A way for users to bulk upgrade across incompatible versions

[epage]

Problem

A lot of incompatible upgrades have a small enough of breakages that a "upgrade all of my incompatible version requirements" would make it easier to go through this process.

This is currently exposed in the cargo upgrade command

Proposed Solution

Tasks

• update breaking
  • Add cargo update --breaking #13979
    • Test cases for cargo update --breaking only-compatible renamed non-semver-operator transitive-dep
  • --precise <breaking>
  • Call for testing
  • Stabilization
• fix(update): Clarify meaning of --aggressive as --recursive #12544
• feat(pkgid): Allow incomplete versions when unambigious #12614
• feat(update): Tell users when they are still behind #13372
• fix(update): Make -p more convenient by being positional #12545
• bugs and identified tasks

Deferred

• Modifying version requirements for compatible upgrades

See #12425 (comment)

Previously, https://internals.rust-lang.org/t/feedback-on-cargo-upgrade-to-prepare-it-for-merging/17101/141

Unresolved questions

• Flag name
  • --breaking to focus on semver (as this is limited to ^)
  • --incompatible applies to all version req operators and doesn't have a good short flag (-i is generally assumed to be "interactive")
  • Dart calls it --major-versions though we have the issue of "absolute major" (x.0.0) and "relative major (also 0.x.0, 0.0.x).
    • budler calls it --major
  • pnpm calls it --latest because of a "latest" tag

Notes

Related

• cargo update should not highlight non-semver-compatible "(latest: v0.23.5)" for indirect dependencies #13908
• A method to update manifest version requirements (like cargo-edit's cargo-upgrade) #10498

[epage]

Below is some background on what we are using to help come up with a design

Care Abouts

Priorities

• Don't break behavior on cargo update
• Don't write out incompatible Cargo.lock and Cargo.toml
• Focus is on end-users solving common problems and not on being a general programmatic CLI that is meant to cover every case
• Be predictable and understandable
  • Can someone unfamiliar with Rust, reading a blog post, predict what different command invocations will do?
  • Preference for not having too similarly named commands
• When higher priorities allow, avoid errors that make users go "if you know exactly what I was asking for then why didn't you just do it?"; those are a sign of issues with the UX.
• Don't be hassle when dealing with intentionally held back dependencies

Primary use cases:

• Want to have simple workflow for "upgrade incompatible dependencies only".
• Want to have simple workflow for bulk lock to latest compatible (already exists as cargo update)
• (medium priority) Selective modify one dependency's version requirement to latest compatible, latest incompatible
• (lower priority) Want to have simple-ish workflow for bulk upgrade to latest compatible
• (lower priority) bulk upgrade all dependencies (could just be two command invocations)

Secondary use cases are:

• Selective modify version requirement to user-provided value
• Upgrade explicitly pinned version requirements

Some open questions we had

• How do we tell when a renamed dependency like tokio_03 is pinned or not?
  • We could just assume all renamed are pinned
  • We could add a dependency field but I'm a bit leery of adding that kind of bookkeeping to the manifest
  • We could force users to --exclude these dependencies but that might be a bit of a pain to always remember to do
  • We could only skip renamed if multiple dependencies exist that point to the same package

Context

Currently, cargo update is focused solely on Cargo.lock editing

• Spans entire dependency tree
• Multiple versions of a package may exist, referenced by name@version
• Deals with exact versions and not version ranges
• Only affects you and not your dependents

Version requirement editing is different in that

• Workspace members only
• May want differences between members
• Supports alternative names for packages
• Affects dependents

And as a reminder of the CLI:

cargo update -p foo  # select a non-ambiguous `foo` and update it
cargo update -p foo@ver  # selects a specific `foo` to update
cargo update -p foo --aggressive  # also update dependencies
cargo update -p foo --precise <ver>  # select a specific version
cargo update --locked  # fail if the lockfile will change

Note: cargo add --locked will also fail if the manifest will change

Some design tools we can consider include

• Renaming a command, making the old name an alias
  • Even if there isn't a culture shift to use the new name, cargo <cmd> --help and cargo --list will point people to the new name
• Versions without the build metadata field is a subset of version requirement syntax, we may be able to do some mixing of them
  • Precedence: using the same foo@ver syntax for versions and version requirements
• Minimal-version resolution being the default mode would make Cargo.lock mostly align with Cargo.toml, making it easier to conflate the two commands (whether merging them or keeping separate but de-emphasizing update)

Interjection

Through this, I realized that the core of my concern with our previous attempts at a single command is that it feels like we are shoehorning new behaviors into cargo update rather than making the behavior cohesive.

