Code Loading

Julia has two mechanisms for loading code:

Code inclusion: e.g. include("source.jl"). Inclusion allows you to split a single program across multiple source files. The expression include("source.jl") causes the contents of the file source.jl to be evaluated in the global scope of the module where the include call occurs. If include("source.jl") is called multiple times, source.jl is evaluated multiple times. The included path, source.jl, is interpreted relative to the file where the include call occurs. This makes it simple to relocate a subtree of source files. In the REPL, included paths are interpreted relative to the current working directory, pwd().
Package loading: e.g. import X or using X. The import mechanism allows you to load a package—i.e. an independent, reusable collection of Julia code, wrapped in a module—and makes the resulting module available by the name X inside of the importing module. If the same X package is imported multiple times in the same Julia session, it is only loaded the first time—on subsequent imports, the importing module gets a reference to the same module. Note though, that import X can load different packages in different contexts: X can refer to one package named X in the main project but potentially different packages named X in each dependency. More on this below.

Code inclusion is quite straightforward: it simply parses and evaluates a source file in the context of the caller. Package loading is built on top of code inclusion and is a lot more complex. Therefore, the rest of this chapter focuses on the behavior and mechanics of package loading.

Note

You only need to read this chapter if you want to understand the technical details of package loading. If you just want to install and use packages, simply use Julia's built-in package manager to add packages to your environment and write import X or using X in your code to load packages that you've added.

A package is a source tree with a standard layout providing functionality that can be reused by other Julia projects. A package is loaded by import X or using X statements. These statements also make the module named X, which results from loading the package code, available within the module where the import statement occurs. The meaning of X in import X is context-dependent: which X package is loaded depends on what code the statement occurs in. The effect of import X depends on two questions:

What package is X in this context?
Where can that X package be found?

Understanding how Julia answers these questions is key to understanding package loading.

Federation of packages

Julia supports federated management of packages. This means that multiple independent parties can maintain both public and private packages and registries of them, and that projects can depend on a mix of public and private packages from different registries. Packages from various registries are installed and managed using a common set of tools and workflows. The Pkg package manager ships with Julia 0.7/1.0 and lets you install and manage your projects' dependencies. It does this by creating and manipulating project files that describe what your project depends on, and manifest files that snapshot exact versions of your project's complete dependency graph.

One consequence of federation is that there cannot be a central authority for package naming. Different entities may use the same name to refer to unrelated packages. This possibility is unavoidable since these entities do not coordinate and may not even know about each other. Because of the lack of a central naming authority, a single project can quite possibly end up depending on different packages that have the same name. Julia's package loading mechanism handles this by not requiring package names to be globally unique, even within the dependency graph of a single project. Instead, packages are identified by universally unique identifiers (UUIDs) which are assigned to them before they are registered. The question "what is X?" is answered by determining the UUID of X.

Since the decentralized naming problem is somewhat abstract, it may help to walk through a concrete scenario to understand the issue. Suppose you're developing an application called App, which uses two packages: Pub and Priv. Priv is a private package that you created, whereas Pub is a public package that you use but don't control. When you created Priv, there was no public package by that name. Subsequently, however, an unrelated package also named Priv has been published and become popular. In fact, the Pub package has started to use it. Therefore, when you next upgrade Pub to get the latest bug fixes and features, App will end up—through no action of yours other than upgrading—depending on two different packages named Priv. App has a direct dependency on your private Priv package, and an indirect dependency, through Pub, on the new public Priv package. Since these two Priv packages are different but both required for App to continue working correctly, the expression import Priv must refer to different Priv packages depending on whether it occurs in App's code or in Pub's code. Julia's package loading mechanism allows this by distinguishing the two Priv packages by context and UUID. How this distinction works is determined by environments, as explained in the following sections.

