puck/lix
0
0
Fork 0
forked from lix-project/lix
Code Pull requests Activity

Updated documentation

Browse source
This commit is contained in:
Eelco Dolstra 2019-05-26 15:59:50 +02:00 committed by Nick Van den Broeck
parent 412684f9dc
commit 8a6704d826
2 changed files with 235 additions and 328 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

555
doc/flakes/design.md
View file

@ -2,92 +2,83 @@
## Goals ## Goals
* To provide Nix repositories with an easy and standard way to * Standard and easy way for Nix repos to reference other Nix repos as
reference other Nix repositories. dependencies
* To allow such references to be queried and updated automatically. * Discoverability: Be able to query and update these references to Nix repos
automatically
* To provide a replacement for `nix-channel`, `NIX_PATH` and Hydra * To provide a replacement for `nix-channel`, `NIX_PATH` and Hydra jobset
jobset definitions. definitions
* To enable reproducible, hermetic evaluation of packages and NixOS * Reproducibility: Evaluate packages and NixOS configurations hermetic by
configurations. default
Things that we probably won't do in the initial iteration: Upcoming but not yet implemented:
* Sophisticated flake versioning, such as the ability to specify * Sophisticated flake versioning, such as the ability to specify version ranges
version ranges on dependencies. on dependencies.
* A way to specify the types of values provided by a flake. For the * A way to specify the types of values provided by a flake. For the most part,
most part, flakes can provide arbitrary Nix values, but there will flakes can provide arbitrary Nix values, but there will be some standard
be some standard attribute names (e.g. `packages` must be a set of attribute names (e.g. `packages` must be a set of installable derivations).
installable derivations).
## Overview ## Overview
* A flake is (usually) a Git repository that contains a file named * A flake is (usually) a Git repository that contains a file named `flake.nix`
`flake.nix` at top-level. at top-level
* Flakes *provide* an attribute set of values, such as packages, * A flake *provides* an attribute set of values, such as packages, Nixpkgs
Nixpkgs overlays, NixOS modules, library functions, Hydra jobs, overlays, NixOS modules, library functions, Hydra jobs, `nix-shell`
`nix-shell` definitions, etc. definitions, etc.
* Flakes can *depend* on other flakes. * Flakes can *depend* on other flakes or other repositories which aren't flakes
* Flakes are referred to using a *flake reference*, which is either a * Flakes are referred to using a *flake reference*, which is either a URL
URL specifying its repository's location specifying its repository's location or an identifier looked up in a *lock
(e.g. `github:NixOS/nixpkgs/release-18.09`) or an identifier file* or *flake registry*.
(e.g. `nixpkgs`) looked up in a *lock file* or *flake
registry*. They can also specify revisions,
e.g. `github:NixOS/nixpkgs/98a2a5b5370c1e2092d09cb38b9dcff6d98a109f`.
* The *flake registry* is a centrally maintained mapping (on * A *flake registry* is a mapping from flake identifiers to flake locations
`nixos.org`) from flake identifiers to flake locations (e.g. `nixpkgs -> github:NixOS/nixpkgs/release-18.09`). There is a centrally
(e.g. `nixpkgs -> github:NixOS/nixpkgs/release-18.09`). maintained flake registry on `nixos.org`.
* A flake can contain a *lock file* (`flake.lock`) used when resolving * A flake can contain a *lock file* (`flake.lock`) used when resolving the
the dependencies in `flake.nix`. It maps flake references to dependencies in `flake.nix`. It maps mutable flake references
references containing revisions (e.g. `nixpkgs -> (e.g. `github:NixOS/nixpkgs/release-18.09`) to references containing revisions
(e.g. `nixpkgs ->
github:NixOS/nixpkgs/98a2a5b5370c1e2092d09cb38b9dcff6d98a109f`). github:NixOS/nixpkgs/98a2a5b5370c1e2092d09cb38b9dcff6d98a109f`).
