Regression of applier build times in Rust 1.57

[nightkr]

Current and expected behavior

1.57 behaviour

1. Check out https://gitlab.com/teozkr/repros/-/tree/kube/rs1.57-compile-regression.
2. rustup update 1.57.0
3. cargo +1.57.0 build
4. Note that the full build gets stuck for several minutes on rs157-compile-regression(bin)
5. touch src/main.rs
6. cargo +1.57.0 build
7. Note that the incremental build takes ~40s (on my machine, scale to yours)

1.56 behaviour

1. rustup update 1.56.1
2. cargo +1.56.1 build
3. Note that the full build completes in ~30s (on my machine, scale to yours)
4. touch src/main.rs
5. cargo +1.56.1 build
6. Note that the incremental build takes ~3s (on my machine, scale to yours)

Possible solution

More of a workaround, but boxing queue:

diff --git a/src/main.rs b/src/main.rs
index a306bba..6e08d03 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -25,7 +25,8 @@ async fn main() {
                 stream::empty::<Result<watcher::Event<ConfigMap>, watcher::Error>>(),
             )),
             (),
-        ),
+        )
+        .boxed(),
     )
     .boxed()
     .for_each(|_| async {})

brings the incremental rebuild time down to ~4s.

Additional context

No response

Environment

Arch Linux
Rustc 1.57 (bad), 1.56.1 (good)
K8s cluster is irrelevant since this is a build issue

Configuration and features

kube = { version = "0.64.0", features = ["runtime"] }
k8s-openapi = { version = "0.13.1", features = ["v1_22"] }

Affected crates

kube-runtime

Would you like to work on fixing this bug?

yes

[nightkr]

Curiously, this only seems to happen for the pattern applier(..., trigger_self(try_flatten_applied(reflector(...)))). Removing any of those steps, or erasing to a trait object (with StreamExt::boxed) seems to step around the issue.

[clux]

Good find. Confirming it happens on mine as well.

$ touch src/main.rs 
$ time cargo +1.57.0 build
   Compiling rs157-compile-regression v0.1.0 (/home/clux/repos/repros/repros)
    Finished dev [unoptimized + debuginfo] target(s) in 1m 18s

real	1m18.798s
user	1m12.865s
sys	0m5.807s

then clean and after building once:

$ touch src/main.rs 
$ time cargo +1.56.1 build
   Compiling rs157-compile-regression v0.1.0 (/home/clux/repos/repros/repros)
    Finished dev [unoptimized + debuginfo] target(s) in 3.12s

real	0m3.158s
user	0m2.700s
sys	0m0.458s

Dramatic difference.

[nightkr]

I can reproduce this with just futures, so this seems more like a rustc issue than kube-rs in particular. Filed rust-lang/rust#91598.

[clux]

This is is presumably no longer an issue with the fix getting into stable. Are you OK with bumping the msrv from this point on?

[nightkr]

Happy to close this and unblock the MSRV, I'm back on stable since 1.59.