rfcs/text/3243-packages-as-optional-n...

19 KiB

Summary

Languages like C++ have open namespaces where anyone can write code in any namespace. In C++'s case, this includes the std namespace and is only limited by convention. In contrast, Rust has closed namespaces which can only include code from the original namespace definition (the crate).

This proposal extends Rust to have partially open namespaces by allowing crate owners to create crates like parent::foo that will be available as part of the crate parent's namespace. To protect the use of open namespaces, the owners of parent has exclusive access to publishing crates in that namespace.

Motivation

While Rust crates are practically unlimited in size, it is a common pattern for organizations to split their projects into many crates, especially if they expect users to only need a fraction of their crates or they have different backwards compatibility guarantees.

For example, unic, tokio, async-std, rusoto all do something like this, with lots of projectname-foo crates. At the moment, it is not necessarily true that a crate named projectname-foo is maintained by projectname, and in some cases that is even desired! E.g. serde has many third party "plugin" crates like serde-xml-rs. Similarly, async-tls is a general crate not specific to the async-std ecosystem.

Regardless, it is nice to have a way to signify "these are all crates belonging to a single project, and you may trust them the same" and discover these related crates. When starting up ICU4X, we came up against this problem: We wanted to be able to publish ICU4X as an extremely modular system of icu-foo or icu4x-foo crates, but it would be confusing to users if third-party crates could also exist there (or take names we wanted to use).

It's worth spending a bit of time talking about "projects" and "organizations", as nebulous as those terms are. This feature is primarily motivated by the needs of "projects". By this I mean a single Rust API developed as multiple crates, for example serde and serde::derive, or icu and icu::provider, or servo::script and servo::layout. One would expect "projects" like this to live under a single Git repository according to the norms of project organization; they are logically a single project and API even if they are multiple crates.

The feature suggested here is unlikely to be used by "organizations" as this would put independent concerns in the same Rust API. By "organizations", I mean a group of people who are coming together to build likely related crates, under the same "brand", likely developed in multiple repos under a GitHub organization.

The motivation here is distinct from the general problem of squatting -- with general squatting, someone else might come up with a cool crate name before you do. However, with projectname-foo crates, it's more of a case of third parties "muscling in" on a name you have already chosen and are using.

Guide-level explanation

The owners of the foo crate may provide other crates under the foo namespace, like foo::bar. For users, this makes its official status clearer and makes it easier to discover.

Users import these crates in Cargo.toml as normal:

[dependencies]
"foo" = "1.0.42"
"foo::bar" = "3.1"

They will then access this through a facade made of foo and all foo::* crates, for example:

let baz = foo::bar::Baz::new();
foo::render(baz);

Some reasons for foos owner to consider using namespaces:

  • Avoid name conflicts with third-party authors (since they are reserved)
  • Improve discoverability of official crates
  • As an alternative to feature flags for optional subsystems
  • When different parts of your API might have different compatibility guarantees

When considering this, keep in mind:

  • Does it makes sense for this new crate to be presented in the foo facade?
  • How likely is a crate to move into or out of the namespace?
    • Moving the crate in or out of a namespace is a breaking change though it can be worked around by having the old crate re-export the new crate but that does add extra friction to the process.
    • There is not currently a mechanism to raise awareness with users that a crate has migrated into or out of a namespace and you might end up leaving users behind.
  • If users import both foo and foo::bar but foo also has a bar item in its API that isn't just foo::bar re-exported, then rustc will error.

Only the owners of foo may create the foo::bar crate (and all owners of foo are implicitly owners of foo::bar). After the foo::bar crate is created, additional per-crate publishers may be added who will be able to publish subsequent versions as usual.

Reference-level explanation

This section will maintain a distinction between "package" (a crates.io package) and "crate" (the actual rust library). The rest of the RFC does not attempt to make this distinction

:: is now considered valid inside package names on Crates.io. For now, we will restrict package names to having a single :: in them, not at the beginning or end of the name, but this can be changed in the future.

When publishing a package foo::bar, if the package does not exist, the following must be true:

  • foo must exist
  • The user publishing the package must be an owner of foo

For the package foo::bar, all owners of foo are always considered owners of foo::bar, however additional owners may be added. People removed from ownership of foo will also lose access to foo::bar unless they were explicitly added as owners to foo::bar.

Crates.io displays foo::bar packages with the name foo::bar, though it may stylistically make the foo part link to the foo package.

The registry index trie may represent subpackages by placing foo::bar as just foo::bar.

rustc will need some changes. When --extern foo::bar=crate.rlib is passed in, rustc will include this crate during resolution as if it were a module bar living under crate foo. If crate foo is also in scope, this will not automatically trigger any errors unless foo::bar is referenced, foo has a module bar, and that module is not just a reexport of crate foo::bar.

The autogenerated lib.name key for such a crate will just be bar, the leaf crate name, and the expectation is that to use such crates one must use --extern foo::bar=bar.rlib syntax. There may be some better things possible here, perhaps foo_bar can be used here.