* The `nix` command uses the flake registry as its default * The `nix` command uses the flake registry as its default installation source.
installation source. For example, `nix build nixpkgs.hello` builds the For example, `nix build nixpkgs.hello` builds the `hello` package provided by
`hello` package provided by the `nixpkgs` flake listed in the the `nixpkgs` flake listed in the registry. `nix` will automatically
registry. `nix` will automatically download/upload the registry and download/upload the registry and flakes as needed.
flakes as needed.
* `nix build` without arguments will build the flake in the current * `nix build` without arguments will build the flake in the current
directory (or some parent). directory (or some parent).
* The command `nix flake update` generates/updates `flake.lock` from * `nix flake update` generates `flake.lock` from `flake.nix`, ignoring the old
`flake.nix`. This should probably also be done automatically when lockfile.
building from a local flake.
* `nixos-rebuild` will build a configuration from a (locked) * `nixos-rebuild` will build a configuration from a (locked) flake. Evaluation
flake. Evaluation will be done in pure mode to ensure there are no is done in pure mode to ensure there are no unaccounted inputs. Thus the
unaccounted inputs. Thus the NixOS configuration can be reproduced NixOS configuration can be reproduced unambiguously from the top-level flake.
unambiguously from the top-level flake.
* Nix code can query flake metadata such as `commitHash` (the Git * Nix code can query flake metadata such as `commitHash` (the Git revision) or
revision) or `date` (the date of the last commit). This is useful `epoch` (the date of the last commit). This is useful for NixOS to compute
for NixOS to compute the NixOS version string (which will be the the NixOS version string (which will be the revision of the top-level
revision of the top-level configuration flake, uniquely identifying configuration flake, uniquely identifying the configuration).
the configuration).
* Hydra jobset configurations will consist of a single flake * Hydra jobset configurations will consist of a single flake reference. Thus we
reference. Thus we can get rid of jobset inputs; any other needed can get rid of jobset inputs; any other needed repositories can be fetched by
repositories can be fetched by the top-level flake. The top-level the top-level flake. The top-level flake can be locked or unlocked; if some
flake can be locked or unlocked; if some dependencies are unlocked, dependencies are unlocked, then Nix will fetch the latest revision for each.
then Nix will fetch the latest revision for each.
## Example flake ## Example flake
A flake is a Git repository that contains a file named Let us look at an example of a `flake.nix` file, here for `dwarffs`, a small
`flake.nix`. For example, here is the `flake.nix` for `dwarffs`, a repository that provides a single package and a single NixOS module.
small repository that provides a single package and a single NixOS
module.
```nix ```nix
{ {
@ -101,23 +92,26 @@ module.
# Some other metadata. # Some other metadata.
description = "A filesystem that fetches DWARF debug info from the Internet on demand"; description = "A filesystem that fetches DWARF debug info from the Internet on demand";
# A list of flake references denoting the flakes that this flake # The flake dependencies. Nix will resolve and fetch these flakes and pass
# depends on. Nix will resolve and fetch these flakes and pass them # them as a function argument to `outputs` below.
# as a function argument to `outputs` below.
# #
# `flake:nixpkgs` denotes a flake named `nixpkgs` which is looked up # "nixpkgs" denotes a flake named `nixpkgs` which is looked up
# in the flake registry, or in `flake.lock` inside this flake, if it # in the flake registry, or in `flake.lock` inside this flake, if it
# exists. # exists.
inputs = [ flake:nixpkgs ]; inputs = [ flake:nixpkgs ];
# An attribute set listing dependencies which aren't flakes, also to be passed as
# a function argument to `provides`.
nonFlakeRequires = {};
# The stuff provided by this flake. Flakes can provide whatever they # The stuff provided by this flake. Flakes can provide whatever they
# want (convention over configuration), but some attributes have # want (convention over configuration), but some attributes have
# special meaning to tools / other flakes: for example, `packages` # special meaning to tools / other flakes. For example, `packages`
# is used by the `nix` CLI to search for packages, and # is used by the `nix` CLI to search for packages, and
# `nixosModules` is used by NixOS to automatically pull in the # `nixosModules` is used by NixOS to automatically pull in the
# modules provided by a flake. # modules provided by a flake.
# #
# `outputs` takes a single argument named `deps` that contains # `outputs` takes a single argument (`deps`) that contains
# the resolved set of flakes. (See below.) # the resolved set of flakes. (See below.)
outputs = deps: { outputs = deps: {
@ -153,7 +147,11 @@ module.
