Re-export `rand-core` members in `rand`?

[dhardy]

Since rand-core is intended to be a separate crate, but with rand depending on it, should some parts of be re-exported in rand?

I argue that the following should be re-exported in rand so that "normal users" do not have to depend on rand-core:

• Rng, CryptoRng, SeedableRng, SeedFromRng traits
• Error struct and ErrorKind enum

The rand_core::impls module is aimed at implementations of Rng so does not need to be re-exported (in my opinion).

~There is one other item: the mock module. Probably it is useful to re-export this from rand.~

[dhardy]

It is also worth mentioning the story for crates implementing Rng / CryptoRng: the intention is that such crates can depend on rand-core but don't need to use rand.

[pitdicker]

Not re-exporting rand_core::impls would almost force crates implementing Rng to depend on rand-core instead of rand, guiding them to the right dependency. 👍

Could you explain why the mock module should be in rand_core?

[dhardy]

I put mock in there to allow tests on the Rng trait. It doesn't really need to there of course.

[dhardy]

Ok, I moved mock from rand_core to rand, so that's not a problem any more. I think the rest is resolved?

[pitdicker]

The re-exports seem fine to me.

I still like to look over rand_core::impls. I now have a branch with 45% wins for fill_bytes with Isaac, and suspect a much smaller win for something like Xorshift should be possible to. And it needs testing. But that is not relevant for this issue...

[burdges]

I know this messes up the crate separation, but did you consider

pub trait Rng : Sample { ... }
pub trait Sample { ... }
impl<R: Rng> Sample for R { ... }

Would that eliminate needing use rand::Sample? I know that a minor thing, and probably not even a good thing for reading/auditing, but it's vaguely relevant to backwards compatibility. It might also create a bunch of double references in trait objects, which maybe slows down &mut R : Rng slightly.

[dhardy]

I don't believe it would eliminate the need to use rand::Sample, no.

Good point about this being a breaking change; I don't see an obvious way around it though (also the crate separation is quite important IMO).

[burdges]

Right, if that existed it'd require an extra pub like pub trait Rng : pub Sample or something.

[dhardy]

Like class extension in classic OO languages? Yes, but I don't think anything like that exists in Rust.

[nixpulvis]

I think it would be a really good exercise to write down the complete desired interfaces for both rand and rand_core in once place. Might make discussions like this a bit easier (at least for me).

Edit: Public interface. I don't care about internal details for the purposes of this comment.

[dhardy]

Check out the documentation: https://dhardy.github.io/rand/

[dhardy]

Important to this topic are: which other parts of rand should be moved to other crates?

• Rng and seeding traits, and error handling code, are being moved to rand-core

• 58 suggests moving all PRNG implementations to external crates

• We could move external RNGs (OsRng and JitterRng) to another crate; the rationale is similar to rand-core: cryptographic code may but able to make use of these without needing other rand code; this implies they could also be moved to rand-core

There are more subdivisions possible, although these appear less useful:

• We could move distribution code and the new Sample trait to another crate, though this seems less useful since rand is heavily dependent on this code
• The rand::sequences module is more independent (see https://github.com/rust-lang-nursery/rand/issues/194)

Finally, rand::random is the most derived feature: it depends on Sample and distribution code for conversion, and on thread_rng which uses StdRng and ReseedingRng with entropy from the external RNGs — about the only things it doesn't depend on are the sequences module, weak_rng, and the Xorshift and ChaCha generators.

Conclusion

• [x] Discuss moving PRNGs: #58
• [ ] Discuss moving external RNGs: TODO
• [ ] Discuss moving the sequences module: TODO, https://github.com/rust-lang-nursery/rand/issues/194 is related
• [x] The rest of rand is too interdependent, thus should stay in a single crate

Re-exports

rand is dependent on rand-core, the external RNGs and some PRNGs. To save users depending on multiple crates, we should import anything from these crates which (1) is likely to be useful (includes most stuff, except perhaps rand-core::impls) and (2) is unlikely to be replaced (this may rule out PRNGs as noted in #58).

[dhardy]

I think this topic is either decided or will be reviewed later, so closing the issue.