Not related to async in Rust, but a 26 sec clean build time for a project with 69 dependencies and no code of its own just to get to the point where we can "do something useful" feels a little wasteful.

I'd say it's pretty unfair to take clean build time to consider doing something useful. 98% of my rust builds are incremental builds which take 10-15 seconds for a project I have been working on for 9 months full time and I think my dependency tree has easily 1000 dependencies.

True, it's not a real problem in most scenarios if incremental compile times are good. I still feel uneasy depending on so many other crates, but this seems to be a level of paranoia that others in the community don't share.

Having 1000 dependencies sounds crazy to me! If there's a bug in even one of them that affects you then there's gonna be a lot of digging to figure out the cause, and possibly multiple pull requests to get it fixed. I think my CPU would spend a bit longer than 10-15 secs on the linker step, too.

Compile/link times are definitely a discussion topic. Here's some quick tips:

  * Incremental compilation helps locally, not so much in CI, unless you save/restore the whole cache which gets large quickly and evicting the out-of-date objects is non-trivial.
  * In CI, sccache helps, but not as much as I'd like: there's a bunch of things that are non-cacheable, notably crates that pull in & compile C/C++ code (I'd trade 2 c-bindings crates against 12 Rust-only crates any day for that reason alone)
  * Splitting stuff across different crates helps parallelizing the build (and caching it better), which may be one of the reasons some projects end up having "1000 dependencies" (although I've rarely seen upwards of 600)
  * For incremental builds, linking does become the bottleneck. Full LTO is especially slow, Thin LTO is better. Switching to LLD improves thing. I'm hopeful that mold will improve things some more.
  * Re multiple PRs: thankfully crate families tend to live in the same repository, so fixing something across both tracing / tracing-subscriber could be a single PR, for example.