Syntax is taken from MyST: https://myst-parser.readthedocs.io/en/latest/using/syntax.html#targets-and-cross-referencing
9.7 KiB
Multiple-output packages
Introduction
The Nix language allows a derivation to produce multiple outputs, which is similar to what is utilized by other Linux distribution packaging systems. The outputs reside in separate Nix store paths, so they can be mostly handled independently of each other, including passing to build inputs, garbage collection or binary substitution. The exception is that building from source always produces all the outputs.
The main motivation is to save disk space by reducing runtime closure sizes; consequently also sizes of substituted binaries get reduced. Splitting can be used to have more granular runtime dependencies, for example the typical reduction is to split away development-only files, as those are typically not needed during runtime. As a result, closure sizes of many packages can get reduced to a half or even much less.
::: {.note} The reduction effects could be instead achieved by building the parts in completely separate derivations. That would often additionally reduce build-time closures, but it tends to be much harder to write such derivations, as build systems typically assume all parts are being built at once. This compromise approach of single source package producing multiple binary packages is also utilized often by rpm and deb. :::
A number of attributes can be used to work with a derivation with multiple outputs. The attribute outputs
is a list of strings, which are the names of the outputs. For each of these names, an identically named attribute is created, corresponding to that output. The attribute meta.outputsToInstall
is used to determine the default set of outputs to install when using the derivation name unqualified.
Installing a split package
When installing a package with multiple outputs, the package’s meta.outputsToInstall
attribute determines which outputs are actually installed. meta.outputsToInstall
is a list whose default installs binaries and the associated man pages. The following sections describe ways to install different outputs.
Selecting outputs to install via NixOS
NixOS provides two ways to select the outputs to install for packages listed in environment.systemPackages
:
-
The configuration option
environment.extraOutputsToInstall
is appended to each package’smeta.outputsToInstall
attribute to determine the outputs to install. It can for example be used to installinfo
documentation or debug symbols for all packages. -
The outputs can be listed as packages in
environment.systemPackages
. For example, the"out"
and"info"
outputs for thecoreutils
package can be installed by includingcoreutils
andcoreutils.info
inenvironment.systemPackages
.
Selecting outputs to install via nix-env
nix-env
lacks an easy way to select the outputs to install. When installing a package, nix-env
always installs the outputs listed in meta.outputsToInstall
, even when the user explicitly selects an output.
::: {.warning}
nix-env
silenty disregards the outputs selected by the user, and instead installs the outputs from meta.outputsToInstall
. For example,
$ nix-env -iA nixpkgs.coreutils.info
installs the "out"
output (coreutils.meta.outputsToInstall
is [ "out" ]
) instead of the requested "info"
.
:::
The only recourse to select an output with nix-env
is to override the package’s meta.outputsToInstall
, using the functions described in . For example, the following overlay adds the "info"
output for the coreutils
package:
self: super:
{
coreutils = super.coreutils.overrideAttrs (oldAttrs: {
meta = oldAttrs.meta // { outputsToInstall = oldAttrs.meta.outputsToInstall or [ "out" ] ++ [ "info" ]; };
});
}
Using a split package
In the Nix language the individual outputs can be reached explicitly as attributes, e.g. coreutils.info
, but the typical case is just using packages as build inputs.
When a multiple-output derivation gets into a build input of another derivation, the dev
output is added if it exists, otherwise the first output is added. In addition to that, propagatedBuildOutputs
of that package which by default contain $outputBin
and $outputLib
are also added. (See .)
In some cases it may be desirable to combine different outputs under a single store path. A function symlinkJoin
can be used to do this. (Note that it may negate some closure size benefits of using a multiple-output package.)
Writing a split derivation
Here you find how to write a derivation that produces multiple outputs.
In nixpkgs there is a framework supporting multiple-output derivations. It tries to cover most cases by default behavior. You can find the source separated in <nixpkgs/pkgs/build-support/setup-hooks/multiple-outputs.sh>
; it’s relatively well-readable. The whole machinery is triggered by defining the outputs
attribute to contain the list of desired output names (strings).
outputs = [ "bin" "dev" "out" "doc" ];
Often such a single line is enough. For each output an equally named environment variable is passed to the builder and contains the path in nix store for that output. Typically you also want to have the main out
output, as it catches any files that didn’t get elsewhere.
::: {.note}
There is a special handling of the debug
output, described at .
:::
“Binaries first”
A commonly adopted convention in nixpkgs
is that executables provided by the package are contained within its first output. This convention allows the dependent packages to reference the executables provided by packages in a uniform manner. For instance, provided with the knowledge that the perl
package contains a perl
executable it can be referenced as ${pkgs.perl}/bin/perl
within a Nix derivation that needs to execute a Perl script.
The glibc
package is a deliberate single exception to the “binaries first” convention. The glibc
has libs
as its first output allowing the libraries provided by glibc
to be referenced directly (e.g. ${stdenv.glibc}/lib/ld-linux-x86-64.so.2
). The executables provided by glibc
can be accessed via its bin
attribute (e.g. ${stdenv.glibc.bin}/bin/ldd
).
The reason for why glibc
deviates from the convention is because referencing a library provided by glibc
is a very common operation among Nix packages. For instance, third-party executables packaged by Nix are typically patched and relinked with the relevant version of glibc
libraries from Nix packages (please see the documentation on patchelf for more details).
File type groups
The support code currently recognizes some particular kinds of outputs and either instructs the build system of the package to put files into their desired outputs or it moves the files during the fixup phase. Each group of file types has an outputFoo
variable specifying the output name where they should go. If that variable isn’t defined by the derivation writer, it is guessed – a default output name is defined, falling back to other possibilities if the output isn’t defined.
$outputDev
is for development-only files. These include C(++) headers (include/
), pkg-config (lib/pkgconfig/
), cmake (lib/cmake/
) and aclocal files (share/aclocal/
). They go to dev
or out
by default.
$outputBin
is meant for user-facing binaries, typically residing in bin/
. They go to bin
or out
by default.
$outputLib
is meant for libraries, typically residing in lib/
and libexec/
. They go to lib
or out
by default.
$outputDoc
is for user documentation, typically residing in share/doc/
. It goes to doc
or out
by default.
$outputDevdoc
is for developer documentation. Currently we count gtk-doc and devhelp books, typically residing in share/gtk-doc/
and share/devhelp/
, in there. It goes to devdoc
or is removed (!) by default. This is because e.g. gtk-doc tends to be rather large and completely unused by nixpkgs users.
$outputMan
is for man pages (except for section 3), typically residing in share/man/man[0-9]/
. They go to man
or $outputBin
by default.
$outputDevman
is for section 3 man pages, typically residing in share/man/man[0-9]/
. They go to devman
or $outputMan
by default.
$outputInfo
is for info pages, typically residing in share/info/
. They go to info
or $outputBin
by default.
Common caveats
-
Some configure scripts don’t like some of the parameters passed by default by the framework, e.g.
--docdir=/foo/bar
. You can disable this by settingsetOutputFlags = false;
. -
The outputs of a single derivation can retain references to each other, but note that circular references are not allowed. (And each strongly-connected component would act as a single output anyway.)
-
Most of split packages contain their core functionality in libraries. These libraries tend to refer to various kind of data that typically gets into
out
, e.g. locale strings, so there is often no advantage in separating the libraries intolib
, as keeping them inout
is easier. -
Some packages have hidden assumptions on install paths, which complicates splitting.