Environments

An environment determines what import X and using X mean in various code contexts and what files these statements cause to be loaded. Julia understands three kinds of environments:

A project environment is a directory with a project file and an optional manifest file. The project file determines what the names and identities of the direct dependencies of a project are. The manifest file, if present, gives a complete dependency graph, including all direct and indirect dependencies, exact versions of each dependency, and sufficient information to locate and load the correct version.
A package directory is a directory containing the source trees of a set of packages as subdirectories. This kind of environment was the only kind that existed in Julia 0.6 and earlier. If X is a subdirectory of a package directory and X/src/X.jl exists, then the package X is available in the package directory environment and X/src/X.jl is the source file by which it is loaded.
A stacked environment is an ordered set of project environments and package directories, overlaid to make a single composite environment in which all the packages available in its constituent environments are available. Julia's load path is a stacked environment, for example.

These three kinds of environment each serve a different purpose:

Project environments provide reproducibility. By checking a project environment into version control—e.g. a git repository—along with the rest of the project's source code, you can reproduce the exact state of the project and all of its dependencies since the manifest file captures the exact version of every dependency.
Package directories provide low-overhead convenience when a project environment isn't needed. Package directories are handy when you have a set of packages that you just want to put somewhere and use them as they are, without having to create and maintain a project environment for them.
Stacked environments allow for augmentation of the primary environment with additional tools. You can push an environment including development tools onto the stack and they will be available from the REPL and scripts but not from inside packages.

As an abstraction, an environment provides three maps: roots, graph and paths. When resolving the meaning of import X, roots and graph are used to determine the identity of X and answer the question "what is X?", while the paths map is used to locate the source code of X and answer the question "where is X?" The specific roles of the three maps are:

roots:name::Symbol ⟶ uuid::UUID
An environment's roots map assigns package names to UUIDs for all the top-level dependencies that the environment makes available to the main project (i.e. the ones that can be loaded in Main). When Julia encounters import X in the main project, it looks up the identity of X as roots[:X].
graph:context::UUID ⟶ name::Symbol ⟶ uuid::UUID
An environment's graph is a multilevel map which assigns, for each context UUID, a map from names to UUIDs, similar to the roots map but specific to that context. When Julia sees import X in the code of the package whose UUID is context, it looks up the identity of X as graph[context][:X]. In particular, this means that import X can refer to different packages depending on context.
paths:uuid::UUID × name::Symbol ⟶ path::String
The paths map assigns to each package UUID-name pair, the location of that package's entry-point source file. After the identity of X in import X has been resolved to a UUID via roots or graph (depending on whether it is loaded from the main project or a dependency), Julia determines what file to load to acquire X by looking up paths[uuid,:X] in the environment. Including this file should create a module named X. Once this package is loaded, i.e. after its first import, any subsequent import resolving to the same uuid will simply create a new binding to the original already-loaded package module.

Each kind of environment defines these three maps differently, as detailed in the following sections.

Note

For ease of understanding, the examples throughout this chapter show full data structures for roots, graph and paths. However, for efficiency, Julia's package loading code does not actually create them. Instead, it queries them through internal APIs and lazily computes only as much of each structure as it needs to load a given package.

Project environments

A project environment is determined by a directory containing a project file called Project.toml, and optionally a manifest file called Manifest.toml. These files may also be called JuliaProject.toml and JuliaManifest.toml, in which case Project.toml and Manifest.toml are ignored. (This allows for coexistence with other tools that might consider files called Project.toml and Manifest.toml significant.) For pure Julia projects, however, the names Project.toml and Manifest.toml are preferred. The roots, graph and paths maps of a project environment are defined as follows.