nixosModules.dwarffs = import ./module.nix deps; nixosModules.dwarffs = import ./module.nix deps;
# Provide a single Hydra job (`hydraJobs.dwarffs`). # Provide a single Hydra job (`hydraJobs.dwarffs`).
hydraJobs = deps.this.packages; hydraJobs.build.x86_64-linux = packages.dwarffs;
# A bunch of things which can be checked (through `nix flake check`) to
# make sure the flake is well-defined.
checks.build = packages.dwarffs;
}; };
} }
``` ```
@ -170,8 +168,11 @@ Similarly, a minimal `flake.nix` for Nixpkgs:
outputs = deps: outputs = deps:
let pkgs = import ./. {}; in let pkgs = import ./. {}; in
let pkgs = import ./. { system = "x86_64-linux"; }; in
{ {
lib = import ./lib; lib = (import ./lib) // {
nixosSystem = import ./nixos/lib/eval-config.nix;
};
builders = { builders = {
inherit (pkgs) stdenv fetchurl; inherit (pkgs) stdenv fetchurl;
@ -180,145 +181,124 @@ Similarly, a minimal `flake.nix` for Nixpkgs:
packages = { packages = {
inherit (pkgs) hello nix fuse nlohmann_json boost; inherit (pkgs) hello nix fuse nlohmann_json boost;
}; };
legacyPkgs = pkgs;
}; };
} }
``` ```
Note that `packages` is an unpolluted set of packages: non-package Note that `packages` is an unpolluted set of packages: non-package values like
values like `lib` or `fetchurl` are not part of it. `lib` or `fetchurl` are not part of it.
## Flake registries
## Flake identifiers Note: If a flake registry contains an entry `nixpkgs -> github:NixOS/nixpkgs`,
then `nixpkgs/release-18.09` will match to become
A flake has an identifier (e.g. `nixpkgs` or `dwarffs`). `github:NixOS/nixpkgs/release-18.09`. This is referred to as "fuzzymatching".
## Flake references ## Flake references
Flake references are a URI-like syntax to specify the physical Flake references are a URI-like syntax to specify the physical location of a
location of a flake (e.g. a Git repository) or to denote a lookup in flake (e.g. a Git repository) or to denote a lookup in the flake registry or
the flake registry or lock file. lock file. There are four options for the syntax:
* `(flake:)?<flake-id>(/rev-or-ref(/rev)?)?` * Flake aliases
A flake alias is a name which requires a lookup in a flake
registry or lock file.
Look up a flake by ID in the flake lock file or in the flake Example: "nixpkgs"
registry. These must specify an actual location for the flake using
the formats listed below. Note that in pure evaluation mode, the
flake registry is empty.
Optionally, the `rev` or `ref` from the dereferenced flake can be * GitHub repositories
overriden. For example, A repository which is stored on GitHub can easily be fetched using this type.
Note:
* Only the code in this particular commit is downloaded, not the entire repo
* By default, the commit to download is the last commit on the `master` branch.
See later for how to change this.
> nixpkgs/19.09 Example: `github:NixOS/nixpkgs`
uses the `19.09` branch of the `nixpkgs` flake's GitHub repository, * `ssh/https/git/file`
while These are generic `FlakeRef`s for downloadding git repositories or tarballs.
> nixpkgs/98a2a5b5370c1e2092d09cb38b9dcff6d98a109f
uses the specified revision. For Git (rather than GitHub)
repositories, both the rev and ref must be given, e.g.