Drawbacks

Namespace root taken

Not all existing projects can transition to using namespaces here. For example, the unicode crate is reserved, so unicode-rs cannot use it as a namespace despite owning most of the unicode-foo crates. In other cases, the "namespace root" foo may be owned by a different set of people than the foo-bar crates, and folks may need to negotiate (async-std has this problem, it manages async-foo crates but the root async crate is taken by someone else). Nobody is forced to switch to using namespaces, of course, so the damage here is limited, but it would be nice for everyone to be able to transition.

Slow migration

Existing projects wishing to use this may need to manually migrate. For example, unic-langid may become unic::langid, with the unic project maintaining unic-langid as a reexport crate with the same version number. Getting people to migrate might be a bit of work, and furthermore maintaining a reexport crate during the (potentially long) transition period will also be some work. Of course, there is no obligation to maintain a transition crate, but users will stop getting updates if you don't.

A possible path forward is to enable people to register aliases, i.e. unic-langid is an alias for unic::langid.

Requires rustc changes

There are alternate solutions below that don't require the language getting more complex and can be done purely at the Cargo level. Unfortunately they have other drawbacks.

Rationale and alternatives

This change solves the ownership problem in a way that can be slowly transitioned to for most projects.

Slash as a separator

For discussions about separator choice, please discuss them in this issue to avoid overwhelming the main RFC thread.

A previous version of the RFC had / as a separator. It would translate it to foo_bar in source, and disambiguated in feature syntax with foo/bar/ vs foo/bar. It had the following drawbacks:

Slashes

So far slashes as a "separator" have not existed in Rust. There may be dissonance with having another non-identifier character allowed on crates.io but not in Rust code. Dashes are already confusing for new users. Some of this can be remediated with appropriate diagnostics on when / is encountered at the head of a path.

Furthermore, slashes are ambiguous in feature specifiers (though a solution has been proposed above for this):

[dependencies]
"foo" = "1"
"foo/std" = { version = "1", optional = true }

[features]
# Does this enable crate "foo/std", or feature "std" of crate "foo"?
default = ["foo/std"]

Dash typosquatting

This proposal does not prevent anyone from taking foo-bar after you publish foo/bar. Given that the Rust crate import syntax for foo/bar is foo_bar, same as foo-bar, it's totally possible for a user to accidentally type foo-bar in Cargo.toml instead of foo/bar, and pull in the wrong, squatted, crate.

We currently prevent foo-bar and foo_bar from existing at the same time. We could do this here as well, but it would only go in one direction: if foo/bar exists, neither foo-bar nor foo_bar will be allowed to be published. However, if foo-bar or foo_bar exist, we would choose to allow foo/bar to be published, because we don't want to limit the use of names within a crate namespace due to crates outside the namespace existing. This limits the "damage" to cases where someone pre-squats foo-bar before you publish foo/bar, and the damage can be mitigated by checking to see if such a clashing crate exists when publishing, if you actually care about this attack vector. There are some tradeoffs there that we would have to explore.

One thing that could mitigate foo/bar mapping to the potentially ambiguous foo_bar is using something like foo::crate::bar or ~foo::bar or foo::/bar in the import syntax.

Using identical syntax in Cargo.toml and Rust source

The / proposal does not require changes to Rust compiler to allow slash syntax (or whatever) to parse as a Rust path. Such changes could be made (though not with slash syntax due to parsing ambiguity, see below for more options); this RFC is attempting to be minimal in its effects on rustc.

However, the divergence between Cargo.toml and rustc syntax does indeed have a complexity cost, and may be confusing to some users. Furthermore, it increases the chances of Dash typosquatting being effective.

Some potential mappings for foo/bar could be:

  • foo::bar
  • foo::crate::bar
  • foo::/bar
  • ~foo::bar

and the like.

Whole crate name vs leaf crate name in Rust source

For discussions about separator choice, please discuss them in this issue to avoid overwhelming the main RFC thread.

It may be potentially better to use just the leaf crate name in Rust source. For example, when using crate foo/bar from Cargo.toml, the Rust code would simply use bar::. Cargo already supports renaming dependencies which can be used to deal with any potential ambiguities here. This also has the added benefit of not having to worry about the separator not parsing as valid Rust.

A major drawback to this approach is that while it addresses the "the namespace is an organization" use case quite well (e.g. unicode/segmentation vs unicode/line-break and rust-lang/libc vs rust-lang/lazy-static, etc), this is rather less amenable to the "the namespace is a project" case (e.g. serde vs serde/derive, icu/datetime vs icu/provider, etc), where the crates are related not just by provenance. In such cases, users may wish to rename the crates to avoid confusion in the code. This may be an acceptable cost.

Separator choice

For discussions about separator choice, please discuss them in this issue to avoid overwhelming the main RFC thread.

