Basic GitHub actions-based CI

[gdeest]

(New version of #1396 from a Servant repository branch, in order to get GitHub actions actually running here.)

The Travis CI has been broken for a while, and it is probably time to fix it so that proper development can continue.

This PR:

• Setups a basic CI based on GitHub actions, with a somewhat limited build matrix.
• Disables cookbook/testing, because servant-quickcheck doesn't build anymore.
• Disables servant-docs on Cabal build, because of some test failures
  • The order of some JSON fields seems to be reversed in the output, need investigation.
• Fix test failures in servant-http-streams when localhost points to an IPv6 address rather than 127.0.0.1.

If this PR is merged, I plan on fixing and re-enabling all the disabled tests one-by-one. I think the first step here is to get back to a green CI to prevent further regressions.

Opening as draft for now because it still needs a bit of cleanup. An example run can be witnessed here:

https://github.com/gdeest/servant/runs/2024096343

[gdeest]

For the record, here is what seems to be missing ATM:

• Support for additional GHC versions (it's just a matter of adding them to the build matrix, and potentially fixing associated issues).
• GHCJS tests for the relevant parts of the codebase.
• Doctests.

[gdeest]

I noticed the following weirdness in servant-client tests:

Preprocessing test suite 'readme' for servant-client-0.18.2..
Building test suite 'readme' for servant-client-0.18.2..
Running 1 test suites...
Test suite readme: RUNNING...
Error: ConnectionError (HttpExceptionRequest Request {
  host                 = "localhost"
  port                 = 8081
  secure               = False
  requestHeaders       = [("Accept","application/json;charset=utf-8,application/json")]
  path                 = "/books"
  queryString          = ""
  method               = "GET"
  proxy                = Nothing
  rawBody              = False
  redirectCount        = 10
  responseTimeout      = ResponseTimeoutDefault
  requestVersion       = HTTP/1.1
  proxySecureMode      = ProxySecureWithConnect
}
 (ConnectionFailure Network.Socket.connect: <socket: 3>: does not exist (Connection refused)))
Test suite readme: PASS
Test suite logged to:
/home/runner/work/servant/servant/dist-newstyle/build/x86_64-linux/ghc-8.8.4/servant-client-0.18.2/t/readme/noopt/test/servant-client-0.18.2-readme.log

This may be another IPv6 / network-related issue. But I am puzzled by the fact that the test suite is marked as “passing” when it is clearly failing. Needs investigation. This is normal. This is a sample program in README.lhs to demonstrate basic usage of servant-client. The program tries to connect on localhost, which fails because there is no server running, and displays the error in the console. This is not a test failure.

[gdeest]

We are hitting: sol/doctest#119

When running doctest the dumb way. There doesn't seem to be an easy way around this (though I am unsure why current setup doesn't seem affected base-compat-batteries is excluded from doctest setup, see below).

I'll give a stab at it using cabal-doctest, and writing dedicated test suites for the doctests. This seems to be the simplest way to invoke GHC with just the required dependencies.

PackageImports would work as well, but this would be quite an obtrusive change to the codebase for the sole purpose of running tests.

[dfithian]

FWIW I was able to get tests to pass by testing twice, with each field.
https://github.com/haskell-servant/servant/pull/1401/files#diff-1d3cfc9328ae9300f3c558d76c2bab79a5eee646734eac7500f828cd318275f1

[gdeest]

That's good to know ! There is a pending PR ( #1397 ) that fixes this problem in a slightly different way. We'll ponder what the best approach is.

[gdeest]

The Travis CI works around the ambiguous module problem with doctests with:

doctest-filter-packages: base-compat-batteries

in cabal.haskell-ci. It gives me hope that this seems to be the only package requiring this ; I may be able to reproduce what current .travis.yml is doing easily in the Github setup without going down the custom setup / dedicated test-suite road (even though I think this might ultimately be better).

[gdeest]

The doctests were actually only failing on MacOS, which happened to run faster than Ubuntu and cancelled the entire build.

Disabling the MacOS builds for now since the Travis CI doesn't have them anyway, and we're only aiming for feature-parity at this point.

[gdeest]

The UVerb cookbook is triggering a failure on GHC < 8.6 because of a usage of DerivingVia (which was introduced in GHC 8.6.1).

We will modify the cookbook not to use this extension.

[gdeest]

CI is fixed, and (AFAIK) has feature-parity with previous setup. Since the Travis setup is basically broken, I'll merge now.

A few trivial changes were required:

• New cookbooks did not compile with older GHC versions, due to to the use of extensions that weren't available at that time (DerivingVia, DerivingStrategies) or to changes in the structure of base (missing imports for Data.Monoid / Data.Semigroup).
• A test was fixed that involved a problem when localhost is resolved to ::1 instead of 127.0.0.1.

A servant-docs test was commented out for now. It is inherently buggy, as it relies on the order in which aeson will render JSON object fields, but aeson uses a HashMap from unordered-containers for representing objects. A fix is proposed in #1397.

Things to improve on:

• Performance: CI is currently taking in the order of ~20 minutes. There is caching in place, but the cache doesn't seem to always be restored.
• GHCJS setup. We're still using hvr's PPA which doesn't seem to be updated anymore (we're stuck with ghcjs-8.4). We could try to use Nix / reflex-platform.
• General ugliness / hacksinherited from Travis CI.