nix/doc/manual/source/store/drv.md

Derivation and Deriving Path

So far, we have covered "inert" store objects. But the point of the Nix store layer is to be a build system. Other system (like Git or IPFS) also store and transfer immutable data, but they don't concern themselves with how that data was created.

This is where Nix distinguishes itself. Derivations represent individual build steps, and deriving paths are needed to refer to the outputs of those build steps before they are built.

Derivation

A derivation is a specification for running an executable on precisely defined input files to repeatably produce output files at uniquely determined file system paths.

What is the natural Unix analog for a build step in action? Answer: a process that will eventually exit, leaving behind some output files. What is the natural way to plan such a step? An execve system call.

A derivation consists of:

• A name

• A set of inputs, a set of deriving paths

• A map of outputs, from names to other data

• The process creation fields: to spawn the arbitrary process which will perform the build step:

  • The builder, a path to an executable

  • A list of arguments

  • A map of environment variables.

• A two-component "system" type (e.g. x86_64-linux) where the executable is to run.

Referencing derivations

Derivations are always referred to by the store path of the store object they are encoded to. See the encoding section for more details on how this encoding works, and thus what exactly what store path we would end up with for a given derivation.

The store path of the store object which encodes a derivation is often called a derivation path for brevity.

Deriving path

Deriving paths are a way to refer to store objects that might not yet be realised. This is necessary because, in general and particularly for content-addressed derivations, the store path of an output is not known in advance. There are two forms:

• [constant]{#deriving-path-constant}: just a store path. It can be made valid by copying it into the store: from the evaluator, command line interface or another st ore.

• [output]{#deriving-path-output}: a pair of a store path to a derivation and an output name.

In pseudo code:

type OutputName = String

data DerivingPath
  = Constant { path : StorePath }
  | Output {
      drvPath : StorePath,
      output  : OutputName,
    }

Parts of a derivation

With both derivations introduced and deriving paths defined, it is now possible to define the parts of a derivation.

Inputs

The inputs are a set of deriving paths, refering to all store objects needed in order to perform this build step.

The information needed for the execve system call will presumably include many store paths:

• The path to the executable is almost surely starts with a store path
• The arguments and environment variables likely contain many other store paths.

But just as we stored the references contained in the file data separately for store objects, so we store the set of inputs separately from the builder, arguments, and environment variables.

Outputs

The outputs are the derivations are the store objects it is obligated to produce.

Outputs are assigned names, and also consistent of other information based on the type of derivation.

Output names can be any string which is also a valid store path name. The store path of the output store object (also called an [output path] for short), has a name based on the derivation name and the output name. Most outputs' store paths have name drvMame + '-' + outputName. However, an output named "out" has a store path with name drvName. This is to allow derivations with a single output to avoid a superfluous -<outputName> in their single output's name when no disambiguation is needed.

Example

A derivation is named hello, and has two outputs, out, and dev

• The derivation's path will be: /nix/store/<hash>-hello.drv.

• The store path of out will be: /nix/store/<hash>-hello.

• The store path of dev will be: /nix/store/<hash>-hello-dev.

Process creation fields

These are the three fields which describe out to spawn the process which (along with any of its own child processes) will perform the build. As state in the derivation introduction, this is everything needed for an execve system call.

Builder

This is the path to an executable that will perform the build and produce the outputs.

Arguments

Command-line arguments to be passed to the builder executable.

Note that these are the arguments after the first argument. The first argument, argv[0], is the "base name" of the builder, as per the usual convention on Unix. See Wikipedia for details.

Environment Variables

Environment variables which will be passed to the builder executable.

Placeholders

Placeholders are opaque values used within the process creation fields to [store objects] for which we don't yet know store paths. The are strings in the form /<hash> that are embedded anywhere within the strings of those fields.

Note

Output Deriving Path exist to solve the same problem as placeholders --- that is, referring to store objects for which we don't yet know a store path. They also have a string syntax, descibed in the encoding section. We could use that syntax instead of /<hash> for placeholders, but its human-legibility would cuse problems.

There are two types of placeholder, corresponding to the two cases where this problem arises:

• [Output placeholder]{#output-placeholder}:

  This is a placeholder for a derivation's own output.

• [Input placeholder]{#input-placeholder}:

  This is a placeholder to a derivation's non-constant input, i.e. an input that is an [output derived path].

Explanation

In general, we need to realise realise a store object in order to be sure to have a store object for it. But for these two cases this is either impossible or impractical:

• In the output case this is impossible:

  We cannot built the output until we have a correct derivation, and we cannot have a correct derivation (without using placeholders) until we have the output path.