> nixpkgs/19.09/98a2a5b5370c1e2092d09cb38b9dcff6d98a109f
* `github:<owner>/<repo>(/<rev-or-ref>)?`
A repository on GitHub. These differ from Git references in that
they're downloaded in a efficient way (via the tarball mechanism)
and that they support downloading a specific revision without
specifying a branch. `rev-or-ref` is either a commit hash (`rev`)
or a branch or tag name (`ref`). The default is `master` if none is
specified. Note that in pure evaluation mode, a commit hash must be
used.
Flakes fetched in this manner expose `rev` and `date` attributes,
but not `revCount`.
Examples:
> github:edolstra/dwarffs
> github:edolstra/dwarffs/unstable
> github:edolstra/dwarffs/41c0c1bf292ea3ac3858ff393b49ca1123dbd553
* > https://<server>/<path>.git(\?attr(&attr)*)?
> ssh://<server>/<path>.git(\?attr(&attr)*)?
> git://<server>/<path>.git(\?attr(&attr)*)?
> file:///<path>(\?attr(&attr)*)?
where `attr` is one of `rev=<rev>` or `ref=<ref>`.
A Git repository fetched through https. Note that the path must end
in `.git`. The default for `ref` is `master`.
Examples:
> https://example.org/my/repo.git
> https://example.org/my/repo.git?ref=release-1.2.3
> https://example.org/my/repo.git?rev=e72daba8250068216d79d2aeef40d4d95aff6666
* > /path.git(\?attr(&attr)*)?
Like `file://path.git`, but if no `ref` or `rev` is specified, the
(possibly dirty) working tree will be used. Using a working tree is
not allowed in pure evaluation mode.
Examples: Examples:
- https://example.org/my/repo.git
- ssh://git@github.com:NixOS/nix.git
- git://github.com/edolstra/dwarffs.git
- file:///home/my-user/some-repo/some-repo.git
- https://releases.nixos.org/nixos/unstable/nixos-19.03pre167858.f2a1a4e93be/nixexprs.tar.xz
- file:///<path>.tar.xz
> /path/to/my/repo * Local, dirty paths
This `FlakeRef` is the equivalent of `file://<path>` used for dirty paths.
> /path/to/my/repo?ref=develop Example: /path/to/my/repo
> /path/to/my/repo?rev=e72daba8250068216d79d2aeef40d4d95aff6666 Notes:
- Each FlakeRef (except for the Path option) allows for a Git revision (i.e.
* > https://<server>/<path>.tar.xz(?hash=<sri-hash>) commit hash) and/or referenceo(i.e. git branch name) to be added. For
tarbals, an SRI hash needs to be added.
> file:///<path>.tar.xz(?hash=<sri-hash>) Examples:
* `"nixpkgs/release-18.09"`
A flake distributed as a tarball. In pure evaluation mode, an SRI * `github:NixOS/nixpkgs/1e9e709953e315ab004951248b186ac8e2306451`
hash is mandatory. It exposes a `date` attribute, being the newest * `git://github.com/edolstra/dwarffs.git?ref=flake&rev=2efca4bc9da70fb001b26c3dc858c6397d3c4817`
file inside the tarball. * file:///<path>.tar.xz(?hash=<sri-hash>)
- In full pure mode, no mutable `FlakeRef`s can be used
Example: * No aliases, because they need to be looked up
* `github` requires a specified `rev`
> https://releases.nixos.org/nixos/unstable/nixos-19.03pre167858.f2a1a4e93be/nixexprs.tar.xz * `ssh/https/git/file` require a specified `ref` _and_ `rev`
* `path` is always mutable
> https://releases.nixos.org/nixos/unstable/nixos-19.03pre167858.f2a1a4e93be/nixexprs.tar.xz?hash=sha256-56bbc099995ea8581ead78f22832fee7dbcb0a0b6319293d8c2d0aef5379397c - Flakes don't need to be top-level, but can also reside in a subdirectory. This is shown by adding `dir=<subdir>` to the `FlakeRef`.
Example: `./foo?dir=bar`
Note: currently, there can be only one flake per Git repository, and
it must be at top-level. In the future, we may want to add a field
(e.g. `dir=<dir>`) to specify a subdirectory inside the repository.
