Is ParAff Actually Alternative?

[eric-corumdigital]

Distributivity: (f <|> g) <*> x = (f <*> x) <|> (g <*> x)

Let r ∷ Ref Int, f = delay 1.0, g = delay 2.0, x = modify (_ + 1) r

Then

  (delay 1.0 <|> delay 2.0) <*> modify (_ + 1) r
= (delay 1.0 <*> modify (_ + 1) r) <|> (delay 2.0 <*> modify (_ + 1) r)

But, by duplicating the Ref.modify (_ + 1) r effect and running both in parallel, r is incremented by 2 rather than 1.

(Yes, I fudged some types here for brevity).

[natefaubion]

I guess that depends on what = we are referring to 😄.

[natefaubion]

More seriously, <|> is a race, and losers are cancelled, so I think this particular example is bad since it won't be run twice.

[eric-corumdigital]

@natefaubion I am using the ParAff instance for both <|> and <*>, so it should run twice.

[natefaubion]

Again, I think this kind of comes down to what you mean be =. If it's observationally equal after interpretation, than distributivity can never be satisfied by anything effectful or referential. That sounds very strict and greatly limits what is Alternative. Maybe there needs to be a middle ground in the hierarchy?

[eric-corumdigital]

The main interest of Aff is effects. I am just looking at how usual effects work, like printing to the console, and making appropriate generalisations.

Strict or not, that just seems to be the way things are.

[artemisSystem]

wouldn't (g <*> x) be cancelled before it could run its modify, though?

[safareli]

no as the <*> is of ParAff which executes both g and x in parallel