Non-deterministic behaviour when overlapping packages are installed

[guest]

See  http://stackoverflow.com/questions/3119782/mysterious-cabal-install-problems.

This user took a standard Haskell Platform install, installed a new version of the time package, and then installed random. The random package was rebuilt and installed against the new time package. However, re-running cabal install random installs random again; it doesn't reach a fixed point.

The problem (I presume) is that when cabal-install merges the package databases it chooses which random-1.0.0.2 to use non-deterministically. In fact it seems to depend on the exact package Ids, because I get different results with GHC 6.12.2 where the package Ids of the two random-1.0.0.2 packages are ordered differently.

I realise there are larger issues about how to do resolution here, but I think at least we should make cabal-install deterministic, and preferably idempotent.

[kosmikus]

I should check if this is still a potential issue, also in combination with the modular solver.