do a lookup of the PyPI package name to a corresponding conda package name (e.g. docker -> docker-py). In cases where there exists no lookup for the package it assumes that the PyPI name, and the conda name are the same.

Channels

# pyproject.toml

[tool.conda-lock]
channels = [
    'conda-forge', 'defaults'
]

Platforms

Like in environment.yml, you can specify default platforms to target:

# pyproject.toml

[tool.conda-lock]
platforms = [
    'osx-arm64', 'linux-64'
]

Extras

If your pyproject.toml file contains optional dependencies/extras these can be referred to by using the --extras flag

# pyproject.toml

[tool.poetry.dependencies]
mandatory = "^1.0"
psycopg2 = { version = "^2.7", optional = true }
mysqlclient = { version = "^1.3", optional = true }

[tool.poetry.extras]
mysql = ["mysqlclient"]
pgsql = ["psycopg2"]

These can be referened as follows

conda-lock --extra mysql --extra pgsql -f pyproject.toml

When generating lockfiles that make use of extras it is recommended to make use of --filename-template covered here.

Filtering extras

By default conda-lock will attempt to solve for ALL extras/categories it discovers in sources. This allows you to render explicit locks with subset of extras, without needing a new solve.

However this does make the assumption that your extras can all be installed in conjunction with each other. If you want extras filtering to happen at the solve stage use the flag --filter-extras

conda-lock --extra incompatiblea --filter-extras -f pyproject.toml

Extra conda dependencies

Since in a pyproject.toml all the definitions are python dependencies if you need to specify some non-python dependencies as well this can be accomplished by adding the following sections to the pyproject.toml

# pyproject.toml

[tool.conda-lock.dependencies]
sqlite = ">=3.34"

pip dependencies

If a dependency refers directly to a URL rather than a package name and version, conda-lock will assume it is pip-installable, e.g.:

# pyproject.toml
[tool.poetry.dependencies]
python = "3.9"
pymage = {url = "https://github.com/MickaelRigault/pymage/archive/v1.0.tar.gz#sha256=11e99c4ea06b76ca7fb5b42d1d35d64139a4fa6f7f163a2f0f9cc3ea0b3c55eb"}

Similarly, if a dependency is explicitly marked with source = "pypi", it will be treated as a pip dependency, e.g.:

[tool.poetry.dependencies]
python = "3.9"
ampel-ztf = {version = "^0.8.0-alpha.2", source = "pypi"}

In both these cases, the dependencies of pip-installable packages will also be installed with pip, unless they were already requested by a conda dependency.

Dockerfile example

In order to use conda-lock in a docker-style context you want to add the lockfile to the docker container. In order to refresh the lock file just run conda-lock again.

Given a file tree like

  Dockerfile
  environment.yaml
* conda-linux-64.lock

You want a dockerfile that is structured something similar to this

# Dockerfile

# Build container
FROM continuumio/miniconda:latest as conda

ADD conda-linux-64.lock /locks/conda-linux-64.lock
RUN conda create -p /opt/env --copy --file /locks/conda-linux-64.lock

# Primary container

FROM gcr.io/distroless/base-debian10

COPY --from=conda /opt/env /opt/env