Why does cabal install reinstall packages already in .cabal/lib

Question

I'm attempting to use CircleCI to build some Haskell projects. Circle automatically caches the /home/ubuntu/.cabal directory after each build and I thought that would speed up the builds, but they seem to take just as long as before---all that's skipped is the download step.

The build instructions I'm using are

cabal update
cabal install --only-dependencies --enable-tests
cabal configure --enable-tests
cabal build
cabal test

and I anticipated that install --only-dependencies step would see the cache and run more quickly.

What might be going wrong?

Answer 1

There was an excellent Reddit discussion of cabal issues a few months back. My recollection is that Haskell has no specified application binary interface, and that the ghc compiler does a lot of aggressive inter-library graph reductions during compilation. As a result, there is no such thing as a stable library installation in a Haskell app -- any library is liable to be dragged down and recompiled against the full set of already-installed libraries.

The bottom line here is that caching the .cabal/lib directory does nothing to speed installation. Recompilation of dependencies will always happen. Your best approach, aside from sandboxing using hsenv, is to specify your full set of dependent libraries when you call cabal install. That way, the dependencies can be calculated ahead of time, and compilation should only occur once.

Haskell gurus, please feel free to correct me if anything I said here is mistaken. I have been out of the Haskell loop for a few months, and I'm aware that cabal/library issues have received a lot of attention lately.