• If I see a cargo update --incompatible on a blog post, can I predict what will happen if you do cargo update? No, because --save is needed to match behavior
• We were trying to make --package be both for package IDs and dependency names to make some of the cargo upgrade workflows work
• We were trying to overload --precise to allow control over version requirements

I also realized that my Windows Updates vs Windows Upgrades analogy for cargo update and cargo upgrade breaks down a little because cargo upgrade can do "upgrades" that are on the level of cargo update (say we call it cargo upgrade --compatible). The difference is in the target audience (yourself vs your dependents)

[epage]

Proposal: cargo update only changes version requirements as a side effect

The primary role of cargo update has been to update your active dependencies (ie Cargo.lock). We do not plan to change that role but give the user control to force it to update in situations that were previously unsupported, particularly updating the Cargo.toml if need be.

Behavior:

• By default, only "safe" updates are allowed (today's behavior)
• cargo update --incompatible / cargo update -i will force (ie update version reqs) update unpinned, incompatible versions (no other dependencies)
  • Yes, this could potentially be called --breaking or something else. The name depends on what expresses the concept clearly especially in light of any pinning behavior we have
  • Keeping -i unused would keep the door for a --interactive to be added
• cargo update -p foo --precise ver will force the update to happen, only erroring if we can't (don't own relevant version reqs, is pinned), even if its incompatible but unpinned. Version requirement is only changed on incompatible.
  • Maybe the error on pinned could be relaxed

Somewhere between deferred and rejected (speaking for myself): Support in cargo update for writing to the manifest for non-breaking changes, like bulk compatible upgrades of version requirements (ie a -save flag) which was one of our lower priority workflows. A --save flag is more about updating versions for your dependents, which while important for having valid lower-bounds on version requirements, doesn't fit with the existing model of cargo update. Maybe in the future we can find a way to express this in cargo update that fits with how it works or maybe another command can take on this role. We just aren't wanting to distract our efforts for handling most of the use cases to handle this one. See #10498

While this tells a cohesive story, a part of me is somewhat concerned that this goes beyond the name update.

Potential related cargo update improvements

• "Rename" --aggressive to --recursive (make --aggressive an alias for the new name) (fix(update): Clarify meaning of --aggressive as --recursive #12544)
• Specialize the parsing of foo@x.y.z so that foo@x will work so long as its unambiguous (which it should be), much like cargo install foo@x.y.z supports auto-selecting the y and z fields (picks latest) (feat(pkgid): Allow incomplete versions when unambigious #12614)
• Consider applying some of the output formatting from cargo upgrade into cargo update (feat(update): Tell users when they are still behind #13372)
  • Provides some of the cargo outdated experience (Make cargo outdated part of cargo itself #4309)
  • Is an alternative to warnings (cargo update should print warnings for incompatible candidates #7167)
• Add a positional package argument to cargo update, removing the need for --package (fix(update): Make -p more convenient by being positional #12545)
  • Should be safe because there are no trailing var args for a wrapped child process and no (and little chance of) flags that accept 0..=1 values

Alternatives

These are alternatives I had considered that help give an idea of what I mean by fitting into cargo.

cargo update always modifies Cargo.toml

• This would be a breaking change
• This would get in the way of people intentionally keeping separate versions from version requirements

We deprecate cargo update and a cargo upgrade always updates both files

• This would get in the way of people intentionally keeping separate versions from version requirements

We migrate to minimal-version resolution by default

• cargo update becomes less useful and we move it out of the spot light.
• A cargo upgrade is added that is focused on editing version requirements

Separate commands for Cargo.lock (update) and Cargo.toml (upgrade)

• Names don't clarify the role each fills
• Much like Debian has apt dist-upgrade, maybe it could be cargo req-update?

Misc Notes

• I don't see us making a distinction between default operator and ^ as we document them as being the same thing and sometimes people use ^ just because
• We are erring on the side of needing cargo update && cargo update --incompatible vs cargo update --incompatible doing both as there are people who do want them separated and running two commands, while annoying, allows us to cover more use cases.

[epage]

This proposal has been up here and on internals for a bit now without any major concerns raised.

@rfcbot fcp merge

[rfcbot]

Team member @epage has proposed to merge this. The next step is review by the rest of the tagged team members:

• @Eh2406
• @Muscraft
• @arlosi
• @ehuss
• @epage
• @joshtriplett
• @weihanglo

No concerns currently listed.

Once a majority of reviewers approve (and at most 2 approvals are outstanding), this will enter its final comment period. If you spot a major issue that hasn't been raised at any point in this process, please speak up!

See this document for info about what commands tagged team members can give me.

[djc]

Having been fairly involved in this discussion via the internals thread, I'm happy to see this move forward in a direction that I wholeheartedly support. @epage thanks for pushing this forward!