## Flake lock files ## Flake lock files
This is a JSON file named `flake.lock` that maps flake identifiers A lockfile is a JSON file named `flake.lock` which contains a forrest of
used in the corresponding `flake.nix` to "immutable" flake references; entries mapping `FlakeRef`s to the immutable `FlakeRef` they were resolved to.
that is, flake references that contain a revision (for Git
repositories) or a content hash (for tarballs).
Example: Example:
```json ```json
{ {
"nixpkgs": "github:NixOS/nixpkgs/41c0c1bf292ea3ac3858ff393b49ca1123dbd553", "nixpkgs": {
"foo": "https://example.org/foo.tar.xz?hash=sha256-56bbc099995ea8581ead78f22832fee7dbcb0a0b6319293d8c2d0aef5379397c" "uri": "github:NixOS/nixpkgs/41c0c1bf292ea3ac3858ff393b49ca1123dbd553",
"content-hash": "sha256-vy2UmXQM66aS/Kn2tCtjt9RwxfBvV+nQVb5tJQFwi8E="
},
"foo": {
"uri": "https://example.org/foo.tar.xz?hash=sha256-56bbc099995ea8581ead78f22832fee7dbcb0a0b6319293d8c2d0aef5379397c",
"content-hash": "sha256-vy2UmXQM66aS/Kn2tCtjt9RwxfBvV+nQVb5tJQFwi8E="
}
} }
``` ```
Lockfiles are used to help resolve the dependencies of a flake.
- `nix build github:<..>` uses the remote lockfile and update it
- `nix build /home/user/dwarffs` uses the local lockfile, updates it and writes the result to file
- `nix flake update <flakeref>` recreates the lockfile from scratch and writes it to file
- `--no-registries` makes the command pure, also when fetching dependencies
- `--no-save-lock-file`: Several commands will update the lockfile (e.g. `nix
build`). This flag prevents the updated lockfile to be written to file.
- `--recreate-lock-file` makes prevents the current lockfile from being used
## `outputs` ## `outputs`
The flake attribute `outputs` is a function that takes an argument The function argument `deps` is an attrset containing all dependencies listed
named `deps` and returns a (mostly) arbitrary attrset of values. Some in `requires` and `nonFlakeRequires` as well as `path` (for the flake's source
of the standard result attributes: code) and an attribute `meta` with:
- `description`
- `commitHash` (not for tarball flakes): The Git commit hash.
- `date`: The timestamp of the most recent commit (for Git repos), or of the
most recently modified file (for tarballs)
- `revCount` (for Git flakes, but not GitHub flakes): The number of ancestors
of the revision. Useful for generating version strings.
* `packages`: A set of installable derivations used by the `nix` The flake attribute `outputs` is a function that takes an argument named `deps`
command. That is, commands such as `nix install` ignore all other and returns an attribute set. Some of the members of this set have protected
flake attributes. names:
* `packages`: A set of installable derivations used by the `nix` command. That
is, commands such as `nix install` ignore all other flake attributes. It
cannot be a nested set.
* `hydraJobs`: Used by Hydra. * `hydraJobs`: Used by Hydra.
@ -329,213 +309,155 @@ of the standard result attributes:
we need to avoid a situation where `nixos-rebuild` needs to fetch we need to avoid a situation where `nixos-rebuild` needs to fetch
its own `nixpkgs` just to do `evalModules`.) its own `nixpkgs` just to do `evalModules`.)
* `devShell`: A specification of a development environment in some TBD * `devShell`: A derivation to create a development environment
format.
The function argument `flakes` is an attrset that contains an * `self`: The result of the flake's output which is passed to itself
attribute for each dependency specified in `inputs`. (Should it Example: `self.outputs.foo` works.
contain transitive dependencies? Probably not.) Each attribute is an
attrset containing the `outputs` of the dependency, in addition to
the following attributes:
* `path`: The path to the flake's source code. Useful when you want to
use non-Nix artifacts from the flake, or if you want to *store* the
source code of the dependency in a derivation. (For example, we
could store the sources of all flake dependencies in a NixOS system
configuration, as a generalization of
`system.copySystemConfiguration`.)
