Change union to be right-biased

[fehrenbach]

I'd quite like this to hold:

fromFoldable (a <> b) == fromFoldable a <> fromFoldable b

fromFoldable is repeated insertion by foldl. I assume the reason is to match Haskell's fromList which uses foldl', which makes sense for lists.
(<>) is union is unionWith const and that simple implementation seems to be the only reason (and Haskell does it that way).

What do people think?

[LiamGoodacre]

My thoughts on this are that I'd rather either not have a direct Semigroup instance or pick (Ord k, Semigroup v) => Semigroup (Map k v) (which, I feel, is a more obvious instance). Then let newtypes deal with the First/Last behaviour. In a similar way that we don't pick an instance for Int when there are multiple useful interpretations.

[fehrenbach]

Good point.

I'm not sure I agree with the "more obvious" part, but still, good point.

I quite like using foldMap (\x -> singleton (f x) x) to index an array which would not work without/with a different Semigroup instance. That's not a good argument against, though. That's an argument for fromFoldableBy or something.

[natefaubion]

This is a 👎 for me. Left-biased union is pretty much standard everywhere in our ecosystem, including Prim.Row.Union, for better or worse. We have so much code at work that depends on the union ordering. I would be ambivalent about the Semigroup changes with coerce and if we had type applications.

[LiamGoodacre]

(Side note: I wouldn't describe Prim.Row.Union as left-biased. It acts more like unionWith (++) - treating a row effectively as a Map String (Array Type) - multiple entries for the same key are preserved left to right.)

[natefaubion]

It is effectively left-biased when paired with Nub, which is very common. Nub technically has the bias, but the behavior of Nub is also consistent (keep the first, ie left-biased value).

This has more of an effect with Record, which is that the first (left-most) type for a row is the actual runtime value.

[LiamGoodacre]

I thought a record type with duplicates keys was uninhabited.

[natefaubion]

No, you can get duplicate labels via union. Record delete and insert are just predicated on Lacks constraints, and you can't construct a record literal that's typed as having duplicate labels.

[hdgarrood]

I’m going to close this, see #4 (comment)