An analysis of semver compliance in Rust finds accidental semver violations are common even in the most popular crates. Human error is not the cause, and better tooling is the way forward.
"Just don't write bugs" ( or "Just don't write semver violations" in this case) is now rightfully recognized as a joke proposition by many (although derivative ability/experience arguments are sometimes still used, UNIRONICALLY). But it's the "better education" (or its sister magic pill "better docs") that still has many believers. So it is still valuable to explicitly make the argument for reliable automated tooling as the only real logical solution. But I digress.
if our Example enum above was #[doc(hidden)], adding a new variant would not have violated semver.
Violations in #[doc(hidden)] items should definitely trigger errors by default IMHO. To give what was a kludge in the first place more powers is not something I would call wise. Not to mention, module source code is just one click away from html docs, and it's also one click/key-combination away from a crate user's editor/IDE with LSP (rust-analyzer).
So
how #[doc(hidden)] items sometimes have semver obligations after all,
I would argue it's always the case, unless the user of the tool explicitly decides it's not.
What do you think #[doc(hidden)] is for, other than declaring something "private" that the language unfortunately doesn't let you declare as truly private right now?
I've mostly seen it used as a way to expose tools to macro APIs. For example, these internal parts of the quote! macro, or these internal parts of the vec! macro. Changing these things shouldn't be considered a semver violation, because they're not really part of the API, even though the quote! macro can't enforce it.
The only other cases I can think of where I've seen #[doc(hidden)] used are even bigger kludges, and the hidden items definitely aren't part of the platonic API, like pre-#[non_exhaustive] crates that wanted to reserve the right to add new variants to their enums.
It would be nice if we had something like pub(macro), which would make something public but only inside macros from the same crate. So they are usable in macros but not part of the public API.
Probably a spicy take, but I think any API being used by a macro should be made public. A macro shouldn't be the only way to do something; it should just be a way to remove the boilerplate required to do it.
I'm arguing (humbly of course) that intended vs. unintended use of what is at the end of the day a part of the public interface shouldn't be taken into consideration by default. Otherwise, other cases can be argued as non-breaking too!
Foo was never meant to be sent to other threads, So, losing Send is not a semver- breaking change!
Exhaustive enum Bar is only meant to be matched exhaustively internally. We say so in the docs. So adding a variant to it is not a semver-breaking change!
And giving more powers to a (kludge) doc attribute just doesn't seem in my eyes to be a generally wise move.
A: cargo-semver is still complaining about this item which I already have cfg-ed under an exp_api crate feature (which I don't want to rename). CI is failing.
B: PRO-TIP: Just slap a #[doc(hidden)] on it and CI will pass!
A: What if I still want to see the docs?
B: We are pushing for --document-hidden-items to stabilize soon. So you can just simply use that!
semver is a nice way to communicate your intent to other humans, but it's never been a great way of programmatically communicating those changes. If you want it to communicate something meaningful programmatically, you'd have to have the version generated automatically, and not rely on a human to do so.
Semver was literally made to communicate programmatically. But people keep using it for "human" communication instead, like the whole "1.0 means stable" thing.
It was made to do so, but failed from the start exactly because humans got involved. semver's ideals can only happen when tooling generates the version number, not humans.
Is that even possible though? Sometimes you need a human to understand if something is a breaking change.
Imagine an API like fn third_planet_from_sun() -> String, and an update is made where the output changes the value to be lowercase instead of capitalized. That should normally be considered a breaking change.
However, imagine fn current_version() -> String. That is by its definition meant to change outputs between versions, so it isn't a breaking change since that's part of its human, documentation based API contract.
Also, what if somw function which returns a String changes, but only one code path that is very hard to hit changes the output? How would a machine find that?
I guess the first example with Earth / version could use some attribute macro so devs can say the output is expected to change across versions, but then there is no way for a program to know what is a breaking change vs expected vs a bug.
To do it 100% probably isn't possible, something something halting problem. However, you'd catch a lot of basic mistakes with proper typing. In your example, the first function should be typed like this: fn third_planet_from_sun() -> Planet, where Planet is an enum. De/serializing it still has the same problem of interpreting an arbitrary string, but at least for deserializing it, you can be loose in what you allow and just lowercase it before matching it to the enum.
There are also plenty of purposeful semver violations. For example, serde makes basically no attempt to follow semver, and any pleas to do otherwise fall on deaf ears.
Not quite. Suppose instead of a single version of serde there's now 46 versions like in https://crates.io/crates/parquet to be able to use instances derived in some other crate X you have to use the same version of serde. Now, how should a crate decide which versions of serde to support?
All 46 and all optional? Supporting that would be painful. Just the last one? crates.io is a cemetery full of dead crates, with this support strategy any handful of crates picked at random are not going to be serde-compatible with each other.
A better solution would be a better support for compile time reflection so serde doesn't have to exist in its current state, but that's got delayed (by big macro conspiracy :)
What's more, there was a recent discussion about why the derive feature is recommended in serde, and one of the points brought up was that the versions for both crates basically have to be equal. I couldn't help but wonder, would this be a problem if the releases actually followed semver? Theoretically, it shouldn't matter what versions you use, as long as they're above a certain minor version and the major versions match. But since everything is a patch, we have to pin the two crates together somehow.