* `meta`: An attrset containing the following:
* `description`
* `commitHash` (or `rev`?) (not for tarball flakes): The Git commit
hash.
* `date`: The timestamp of the most recent commit (for Git
repositories), or the timestamp of the most recently modified file
(for tarballs).
* `revCount` (for Git flakes, but not GitHub flakes): The number of
ancestors of the revision. Useful for generating version strings.
## Non-flake dependencies
It may be useful to pull in repositories that are not flakes
(i.e. don't contain a `flake.nix`). This could be done in two ways:
* Allow flakes not to have a `flake.nix` file, in which case it's a
flake with no inputs and no outputs. The downside of this
approach is that we can't detect accidental use of a non-flake
repository. (Also, we need to conjure up an identifier somehow.)
* Add a flake attribute to specifiy non-flake dependencies, e.g.
> nonFlakeInputs.foobar = github:foo/bar;
## Flake registry ## Flake registry
The flake registry maps flake IDs to flake references (where the A flake registry is a JSON file mapping flake references to flake references.
latter cannot be another indirection, i.e. it must not be a The default/global registry is kept at
`flake:<flake-id>` reference). `https://github.com/NixOS/flake-registry/blob/master/flake-registry.json` and
looks like this:
The default registry is kept at
`https://nixos.org/flake-registry.json`. It looks like this:
```json ```json
{ {
"version": 1,
"flakes": { "flakes": {
"dwarffs": { "dwarffs": {
"uri": "github:edolstra/dwarffs/flake" "uri": "github:edolstra/dwarffs/flake"
}, },
"nix": {
"uri": "github:NixOS/nix/flakes"
},
"nixpkgs": { "nixpkgs": {
"uri": "github:NixOS/nixpkgs/release-18.09" "uri": "github:edolstra/nixpkgs/release-19.03"
},
"hydra": {
"uri": "github:NixOS/hydra/flake"
},
"patchelf": {
"uri": "github:NixOS/patchelf"
} }
} },
"version": 1
} }
``` ```
Nix automatically (re)downloads the registry. The downloaded file is a Nix automatically (re)downloads this file whenever you have network access. The
GC root so the registry remains available if nixos.org is unreachable. downloaded file is a GC root so the registry remains available if nixos.org is
TBD: when to redownload? unreachable.
In addition to a global registry, there is also a user registry stored in
`~/.config/nix/registry.json`.
## Nix UI ## Nix UI
Commands for registry / user flake configuration: There is a list of new commands added to the `nix` CLI:
* `nix flake list`: Show all flakes in the registry. * `nix flake list`: Show all flakes in the registry
* `nix flake add <flake-ref>`: Add or override a flake to/in the * `nix flake add <alias FlakeRef> <resolved FlakeRef>`: Add or override a flake
user's flake configuration (`~/.config/nix/flakes.nix`). For to/in the user flake registry.
example, `nix flake add nixpkgs/nixos-18.03` overrides the `nixpkgs`
flake to use the `nixos-18.03` branch. There should also be a way to
add multiple branches/revisions of the same flake by giving them a
different ID, e.g. `nix flake add --id nixpkgs-ancient
nixpkgs/nixos-16.03`).
* `nix flake remove <flake-id>`: Remove a flake from the user's flake * `nix flake remove <alias FlakeRef>`: Remove a FlakeRef from the user flake
configuration. Any flake with the same ID in the registry remains registry.
available.
* `nix flake lock <flake-id>`: Lock a flake. For example, `nix flake * `nix flake pin <alias FlakeRef>`: Look up to which immutable FlakeRef the
lock nixpkgs` pins `nixpkgs` to the current revision. alias FlakeRef maps to currently, and store that map in the user registry.
Example: `nix flake pin github:NixOS/nixpkgs` will create an entry
`github:NixOS/nixpkgs ->
github:NixOS/nixpkgs/444f22ca892a873f76acd88d5d55bdc24ed08757`.