A different separator might make more sense. See the previous section for more on the original proposal of / as a separator.

We could continue to use / but also use @, i.e. have crates named @foo/bar. This is roughly what npm does and it seems to work. The @ would not show up in source code, but would adequately disambiguate crates and features in Cargo.toml and in URLs.

We could perhaps have foo-* get autoreserved if you publish foo, as outlined in https://internals.rust-lang.org/t/pre-rfc-hyper-minimalist-namespaces-on-crates-io/13041 . I find that this can lead to unfortunate situations where a namespace traditionally used by one project (e.g. async-*) is suddenly given over to a different project (the async crate). Furthermore, users cannot trust foo-bar to be owned by foo because the vast number of grandfathered crates we will have.

Triple colons could work. People might find it confusing, but foo:::bar evokes Rust paths without being ambiguous.

We could use ~ which enables Rust code to directly name namespaced packages (as ~ is no longer used in any valid Rust syntax). It looks extremely weird, however.

We could use dots (foo.bar). This does evoke some similarity with Rust syntax, however there are ambiguities: foo.bar in Rust code could either mean "the field bar of local/static foo" or it may mean "the crate foo.bar".

Note that unquoted dots have semantic meaning in TOML, and allowing for unquoted dots would freeze the list of dependency subfields allowed (to version, git, branch, features, etc).

We could reverse the order and use @, i.e. foo/bar becomes bar@foo. This might be a tad confusing, and it's unclear how best to surface this in the source.

User / org namespaces

Another way to handle namespacing is to rely on usernames and GitHub orgs as namespace roots. This ties crates.io strongly to Github -- currently while GitHub is the only login method, there is nothing preventing others from being added.

Furthermore, usernames are not immutable, and that can lead to a whole host of issues.

The primary goal of this RFC is for project ownership, not org ownership, so it doesn't map cleanly anyway.

Feature Flags

This proposal allows for optional subsystems. This can be created today with feature flags by adding a dependency as optional and re-exporting it.

Draw backs to feature flags

  • Solutions for documenting feature flags are limited
  • Feature flags can be cumbersome to work with for users
  • A semver breakage in the optional-subsystem crate is a semver breakage in the namespace crate
  • The optional-subsystem crate cannot depend on the namespace crate
  • There is limited tooling for crate authors to test feature combinations especially in workspaces with feature unification and its slow (re-running all tests even if they aren't relevant)

Prior art

This proposal is basically the same as https://internals.rust-lang.org/t/pre-rfc-packages-as-namespaces/8628 and https://internals.rust-lang.org/t/pre-rfc-idea-cratespaces-crates-as-namespace-take-2-or-3/11320 .

Namespacing has been discussed in https://internals.rust-lang.org/t/namespacing-on-crates-io/8571 , https://internals.rust-lang.org/t/pre-rfc-domains-as-namespaces/8688, https://internals.rust-lang.org/t/pre-rfc-user-namespaces-on-crates-io/12851 , https://internals.rust-lang.org/t/pre-rfc-hyper-minimalist-namespaces-on-crates-io/13041 , https://internals.rust-lang.org/t/blog-post-no-namespaces-in-rust-is-a-feature/13040/4 , https://internals.rust-lang.org/t/crates-io-package-policies/1041/37, https://internals.rust-lang.org/t/crates-io-squatting/8031, and many others.

Python has a similar coupling of top-level namespaces and modules with the filesystem. Users coming from other packaging systems, like Perl, wanted to be able to split up a package under a common namespace. A hook to support this was added in Python 2.3 (see PEP 402). In PEP 420 they formalized a convention for packages to opt-in to sharing a namespace. Differences:

  • Python does not have a coupling between package names and top-level namespaces so there is no need for extending the package name format or ability to extend their registry for permissions support.
  • In Python, nothing can be in the namespace package while this RFC allows the namespace package to also provide an API.

Unresolved questions

Deferred to tracking issue to be resolved pre-stabilization:

  • How exactly should the Cargo.toml lib.name key work in this world, and how does that integrate with --extern and -L and sysroots?
  • Should we allow renames like "foo::bar" = { package = "foo_bar", version = "1.0" } in Cargo.toml?
  • How precisely should this be represented in the index trie?
  • How we should name the .crate file / download URL

Third-parties, like Linux distributions, will need to decide how to encode cargo package names in their distribution package names according to their individual rules. Compared to existing ecosystems with namespaces that they package, the only new wrinkle is that there can be 0-1 namespace levels.

Future possibilities

We can allow multiple layers of nesting if people want it.

FAQ

What if I don't want to publish my crate under a namespace?

You don't have to, namespaces are completely optional when publishing.

Does this stop people from squatting on coolcratename?

No, this proposal does not intend to address the general problem of squatting (See crates.io's policy, a lot of this has been discussed many times before). Instead, it allows people who own an existing crate to publish sub-crates under the same namespace. In other words, if you own coolcratename, it stops people from squatting coolcratename::derive.