The roots map of the environment is determined by the contents of the project file, specifically, its top-level name and uuid entries and its [deps] section (all optional). Consider the following example project file for the hypothetical application, App, as described earlier:

name = "App"
uuid = "8f986787-14fe-4607-ba5d-fbff2944afa9"

[deps]
Priv = "ba13f791-ae1d-465a-978b-69c3ad90f72b"
Pub  = "c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1"

This project file implies the following roots map, if it was represented by a Julia dictionary:

roots = Dict(
    :App  => UUID("8f986787-14fe-4607-ba5d-fbff2944afa9"),
    :Priv => UUID("ba13f791-ae1d-465a-978b-69c3ad90f72b"),
    :Pub  => UUID("c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1"),
)

Given this roots map, in App's code the statement import Priv will cause Julia to look up roots[:Priv], which yields ba13f791-ae1d-465a-978b-69c3ad90f72b, the UUID of the Priv package that is to be loaded in that context. This UUID identifies which Priv package to load and use when the main application evaluates import Priv.

The dependency graph of a project environment is determined by the contents of the manifest file, if present. If there is no manifest file, graph is empty. A manifest file contains a stanza for each of a project's direct or indirect dependencies, including for each one, its UUID and a source tree hash or an explicit path to the source code. Consider the following example manifest file for App:

[[Priv]] # the private one
deps = ["Pub", "Zebra"]
uuid = "ba13f791-ae1d-465a-978b-69c3ad90f72b"
path = "deps/Priv"

[[Priv]] # the public one
uuid = "2d15fe94-a1f7-436c-a4d8-07a9a496e01c"
git-tree-sha1 = "1bf63d3be994fe83456a03b874b409cfd59a6373"
version = "0.1.5"

[[Pub]]
uuid = "c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1"
git-tree-sha1 = "9ebd50e2b0dd1e110e842df3b433cb5869b0dd38"
version = "2.1.4"

  [Pub.deps]
  Priv = "2d15fe94-a1f7-436c-a4d8-07a9a496e01c"
  Zebra = "f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62"

[[Zebra]]
uuid = "f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62"
git-tree-sha1 = "e808e36a5d7173974b90a15a353b564f3494092f"
version = "3.4.2"

This manifest file describes a possible complete dependency graph for the App project:

There are two different Priv packages that the application needs—a private one which is a direct dependency and a public one which is an indirect dependency through Pub:
- The private Priv depends on the Pub and Zebra packages.
- The public Priv has no dependencies.
The application also depends on the Pub package, which in turn depends on the public Priv and the same Zebra package which the private Priv package depends on.

This dependency graph represented as a dictionary, looks like this:

graph = Dict{UUID,Dict{Symbol,UUID}}(
    # Priv – the private one:
    UUID("ba13f791-ae1d-465a-978b-69c3ad90f72b") => Dict{Symbol,UUID}(
        :Pub   => UUID("c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1"),
        :Zebra => UUID("f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62"),
    ),
    # Priv – the public one:
    UUID("2d15fe94-a1f7-436c-a4d8-07a9a496e01c") => Dict{Symbol,UUID}(),
    # Pub:
    UUID("c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1") => Dict{Symbol,UUID}(
        :Priv  => UUID("2d15fe94-a1f7-436c-a4d8-07a9a496e01c"),
        :Zebra => UUID("f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62"),
    ),
    # Zebra:
    UUID("f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62") => Dict{Symbol,UUID}(),
)

Given this dependency graph, when Julia sees import Priv in the Pub package—which has UUID c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1—it looks up:

graph[UUID("c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1")][:Priv]

and gets 2d15fe94-a1f7-436c-a4d8-07a9a496e01c, which indicates that in the context of the Pub package, import Priv refers to the public Priv package, rather than the private one which the app depends on directly. This is how the name Priv can refer to different packages in the main project than it does in one of its package's dependencies, which allows for duplicate names in the package ecosystem.

What happens if import Zebra is evaluated in the main App code base? Since Zebra does not appear in the project file, the import will fail even though Zebradoes appear in the manifest file. Moreover, if import Zebra occurs in the public Priv package—the one with UUID 2d15fe94-a1f7-436c-a4d8-07a9a496e01c—then that would also fail since that Priv package has no declared dependencies in the manifest file and therefore cannot load any packages. The Zebra package can only be loaded by packages for which it appear as an explicit dependency in the manifest file: the Pub package and one of the Priv packages.

