Cannot configure containers with tests and benchmarks enabled

[m-renaud]

I'm not sure if this is a containers problem or cabal problem, but cabal configure --enable-tests --enable-benchmarks is unable to solve the constraints:

$ cabal configure --enable-tests --enable-benchmarks

Resolving dependencies...
Warning: solver failed to find a solution:
Could not resolve dependencies:
rejecting: containers-0.5.10.2:!bench (constraint from config file, command
line flag, or user target requires opposite flag selection)
trying: containers-0.5.10.2:*bench
unknown package: random (dependency of containers-0.5.10.2:*bench)
Dependency tree exhaustively searched.
Trying configure anyway.
Configuring containers-0.5.10.2...
cabal: Encountered missing dependencies:
ChasingBottoms -any,
HUnit -any,
QuickCheck >=2.7.1,
criterion >=0.4.0 && <1.3,
random <1.2,
test-framework >=0.3.3,
test-framework-hunit -any,
test-framework-quickcheck2 >=0.2.9

This causes our instructions in CONTRIBUTING.md to be more complex than they should be.

[hvr]

kinda related discussion at haskell/cabal#4981 (comment)

[sjakobi]

With regards to benchmarks I think we should switch from criterion to gauge once a new release of gauge has been made. The upstream version of gauge has neither direct nor transitive dependencies on containers.

[treeowl]

@sjakobi, how does gauge compare to criterion? What else do we gain/lose? How hard would it be to port the existing benchmark suite?

[treeowl]

@sjakobi, I took a glance at gauge. I think it would be good to add support for it, but it's clearly less capable than criterion, so I'm reluctant to make a hard switch. I also must say I'm not entirely sold on the way it cuts down dependencies, seeming to focus for now on human-readable output. But I can probably be convinced.

[sjakobi]

gauge is just fork of criterion with much fewer dependencies and a few additional features.

how does gauge compare to criterion?

See https://github.com/vincenthz/hs-gauge#features-compared-to-criterion.

What else do we gain/lose?

gauge only supports GHC versions >= 7.8.

How hard would it be to port the existing benchmark suite?

So far it's been as simple as replacing the imports of Criterion and Criterion.Main with import Gauge.

it's clearly less capable than criterion

What features do you miss the most?

[sjakobi]

Here's a trick for building tests and benchmarks with cabal:

mv containers.cabal kontainers.cabal
sed -i 's/name: containers/name: kontainers/1' kontainers.cabal 
sed -i 's/containers,/kontainers,/g' kontainers.cabal

This renames the package to kontainers, removing the circular dependency.

[sjakobi]

This has been fixed in #626.

To avoid the test dependencies getting rebuilt with every library change, it's enough to remove containers from the packages stanza in the cabal.project. It would be good to document this properly though…

[sjakobi]

@m-renaud Can this issue be closed?