The stack CI already tests against 9.2.7 instead of 9.2.5 (as the cabal
file said), since we just specify the major version of the LTS resolver,
and lts-20 is up to 9.2.7 by now. Not much has changed since 9.2.5, but
update the haskell-ci workflow regardless. Likewise, 9.4 can be bumped
one minor version, so do that as well.
hlint 3.5 is built against libtinfo6 and works on Ubuntu 22.04 without
needing to install libtinfo5.
Fixes: b7afb0c0bc3f ("ci: Pin runner in hlint, nix workflows to ubuntu-20.04")
Invoking cabal in the project directory with a cabal.project file might
fail on unrelated dependencies missing, see
4f539734be
Doesn't fail in this repo, but let's keep the workflows in sync.
Both are now broken with ubuntu-22.04. Nix:
error: could not set permissions on '/nix/var/nix/profiles/per-user' to 755: Operation not permitted
hlint:
/opt/hostedtoolcache/hlint/3.4.1/x64/hlint: error while loading shared libraries: libtinfo.so.5: cannot open shared object file: No such file or directory
Breaks `apt update` on the ubuntu-22.04 runner, which is what
ubuntu-latest resolves to these days.
Fixes: 7312d6f3f3f7 ("ci: Drop GHC installation via apt")
haskell-ci by default passes --haddock-all to cabal v2-haddock, which
builds docs for all components and the …-docs.tar.gz tarball ends up
containing the docs for the last component, which happens to be tests.
We need the tarball to contain the library docs, for upload to Hackage.
Fixes: 7cee1915169d ("CI: Remove hlint from haskell-ci")
The apt repo hosting these packages has been unmaintained for quite a
while, and we've excluded ~/.stack/programs from the cache long time
ago, so we can just let stack handle the installation. It's not much
slower than installing via apt, and even if it was, some matrix jobs
need to use the slow path anyway.
Related: 9fce3805fcf2 ("ci: Use system GHC in Stack to not waste GH Actions cache space")
Related: 7d10e470d71d ("ci: Avoid caching GHC")
+ Prefer GHC 8.10.7 to 8.10.4, as versions seem to have stabilised now.
+ Add support for Stackage LTS 19, which ships with GHC 9.0.2.
+ Since a new version of 9.2 has been released, prefer 9.2.2 over 9.2.1.
Also, explicitly pass -XCPP to hlint; see [1].
[1]: https://github.com/ndmitchell/hlint/issues/1360
During the release of xmonad 0.17.0, I realized that we need to be able
to upload candidates before tagging the release on GitHub, because there
might be issues with the tarball and Hackage may reject it. When that
happened, I had to remove the release, delete the tag, upload the
candidate manually to see what's wrong with it, try to fix it, upload it
manually again, and so on.
This commit swaps the logic: when the workflow is invoked manually, it
uploads the candidate. This can be done multiple times, and once
everything is fine, the release can finally be tagged and it's released
to Hackage proper. The only disadvantage is that we need to remember to
try uploading the candidate. Not sure if there's a perfect solution…
This was temporarily disabled in 229d52ff07 and 086db3123b because we
couldn't keep xmonad-contrib backwards compatible with xmonad 0.15 any
more, but xmonad 0.17.0 is out now and we can try to stay compatible for
a while.
We install GHC using apt, so stack shouldn't install it, but should it
ever end up installing it anyway (version mismatch, version unavailable
in hvr/ghc ppa, …), we don't want it wasting valuable cache space.
With "nix build", just the build of xmonad-contrib itself takes 3
minutes (it builds twice, the second build with profiling enabled), so
it ends up being almost 6 minutes in total, making this workflow the
slowest one.
Building xmonad-contrib benefits greatly from compiling modules in
parallel. Vanessa McHale suggested setting ghc-options in
xmonad-contrib.cabal in #431, but that's bad practice—these options
should be set by the user/environment where it's being built, not
selfishly hardcoded for specific package. In stack.yaml and
cabal.haskell-ci, we know for sure that xmonad-contrib is the terminal
(leaf) library, so we can set it there and speed up builds in our CI and
also for anyone who uses our stack.yaml (not that anyone should).
Closes: https://github.com/xmonad/xmonad-contrib/pull/431
Turns out `commit: master` doesn't really mean latest master, because
it's cached. It sometimes works, because when the GHA cache comes from
a job that depended on xmonad-0.15, then the master archive isn't cached
and is fetched. When the cache comes from a job that previously fetched
master, it's not fetched again.
The least ugly way to fix it I found is to use the `[<refname>]@{<date>}`
git revision format, and substitute current date/time before invoking
stack. To make stack-master.yaml valid without the substitution, we use
`master@{today}` which is a valid git revision that GitHub understands.
GitHub Actions writes caches on the first miss and then never updates
them again. If the CI is used frequently, the caches never expire and
as they get old, become less useful.
To avoid this, force refreshing the caches once a month.
When building with an LTS version that has exactly the dependencies we
need (xmonad-0.15), stack doesn't need to download the Hackage index. If
GitHub Actions cache locking chooses this job as the one that writes the
cache, then the "stack-pantry-Linux" cache entry stays empty, possibly
forever.
Force Hackage index update to prevent this from happening.
This further reduces our usage of cache storage by caching the metadata
only once per repo. Now the metadata cache is ~250MB and the individual
caches with built dependencies are ~10M each.
Related: cfe99998fc
Related: https://github.com/xmonad/X11/pull/75
