Variant configuration#
rattler-build
can automatically build multiple variants of a given package.
For example, a Python package might need multiple variants per Python version
(especially if it is a binary package such as numpy
).
For this use case, one can specify variant configuration files. A variant configuration file has 2 special entries and a list of packages with variants. For example:
# special entry #1, the zip keys
zip_keys:
- [python, numpy]
# special entry #2, the pin_run_as_build key
pin_run_as_build:
numpy:
max_pin: 'x.x'
# entries per package version that users are interested in
python:
# Note that versions are _strings_ (not numbers)
- "3.8"
- "3.9"
- "3.10"
numpy:
- "1.12"
- "1.12"
- "1.20"
If we have a recipe, that has a build
, host
or run
dependency on python
we will build multiple variants of this package, one for each configured
python
version ("3.8", "3.9" and "3.10").
For example:
... will be rendered as (for the first variant):
Note that variants are only applied if the requirement doesn't specify any
constraints. If the requirement would be python >3.8,<3.10
then the variant entry
would be ignored.
Automatic variants.yaml
discovery#
rattler-build
automatically includes the variant configuration from a variants.yaml
file next to a recipe.
Use the --ignore-recipe-variants
option to disable automatic discovery of variants.yaml
files next to the recipes.
To include a variant config file from another location or include multiple configuration files use the --variant-config
option:
rattler-build build --variant-config ~/user_variants.yaml --variant-config /opt/rattler-build/global_variants.yaml --recipe myrecipe.yaml
Package hash from variant#
You might have wondered what the role of the build string is. The build string is (if not explicitly set) computed from the variant configuration. It serves as a mechanism to discern different build configurations that produce a package with the same name and version.
The hash is computed by dumping all of the variant configuration values that are used by a given recipe into a JSON file, and then hashing that JSON file.
For example, in our python
example, we would get a variant configuration file that looks something like:
This JSON string is then hashed with the MD5 hash algorithm, and produces the hash.
For certain packages (such as Python packages) special rules exists, and the py<Major.Minor>
version is prepended to the hash, so that the final hash
would look something like py38h123123
.
Zip keys#
Zip keys modify how variants are combined. Usually, each variant key that has multiple entries is expanded to a build matrix. For example, if we have:
...then we obtain 4 variants for a recipe that uses both numpy
and python
:
However, if we use the zip_keys
and specify:
...then the versions are "zipped up" and we only get 2 variants. Note that
both python
and numpy
need to specify the exact same number of versions
to make this work.
The resulting variants with the zip applied are:
Pin run as build#
The pin_run_as_build
key allows the user to inject additional pins. Usually, the run_exports
mechanism is used to
specify constraints for runtime dependencies from build time dependencies, but pin_run_as_build
offers a mechanism
to override that if the package does not contain a run exports file.
For example:
If we now have a recipe that uses libcurl
in the host
and run
dependencies like:
During resolution, libcurl
might be evaluated to libcurl 8.0.1 h13284
. Our new runtime dependency then
looks like:
Prioritizing variants#
You might produce multiple variants for a package, but want to define a priority for a given variant. The variant with the highest priority would be the default package that is selected by the resolver.
There are two mechanisms to make this possible: mutex
packages and the down_prioritize_variant
option in the recipe.
The down_prioritize_variant
option#
Note
It is not always necessary to use the down_prioritize_variant
option - only if the solver has no other way to
prefer a given variant. For example, if you have a package that has multiple variants for different Python versions,
the solver will automatically prefer the variant with the highest Python version.
The down_prioritize_variant
option allows you to specify a variant that should be down-prioritized. For example:
build:
variant_config:
use_keys:
# use cuda from the variant config, e.g. to build multiple CUDA variants
- cuda
# this will down-prioritize the cuda variant versus other variants of the package
down_prioritize_variant: ${{ 1 if cuda else 0 }}
Mutex packages#
Another way to make sure the right variants are selected are "mutex" packages. A mutex package is a package that is mutually exclusive. We use the fact that only one package of a given name can be installed at a time (the solver has to choose).
A mutex package might be useful to make sure that all packages that depend on BLAS are compiled against the same BLAS implementation.
The mutex package will serve the purpose that "openblas
" and "mkl
" can never be installed at the same time.
We could define a BLAS mutex package like this:
And then the recipe.yaml
for the mutex
package could look like this:
package:
name: blas_mutex
version: 1.0
build:
string: ${{ blas_variant }}${{ hash }}_${{ build_number }}
variant_config:
# make sure that `openblas` is preferred over `mkl`
down_prioritize_variant: ${{ 1 if blas_variant == "mkl" else 0 }}
This will create two package: blas_mutex-1.0-openblas
and blas_mutex-1.0-mkl
.
Only one of these packages can be installed at a time because they share the same name.
The solver will then only select one of these two packages.
The blas
package in turn should have a run_export
for the blas_mutex
package, so that any package
that links against blas
also has a dependency on the correct blas_mutex
package:
package:
name: openblas
version: 1.0
requirements:
# any package depending on openblas should also depend on the correct blas_mutex package
run_export:
# Add a run export on _any_ version of the blas_mutex package whose build string starts with "openblas"
- blas_mutex * openblas*
Then the recipe of a package that wants to build two variants, one for openblas
and
one for mkl
could look like this: