Comment by saghm
15 hours ago
> At least they're pinned though.
Frustratingly, they're not by default though; you need to explicitly use `--locked` (or `--frozen`, which is an alias for `--locked --offline`) to avoid implicit updates. I've seen multiple teams not realize this and get confused about CI failures from it.
The implicit update surface is somewhat limited by the fact that versions in Cargo.toml implicitly assume the `^` operator on versions that don't specify a different operator, so "1.2.3" means "1.2.x, where x >= 3". For reasons that have never been clear to me, people also seem to really like not putting the patch version in though and just putting stuff like "1.2", meaning that anything other than a major version bump will get pulled in.
> The implicit update surface is somewhat limited by the fact that versions in Cargo.toml implicitly assume the `^` operator on versions that don't specify a different operator, so "1.2.3" means "1.2.x, where x >= 3". For reasons that have never been clear to me, people also seem to really like not putting the patch version in though and just putting stuff like "1.2", meaning that anything other than a major version bump will get pulled in.
Not quite: "1.2.3" = "^1.2.3" = ">=1.2.3, <2.0.0" in Cargo [0], and "1.2" = "^1.2.0" = ">=1.2.0, <2.0.0", so you get the "1.x.x" behavior either way. If you actually want the "1.2.x" behavior (e.g., I've sometimes used that behavior for gmp-mpfr-sys), you should write "~1.2.3" = ">=1.2.3, <1.3.0".
[0] https://doc.rust-lang.org/cargo/reference/specifying-depende...
I don't know how I got this wrong because I literally went and looked at that page to try to remind myself, but I somehow misread it, because you're definitely right. This probably isn't the first time I've gotten this wrong either.
From thinking it through more closely, it does actually seem like it might be a little safer to avoid specifying the patch version; it seems like putting 1.2.3 would fail to resolve any valid version in the case that 1.2.2 is the last non-yanked version and 1.2.3 is yanked. I feel like "1.2.3" meaning "~1.2.3" would have been a better default, since it at least provides some useful tradeoff compared to "1.2", but with the way it actually works, it seems like putting a full version with no operator is basically worse than either of the other options, which is disappointing.
Are we talking about `cargo build` here? Because my understanding is that if a lockfile is present and `Cargo.toml` hasn't changed since the lockfile was created then the build is guaranteed to use the versions in the lockfile.
If however `Cargo.toml` has changed then `cargo build` will have to recalculate the lockfile. Hence why it can be useful to be explicit about `cargo build --locked`.
Is there a plan to change this? I don't see why --locked shouldn't be the default
I haven't heard anything about this, but I really wish it was there by default. I don't think the way it works right now fits anyone's expectations of what the lockfile is supposed to do; the whole point of storing the resolved versions in a file is to, well, lock them, and implicitly updating them every time you build doesn't do that.
As one of the original authors of Cargo, I agree. lockfiles are for apps and CLIs are apps. QED.
Since you're here, and you happened to indirectly allude to something that seems to have become increasingly common in the Rust world nowadays, I can't help but be curious about your thoughts on libraries checking their lockfiles into version control. It's not totally clear to me exactly when or why it became widespread, but it used to be relatively rare for me to see in open source libraries in the first few post-1.0 years of Rust, whereas at this point I think it's more common for me to see than not.
Do you think it's an actively bad practice, completely benign, or something in between where it makes sense in some cases but probably should still be avoided in others? Offhand, the only variable I can think of that might influence a different choice is that maybe closed-source packages been reused within a company (especially if trying to interface with other package management systems, which I saw firsthand when working at AWS but I'm guessing is something other large companies would also run into), but I'm curious if there are other names nuances I haven't thought of
1 reply →
It should be fine to do this according to semver as long as the major version is above zero.
Sure, but according to semver it's also totally fine to change a function that returns a Result to start returning Err in cases that used to be Ok. Semver might be ae to project from your Rust code not compiling after you update, but it doesn't guarantee it will do the same thing the next time you run it. While changes like that could still happen in a patch release, I'd argue that you're losing nothing by forgoing new API features if all you're doing is recompiling the existing code you have without making any changes, so only getting patches and manually updating for anything else is a better default. (That said, one of the sibling comments pointed out I was actually wrong about the implicit behavior of Cargo dependencies, so what I recommended doesn't protect from anything, but not for the reasons it sounds like you were thinking).
Some people might argue that changing a function to return an error where it didn't previously would be a breaking change; I'd argue that those people are wrong about what semver means. From what I can tell, people having their own mental model of semver that conflicts with the actual specification is pretty common. Most of the time when I've had coworkers claim that semver says something that actively conflicts with what it says, after I point out the part of the spec that says something else, they end up still advocating for what they originally had said. This is fine, because there's nothing inherently wrong with a version schema other than semver, but I try to push back when the term itself gets used incorrectly because it makes discussions much more difficult than they need to be.