Covary F in Message types.

[zarthross]

Adding a covary function to Message and its subtypes that matches the Stream.covary function.

It's implemented as a type cast to avoid the costs of creating a new Request object for something that should be safe to do as a cast.

Alternative implementation without a type cast, (but including an extra allocation of Request is here

[rossabaker]

I thought we had this already. I prefer this to the allocating version.

To gain confidence, we could try this in a spec with a pair of effects that work, and using illTyped, one that doesn't.

[rossabaker]

Out of curiosity, do you need this for rho, or other reasons?

[zarthross]

I'll write up some spec's later, wanted to get buy in before putting in the effort.

I didn't need it for Rho.

[hamnis]

Would this be useful on Media ?

[rossabaker]

Good call, @hamnis. It would thread through Part then, too.

[diesalbla]

Rather than adding a covary method, can we just make the type parameter covariant?

sealed trait Message[+F[_]]

[rossabaker]

Rather than adding a covary method, can we just make the type parameter covariant?

sealed trait Message[+F[_]]

Maybe. It would probably want to start in Media. I expect some operations that take typeclass constraints would have to move to something like fs2's InvariantOps (e.g., as, attemptAs). I don't know what other fallout there would be. Do you want to poke at it and see how sweeping the change is?

[diesalbla]

Sure. Do you mind keeping this PR on hold for a little while, until I have tried those changes?

[rossabaker]

We're already in an RC cycle, with hopes of the release next week that many people have been awaiting for too long. My inclination is to proceed with @zarthross' approach in 0.21, and consider the covariance in the next release.

[zarthross]

Per feedback, I've added covary up to Media so its now on Part as well, and I've added some specs that are hopefully sufficient.

[rossabaker]

Nice work. I'll open a separate ticket so we don't lose @diesalbla's idea for the future, but I think this solves the problem now.

The test failures are unrelated. Two of the newer ones are acting flaky on us.