The paths map of a project environment is also determined by the manifest file if present and is empty if there is no manifest. The path of a package uuid named X is determined by these two rules:

If the manifest stanza matching uuid has a path entry, use that path relative to the manifest file.
Otherwise, if the manifest stanza matching uuid has a git-tree-sha1 entry, compute a deterministic hash function of uuid and git-tree-sha1—call it slug—and look for packages/X/$slug in each directory in the Julia DEPOT_PATH global array. Use the first such directory that exists.

If applying these rules doesn't find a loadable path, the package should be considered not installed and the system should raise an error or prompt the user to install the appropriate package version.

In the example manifest file above, to find the path of the first Priv package—the one with UUID ba13f791-ae1d-465a-978b-69c3ad90f72b—Julia looks for its stanza in the manifest file, sees that it has a path entry, looks at deps/Priv relative to the App project directory—let's suppose the App code lives in /home/me/projects/App—sees that /home/me/projects/App/deps/Priv exists and therefore loads Priv from there.

If, on the other hand, Julia was loading the otherPriv package—the one with UUID 2d15fe94-a1f7-436c-a4d8-07a9a496e01c—it finds its stanza in the manifest, see that it does not have a path entry, but that it does have a git-tree-sha1 entry. It then computes the slug for this UUID/SHA-1 pair, which is HDkr (the exact details of this computation aren't important, but it is consistent and deterministic). This means that the path to this Priv package will be packages/Priv/HDkr/src/Priv.jl in one of the package depots. Suppose the contents of DEPOT_PATH is ["/users/me/.julia", "/usr/local/julia"]; then Julia will look at the following paths to see if they exist:

/home/me/.julia/packages/Priv/HDkr/src/Priv.jl
/usr/local/julia/packages/Priv/HDkr/src/Priv.jl

Julia uses the first of these that exists to load the public Priv package.

Here is a representation of the paths map for the App project environment:

paths = Dict{Tuple{UUID,Symbol},String}(
    # Priv – the private one:
    (UUID("ba13f791-ae1d-465a-978b-69c3ad90f72b"), :Priv) =>
        # relative entry-point inside `App` repo:
        "/home/me/projects/App/deps/Priv/src/Priv.jl",
    # Priv – the public one:
    (UUID("2d15fe94-a1f7-436c-a4d8-07a9a496e01c"), :Priv) =>
        # package installed in the system depot:
        "/usr/local/julia/packages/Priv/HDkr/src/Priv.jl",
    # Pub:
    (UUID("c07ecb7d-0dc9-4db7-8803-fadaaeaf08e1"), :Pub) =>
        # package installed in the user depot:
        "/home/me/.julia/packages/Pub/oKpw/src/Pub.jl",
    # Zebra:
    (UUID("f7a24cb4-21fc-4002-ac70-f0e3a0dd3f62"), :Zebra) =>
        # package installed in the system depot:
        "/usr/local/julia/packages/Zebra/me9k/src/Zebra.jl",
)

This example map includes three different kinds of package locations:

The private Priv package is "vendored" inside the App repository.
The public Priv and Zebra packages are in the system depot, where packages installed and managed by the system administrator live. These are available to all users on the system.
The Pub package is in the user depot, where packages installed by the user live. These are only available to the user who installed them.

Package directories

Package directories provide a kind of environment that approximates package loading in Julia 0.6 and earlier, and which resembles package loading in many other dynamic languages. The set of packages available in a package directory corresponds to the set of subdirectories it contains that look like packages: if X/src/X.jl is a file in a package directory, then X is considered to be a package and X/src/X.jl is the file Julia loads to get X. Which packages can "see" each other as dependencies depends on whether they contain project files, and if they do, on what appears in those project files' [deps] sections.

