It would let us have some cake. Users would be able to test against 0.9, in theory. But they'd have to do it intentionally. And Stack-based projects would probably need some shenanigans to deal with a version of containers that doesn't come with GHC. So I really think that avoiding these subtle bugs requires at least a full GHC release cycle.

On Feb 13, 2018 5:48 PM, "Herbert Valerio Riedel" <hvriedel@gmail.com> wrote:

Hi,

On 2018-02-13 at 17:41:25 -0500, David Feuer wrote:
> We may not have to be *quite* as long-winded as I initially suggested,
> but I don't think your way is sufficiently long-winded. The advantage
> of removing the instances first and then adding them back is that the
> version(s) without the instances will make compilation of every single
> package that uses them fail. The way you suggest, if someone has a
> package using containers or unordered-containers, they don't even have
> a simple way to find out whether they need to make changes. On the
> opposite side, Chris Wong's suggestion would let us be very explicit,
> with a type error that warns users that the instance is missing
> *because it is changing* and that no code using the instance will work
> for both 0.* and 1.* versions.

Alright, then let's do a little Gedankenexperiment; what if you release
a containers-0.9.0.0 (which drops the instances) and a
containers-1.0.0.0 (which 'adds back' the desired new instances) on the
same day!

...wouldn't this allow us to have the cake and eat it too? ;-)