Cabal should use the minimum version that satisfies thedependencies, not the maximum

Simon Marlow simonmar at microsoft.com
Mon Aug 8 04:13:11 EDT 2005


On 08 August 2005 07:45, Brian Smith wrote:

> Let's say we have multiple versions of a package installed:
>     a-1.0
>     a-1.0.1
>     a-1.1
>     a-2.0
> and that we have the following Build-Depends in a package description:
>     Build-Depends: a >= 1.0
> 
> Now, when I configure, Cabal will choose a-2.0 as the package to build
> with. However, it makes more sense to choose a-1.0 by default instead.

I don't agree, there is just as much potential for breakage with this
choice, and it will cause problems if someone is trying to upgrade their
system incrementally.  If I have both a-1.0 and a-2.0 installed, I
probably don't want any new dependencies on a-1.0, because I want to
delete it at some point.

> * Currently, if we change the program to require a feature that is in
> a-2.0, but not in a-1.0, then the build will succeed for _us_, but it
> will fail unexpectedly for someone else that has only a-1.0. However,
> if Cabal instead chose the minimum version available, then, _our_
> build would fail, and we would know that we need to either remove the
> dependency on the a-2.0 feature or increase the version specified in
> Build-depends.

Conversely, if the author builds with a-1.0, and the recipient builds
with a-2.0 and version 2.0 is not backwards compatible, then the package
will break.

If the package depends on a-1.0 rather than a>=1.0, then it should say
so.

> * a-2.0 is not guarenteed to be 100% backwards-compatible with a-1.0.
> Even a-1.0.1 is not guarenteed to be backwards-compatible. So, if the
> original author was really testing with a-1.0, then we are more likely
> to run into problems with a-2.0 (or even a-1.0.1) than a-1.0.

This is true, but most people will not have multiple versions of a
package installed.  Furthermore, we'd like to discover problems early,
and this just hides the bug in the dependency specification.

Cheers,
	Simon


More information about the Libraries mailing list