Commands for creating/modifying a flake: * `nix flake init`: Create a `flake.nix` in the current directory
* `nix flake init`: Create a `flake.nix` in the current directory. * `nix flake update`: Recreate the lock file from scratch, from the `flake.nix`.
* `nix flake update`: Update the lock file for the `flake.nix` in the
current directory. In most cases, this should be done
automatically. (E.g. `nix build` should automatically update the
lock file is a new dependency is added to `flake.nix`.)
* `nix flake check`: Do some checks on the flake, e.g. check that all * `nix flake check`: Do some checks on the flake, e.g. check that all
`packages` are really packages. `packages` are really packages.
* `nix flake clone`: Do a Git clone of the flake repository. This is a * `nix flake clone`: `git clone` the flake repo
convenience to easily start hacking on a flake. E.g. `nix flake
clone dwarffs` clones the `dwarffs` GitHub repository to `./dwarffs`.
TODO: maybe the first set of commands should have a different name
from the second set.
Flags / configuration options: Flags / configuration options:
* `--flakes (<flake-id>=<flake-ref>)*`: add/override some flakes. * `--flakes (<alias FlakeRef>=<resolved FlakeRef>)*`: add/override some
FlakeRef
* (In `nix`) `--flake <flake-ref>`: set the specified flake as the * `--flake <flake-ref>`: set the specified flake as the installation source
installation source. E.g. `nix build --flake ./my-nixpkgs hello`. E.g. `nix build --flake ./my-nixpkgs hello`.
The default installation source in `nix` is the `packages` from all The default installation source in `nix` is the `packages` from all flakes in
flakes in the registry, that is: the registry, that is:
``` ```
builtins.mapAttrs (flakeName: flakeInfo: builtins.mapAttrs (flakeName: flakeInfo:
(getFlake flakeInfo.uri).${flakeName}.outputs.packages or {}) (getFlake flakeInfo.uri).${flakeName}.outputs.packages or {})
builtins.flakeRegistry builtins.flakeRegistry
``` ```
(where `builtins.flakeRegistry` is the global registry with user where `builtins.flakeRegistry` is the global registry with user overrides
overrides applied, and `builtins.getFlake` downloads a flake and applied, and `builtins.getFlake` downloads a flake and resolves its
resolves its dependencies.) dependencies.
It may be nice to extend the default installation source with the
`packages` from the flake in the current directory, so that
> nix build hello
does something similar to the old
> nix-build -A hello
Specifically, it builds `packages.hello` from the flake in the current
directory. Of course, this creates some ambiguity if there is a flake
in the registry named `hello`.
Maybe the command
> nix dev-shell
should do something like use `outputs.devShell` to initialize the
shell, but probably we should ditch `nix shell` / `nix-shell` for
direnv.
## Pure evaluation and caching ## Pure evaluation and caching
Flake evaluation should be done in pure mode. Thus: Flake evaluation is done in pure mode. Thus:
* Flakes cannot do `NIX_PATH` lookups via the `<...>` syntax. * Flakes cannot use `NIX_PATH` via the `<...>` syntax.
* They can't read random stuff from non-flake directories, such as * Flakes cannot read random stuff from non-flake directories, such as
`~/.nix/config.nix` or overlays. `~/.nix/config.nix` or overlays.
This enables aggressive caching or precomputation of Nixpkgs package This enables aggressive caching or precomputation of Nixpkgs package sets. For
sets. For example, for a particular Nixpkgs flake closure (as example, for a particular Nixpkgs flake closure (as identified by, say, a hash
identified by, say, a hash of the fully-qualified flake references of the fully-qualified flake references after dependency resolution) and system
after dependency resolution) and system type, an attribute like type, an attribute like `packages.hello` should always evaluate to the same
`packages.hello` should always evaluate to the same derivation. So we derivation. So we can:
can:
* Keep a local evaluation cache (say `~/.cache/nix/eval.sqlite`) * Keep a local evaluation cache (say `~/.cache/nix/eval-cache-v1.sqlite`)
mapping `(<flake-closure-hash, <attribute>) -> (<drv-name>, mapping `(<flake-closure-hash, <attribute>) -> (<drv-name>,
<drv-output-paths>, <whatever other info we want to cache>)`. <drv-output-paths>, <whatever other info we want to cache>)`.
