New 'cargotest' builders for CI will keep Rust working better


We love tests. Oh my, do we love tests. I love tests. We’ve got more tests.

I’ve added two more auto builders to our CI. The ‘auto’ builders are those that bors runs before merging and prevent Rust from breaking, and each represents a distinct configuration for Rust. Today we test Rust in 36 configurations every pull request, 31 of which are required to pass before the PR can land.

If you haven’t seen the wall of builders it is staggering.

These two new builders are called auto-linux-64-cargotest and auto-win-msvc-64-cargotest and they do something new.

They run a tool in the new build system called ‘cargotest’, which does one simple thing: downloads revisions of a few key crates, builds them and tests them. The crates it builds are chosen because they are prominent projects with broad dependencies. Today they are just two: Cargo and Iron. Between them are 76 transitive dependencies, the most battle-tested crates in the ecosystem, one of which is winapi.

bors gates on the cargotest builders, so it will be very hard to break any of these crates without some deep soul searching.



Just an idea, could bors note compile time and compiler output on those crates and give a nod in the github comments if

  • compile times changed more than a given threshold (say 10%)
  • output changed (preferably with a diff)

That way we’ll know of both compile time regressions and changes in the warnings before merging.


Automated regression detection is really, really hard. 10% is a decent threshold, but I wouldn’t be surprised if noise can nevertheless cause many false positives and temporarily hide true regressions. So it should probably be a “heads up” notification, not a gating criterion, but we already have in that direction, little point in duplicating efforts.


That’s what I meant with ‘nod’. The benefit would be faster turnaround time. If we already have the builders, it would be nice to get the most benefit out of them.


Great! Out of curiosity, I’d love to know which builders in particular (if any) tend to fail the most often, i.e. which platforms present the largest support burden.


I’d like to do a performance bot but to run it in a simulator for consistency.

We had a performance bot long ago that I think even gated on regressions, but it was useless because of the variability.

1 Like

Would love to see Diesel added to this eventually. We tend to stress the edges of the system with our trait usage, and have hit regressions w/ Trait visibility or something similar almost every month. (Worth noting that the core Diesel crate doesn’t rely on syntax extensions or anything unstable)


Adding a new repo should in theory be pretty easy actually, if you want to just add an entry here that should in theory be all that’s necessary!

closed #9

This topic was automatically closed 90 days after the last reply. New replies are no longer allowed.