The roots map is determined by the subdirectories X of a package directory for which X/src/X.jl exists and whether X/Project.toml exists and has a top-level uuid entry. Specifically :X => uuid goes in roots for each such X where uuid is defined as:

If X/Project.toml exists and has a uuid entry, then uuid is that value.
If X/Project.toml exists and but does not have a top-level UUID entry, uuid is a dummy UUID generated by hashing the canonical path of X/Project.toml.
If X/Project.toml does not exist, then uuid is the all-zero nil UUID.

The dependency graph of a project directory is determined by the presence and contents of project files in the subdirectory of each package. The rules are:

If a package subdirectory has no project file, then it is omitted from graph and import statements in its code are treated as top-level, the same as the main project and REPL.
If a package subdirectory has a project file, then the graph entry for its UUID is the [deps] map of the project file, which is considered to be empty if the section is absent.

As an example, suppose a package directory has the following structure and content:

Aardvark/
    src/Aardvark.jl:
        import Bobcat
        import Cobra

Bobcat/
    Project.toml:
        [deps]
        Cobra = "4725e24d-f727-424b-bca0-c4307a3456fa"
        Dingo = "7a7925be-828c-4418-bbeb-bac8dfc843bc"

    src/Bobcat.jl:
        import Cobra
        import Dingo

Cobra/
    Project.toml:
        uuid = "4725e24d-f727-424b-bca0-c4307a3456fa"
        [deps]
        Dingo = "7a7925be-828c-4418-bbeb-bac8dfc843bc"

    src/Cobra.jl:
        import Dingo

Dingo/
    Project.toml:
        uuid = "7a7925be-828c-4418-bbeb-bac8dfc843bc"

    src/Dingo.jl:
        # no imports

Here is a corresponding roots structure, represented as a dictionary:

roots = Dict{Symbol,UUID}(
    :Aardvark => UUID("00000000-0000-0000-0000-000000000000"), # no project file, nil UUID
    :Bobcat   => UUID("85ad11c7-31f6-5d08-84db-0a4914d4cadf"), # dummy UUID based on path
    :Cobra    => UUID("4725e24d-f727-424b-bca0-c4307a3456fa"), # UUID from project file
    :Dingo    => UUID("7a7925be-828c-4418-bbeb-bac8dfc843bc"), # UUID from project file
)

Here is the corresponding graph structure, represented as a dictionary:

graph = Dict{UUID,Dict{Symbol,UUID}}(
    # Bobcat:
    UUID("85ad11c7-31f6-5d08-84db-0a4914d4cadf") => Dict{Symbol,UUID}(
        :Cobra => UUID("4725e24d-f727-424b-bca0-c4307a3456fa"),
        :Dingo => UUID("7a7925be-828c-4418-bbeb-bac8dfc843bc"),
    ),
    # Cobra:
    UUID("4725e24d-f727-424b-bca0-c4307a3456fa") => Dict{Symbol,UUID}(
        :Dingo => UUID("7a7925be-828c-4418-bbeb-bac8dfc843bc"),
    ),
    # Dingo:
    UUID("7a7925be-828c-4418-bbeb-bac8dfc843bc") => Dict{Symbol,UUID}(),
)

A few general rules to note:

A package without a project file can depend on any top-level dependency, and since every package in a package directory is available at the top-level, it can import all packages in the environment.
A package with a project file cannot depend on one without a project file since packages with project files can only load packages in graph and packages without project files do not appear in graph.
A package with a project file but no explicit UUID can only be depended on by packages without project files since dummy UUIDs assigned to these packages are strictly internal.