* Download a precomputed cache * Download a precomputed cache, e.g.
(e.g. `https://releases.nixos.org/eval/<flake-closure-hash>.sqlite`). So `https://releases.nixos.org/eval/<flake-closure-hash>.sqlite`. So a command
a command like `nix search` could avoid evaluating Nixpkgs entirely. like `nix search` could avoid evaluating Nixpkgs entirely.
Of course, this doesn't allow overlays. With pure evaluation, the only Of course, this doesn't allow overlays. With pure evaluation, the only way to
way to have these is to define a top-level flake that depends on the have these is to define a top-level flake that depends on the Nixpkgs flake and
Nixpkgs flake and somehow passes in a set of overlays. somehow passes in a set of overlays.
TODO: in pure mode we have to pass the system type explicitly!
## Hydra jobset dependencies ## Hydra jobset dependencies
Hydra can use the flake dependency resolution mechanism to fetch Hydra can use the flake dependency resolution mechanism to fetch dependencies.
dependencies. This allows us to get rid of jobset configuration in the This allows us to get rid of jobset configuration in the web interface: a
web interface: a jobset only requires a flake reference. That is, *a jobset only requires a flake reference. That is, a jobset *is* a flake. Hydra
jobset is a flake*. Hydra then just builds the `hydraJobs` attrset then just builds the `hydraJobs` attrset
`provide`d by the flake. (It omitted, maybe it can build `packages`.)
## NixOS system configuration ## NixOS system configuration
NixOS currently contains a lot of modules that really should be moved NixOS currently contains a lot of modules that really should be moved into
into their own repositories. For example, it contains a Hydra module their own repositories. For example, it contains a Hydra module that duplicates
that duplicates the one in the Hydra repository. Also, we want the one in the Hydra repository. Also, we want reproducible evaluation for
reproducible evaluation for NixOS system configurations. So NixOS NixOS system configurations. So NixOS system configurations should be stored as
system configurations should be stored as flakes in (local) Git flakes in (local) Git repositories.
repositories.
`my-system/flake.nix`: `my-system/flake.nix`:
```nix ```nix
{ {
name = "my-system";
epoch = 201906;
inputs =
[ "nixpkgs/nixos-18.09"
"dwarffs"
"hydra"
... lots of other module flakes ...
];
outputs = flakes: { outputs = flakes: {
nixosSystems.default = nixosSystems.default =
flakes.nixpkgs.lib.evalModules { flakes.nixpkgs.lib.evalModules {
@ -549,13 +471,6 @@ repositories.
]; ];
}; };
}; };
inputs =
[ "nixpkgs/nixos-18.09"
"dwarffs"
"hydra"
... lots of other module flakes ...
];
} }
``` ```
@ -563,5 +478,5 @@ We can then build the system:
``` ```
nixos-rebuild switch --flake ~/my-system nixos-rebuild switch --flake ~/my-system
``` ```
This performs dependency resolution starting at `~/my-system/flake.nix` This performs dependency resolution starting at `~/my-system/flake.nix` and
and builds the `system` attribute in `nixosSystems.default`. builds the `system` attribute in `nixosSystems.default`.

8
src/nix/flake.cc
View file

@ -128,14 +128,6 @@ static void printNonFlakeInfo(const NonFlake & nonFlake)
printSourceInfo(nonFlake.sourceInfo); printSourceInfo(nonFlake.sourceInfo);
} }
static nlohmann::json nonFlakeToJson(const NonFlake & nonFlake)
{
nlohmann::json j;
j["id"] = nonFlake.alias;
sourceInfoToJson(nonFlake.sourceInfo, j);
return j;
}
// FIXME: merge info CmdFlakeInfo? // FIXME: merge info CmdFlakeInfo?
struct CmdFlakeDeps : FlakeCommand struct CmdFlakeDeps : FlakeCommand
{ {