@mrdziuban Matt, thanks for the try! Unfortunately, this implementation just does not work in a concurrent environment.

Using a global mutable variable (which is what a Ref is) being rewritten within local scopes is unsafe. This sample would show it:

span("parent") {
  printSpanId >>
  List(
    span("child1")(printSpanParentId),
    span("child2")(printSpanParentId),
    span("child3")(printSpanParentId),
    span("child4")(printSpanParentId),
    span("child5")(printSpanParentId)
  ).parSequence_
}

If you run it against IO + Ref, it will spawn some child spans NOT from the parent span, which is incorrect.

If you run it against Kleisli + IO, it will work correctly.

Ahh okay, thanks for the feedback @catostrophe! Do you think an IOLocal solution like natchez's would work? And if so, would adding a new cats-effect submodule for it work?

As for implementation via IOLocal, this would definitely work. Although you need to take into account something before you start implementing it.

We did some work to allow one to use t4c with a compound context (let's name it Ctx), where a Span[F] is just a part of it. natchez impl doesn't care about it. So please don't copy it as is.

I would also like to have it as a separate module since it's CE3-runtime-specific, and we depend only on ce3-kernel.

Closing this one. You are welcome to open a PR with implementation via IOLocal.

As I mentioned before, this should be a separate module. I thought again regarding Ctx and I think it does not relate to IOLocal implementation. You can copy and adapt natchez's solution.