• In the input case this is impractical:

  We an always built a dependency, and then refer to its output by store path, but by doing so we loose the ability for a derivation graph to describe an entire build plan consisting of multiple build steps.

Note

The current method of creating hashes which we substitute for string fields should be seen as an artifact of the current "ATerm" serialization format. In order to be more explicit, and avoid gotchas analogous to SQL injection, we ought to consider switching two a different format where we explicitly use a syntax for the concatenation of plain strings and [deriving paths] written more explicitly.

System

The system type on which the builder executable is meant to be run.

A necessary condition for Nix to schedule a given derivation on given Nix instance is for the "system" of that derivation to match that instance's system configuration option.

By putting the system in each derivation, Nix allows heterogenous build plans, where not all steps can be run on the same machine or same sort of machine. A Nix isntance scheduling builds can automatically build on other platforms by forwarding build requests to other Nix instances.

Encoding

Derivation

There are two formats, documented separately:

• The legacy "ATerm" format

• The experimental JSON format

Every derivation has a canonical choice of encoding used to serialize it to a store object. This ensures that there is a canonical store path used to refer to the derivation, as described in Referencing derivations.

Note

Currently, the canonical encoding for every derivation is the "ATerm" format, but this is subject to change for types derivations which are not yet stable.

Regardless of the format used, when serializing to store objects, content-addressing is always used.

In the common case the inputs to store objects are either:

• constant deriving paths for content-addressed source objects, which are "initial inputs" rather than the outputs of some other derivation

• the outputs of other derivations abiding by this same invariant.

This common case makes for the following useful property: when we serialize such a derivation graph to store objects, the resulting closures are entirely content-addressed.

Here is a sketch at the proof of this:

• The inputs which are constant deriving paths become references of the serialized derivations, but they are content-addressed per the above.

• For inputs which are output deriving paths, we cannot directly reference the input because in general it is not built yet. We instead "peel back" the output deriving path to take its underlying serialized derivation (the drvPath field), and reference that. Since it is a derivation, it must be content-addressed

• There are no other ways a store object would end up in an input closure. The references of a derivation in store object form always come from solely from the inputs of the derivation.

Deriving Path

• constant

  Constant deriving paths are encoded simply as the underlying store path is. Thus, we see that every encoded store path is also a valid encoded (constant) deriving path.

• output

  Output deriving paths are encoded by

  • encoding of a store path referring to a derivation

  • a ^ separator (or ! in some legacy contexts)

  • the name of an output of the previously referred derivation

  Example

  /nix/store/lxrn8v5aamkikg6agxwdqd1jz7746wz4-firefox-98.0.2.drv^out
  

  This parses like so:

  /nix/store/lxrn8v5aamkikg6agxwdqd1jz7746wz4-firefox-98.0.2.drv^out
  |------------------------------------------------------------| |-|
  store path (usual encoding)                                    output name
                                                            |--|
                                                            note the ".drv"
  

Extending the model to be higher-order

Experimental feature: dynamic-derivations

We can apply the same extension discussed for the abstract model to the concrete model. Again, only the data type for Deriving Paths needs to be modified. Derivations are the same except for using the new extended deriving path data type.

type OutputName = String

data DerivingPath
  = Constant { storeObj : StorePath }
  | Output {
      drv    : DerivingPath, -- Note: changed
      output : OutputName,
    }

Now, the drv field of Output is itself a DerivingPath instead of a StorePath.

Under this extended model, DerivingPaths are thus inductively built up from an Constant, contains in 0 or more outer Outputs.

Encoding

The encoding is adjusted in a very simplest way, merely displaying the same

Example

/nix/store/lxrn8v5aamkikg6agxwdqd1jz7746wz4-firefox-98.0.2.drv^foo.drv^bar.drv^out
|----------------------------------------------------------------------------| |-|
inner deriving path (usual encoding)                                           output name
|--------------------------------------------------------------------| |-----|
even more inner deriving path (usual encoding)                         output name
|------------------------------------------------------------| |-----|
innermost constant store path (usual encoding)                 output name

Extra extensions

__structuredAttrs

Historically speaking, most users of Nix made GNU Bash with a script the command run, regardless of what they were doing. Bash variable are automatically created from env vars, but bash also supports array and string-keyed map variables in addition to string variables. People also usually create derivations using language which also support these richer data types. It was thus desired a way to get this data from the language "planning" the derivation to language to bash, the language evaluated at "run time".

__structuredAttrs does this by smuggling inside the core derivation format a map of named richer data. At run time, this becomes two things:

1. A JSON file containing that map.
2. A bash script setting those variables.

The bash command can be passed a script which will "source" that Nix-created bash script, setting those variables with the richer data. The outer script can then do whatever it likes with those richer variables as input.