stack 1.7.1 still has problems finding build tools ("The program 'happy' is required")

[simonmichael]

General summary/comments (optional)

hledger appveyor builds are failing due to not being able to find "happy" when building pretty-show, haskell-src-exts, language-javascript. simonmichael/hledger#832

I saw this a month ago and tried several things to fix it, and probably thought I had solved it by upgrading to stack 1.7.1, but now it seems not.

Possibly related:
#595
#3178
#3249

Steps to reproduce

I don't have a short repro, instead I'll point to appveyor logs.

Expected

Successful builds of pretty-show, haskell-src-exts, language-javascript.

Actual

These packages failed to build due to "missing happy", even with stack 1.7.1. Eg see
https://ci.appveyor.com/project/simonmichael/hledger/build/master-460.

Some more details:

hledger build history: https://ci.appveyor.com/project/simonmichael/hledger/history.
Only the full, >30m builds are building pretty-show etc.

The last full build that succeeded was https://ci.appveyor.com/project/simonmichael/hledger/build/master-451 on 26 May. The stack version is not reported, but it fetched http://www.stackage.org/stack/windows-x86_64, and stack 1.7.1 was released on 27 Apr, so I would have thought it was using stack 1.7.1. Unless there was some path issue.

The next full build, 456 on 30 May, failed with No packages found in snapshot which provide a "happy" executable, which is a build-tool dependency of "pretty-show" warnings and The program 'happy' version >=1.19 is required but it could not be found errors. So did build 460 and later, and these report the stack version as definitely 1.7.1. https://ci.appveyor.com/project/simonmichael/hledger/build/master-460.

Builds 459, 462 and 465 were non-full builds that succeeded by luck, appveyor apparently picks randomly from several build instances each with their own cache state.

Stack version

1.7.1

Method of installation

Official binary, downloaded from stackage.org or fpcomplete's package repository

[mihaimaruseac]

Can you also try with the stack from HEAD? stack upgrade --git and then build again.

I think this has been solved on master, accidentally, but I'm not completely sure.

[simonmichael]

I'll report back.. it's taking a while.

[simonmichael]

I can still reproduce with stack-git: https://ci.appveyor.com/project/simonmichael/hledger/build/master-497

[snoyberg]

As part of #3922, I was discussing the behavior of the build tool detection with @bitemyapp. It might make sense to copy the behavior that cabal-install 2.0 (or 2.2?) added for build tool detection, which unless I'm mistaken comes down to something like:

• For new cabal files, build tools must be specified including the package they come from
• For old cabal files, there's a hard-coded mapping from build tool name to the package it comes from

That would likely bypass many cases where it becomes difficult/impossible to determine which package was intended, as well as allow us to much more lazily parse cabal files (since we needn't know all executables provided by all packages in a snapshot).

[snoyberg]

BTW, adding an explicit stack build happy to your CI scripts should be a decent workaround in the short term unless I'm mistaken @simonmichael.

[snoyberg]

Sorry, last comment for a bit: do you have any kind of local repro for this? I haven't been able to trigger it on my OS X system at least.