Observe the following specific instances of these rules in our example:

Aardvark can import on any of Bobcat, Cobra or Dingo; it does import Bobcat and Cobra.
Bobcat can and does import both Cobra and Dingo, which both have project files with UUIDs and are declared as dependencies in Bobcat's [deps] section.
Bobcat cannot depend on Aardvark since Aardvark does not have a project file.
Cobra can and does import Dingo, which has a project file and UUID, and is declared as a dependency in Cobra's [deps] section.
Cobra cannot depend on Aardvark or Bobcat since neither have real UUIDs.
Dingo cannot import anything because it has a project file without a [deps] section.

The paths map in a package directory is simple: it maps subdirectory names to their corresponding entry-point paths. In other words, if the path to our example project directory is /home/me/animals then the paths map could be represented by this dictionary:

paths = Dict{Tuple{UUID,Symbol},String}(
    (UUID("00000000-0000-0000-0000-000000000000"), :Aardvark) =>
        "/home/me/AnimalPackages/Aardvark/src/Aardvark.jl",
    (UUID("85ad11c7-31f6-5d08-84db-0a4914d4cadf"), :Bobcat) =>
        "/home/me/AnimalPackages/Bobcat/src/Bobcat.jl",
    (UUID("4725e24d-f727-424b-bca0-c4307a3456fa"), :Cobra) =>
        "/home/me/AnimalPackages/Cobra/src/Cobra.jl",
    (UUID("7a7925be-828c-4418-bbeb-bac8dfc843bc"), :Dingo) =>
        "/home/me/AnimalPackages/Dingo/src/Dingo.jl",
)

Since all packages in a package directory environment are, by definition, subdirectories with the expected entry-point files, their paths map entries always have this form.

Environment stacks

The third and final kind of environment is one that combines other environments by overlaying several of them, making the packages in each available in a single composite environment. These composite environments are called environment stacks. The Julia LOAD_PATH global defines an environment stack—the environment in which the Julia process operates. If you want your Julia process to have access only to the packages in one project or package directory, make it the only entry in LOAD_PATH. It is often quite useful, however, to have access to some of your favorite tools—standard libraries, profilers, debuggers, personal utilities, etc.—even if they are not dependencies of the project you're working on. By pushing an environment containing these tools onto the load path, you immediately have access to them in top-level code without needing to add them to your project.

The mechanism for combining the roots, graph and paths data structures of the components of an environment stack is simple: they are simply merged as dictionaries, favoring earlier entries over later ones in the case of key collisions. In other words, if we have stack = [env₁, env₂, …] then we have:

roots = reduce(merge, reverse([roots₁, roots₂, …]))
graph = reduce(merge, reverse([graph₁, graph₂, …]))
paths = reduce(merge, reverse([paths₁, paths₂, …]))

The subscripted rootsᵢ, graphᵢ and pathsᵢ variables correspond to the subscripted environments, envᵢ, contained stack. The reverse is present because merge favors the last argument rather than first when there are collisions between keys in its argument dictionaries. That's all there is to stacked environments. There are a couple of noteworthy features of this design:

The primary environment—i.e. the first environment in a stack—is faithfully embedded in a stacked environment. The full dependency graph of the first environment in a stack is guaranteed to be included intact in the stacked environment including the same versions of all dependencies.
Packages in non-primary environments can end up using incompatible versions of their dependencies even if their own environments are entirely compatible. This can happen when one of their dependencies is shadowed by a version in an earlier environment in the stack.

Since the primary environment is typically the environment of a project you're working on, while environments later in the stack contain additional tools, this is the right tradeoff: it's better to break your dev tools but keep the project working. When such incompatibilities occur, you'll typically want to upgrade your dev tools to versions that are compatible with the main project.

Conclusion

Federated package management and precise software reproducibility are difficult but worthy goals in a package system. In combination, these goals lead to a more complex package loading mechanism than most dynamic languages have, but it also yields scalability and reproducibility that is more commonly associated with static languages. Fortunately, most Julia users can remain oblivious to the technical details of code loading and simply use the built-in package manager to add a package X to the appropriate project and manifest files and then write import X to load X without a further thought.