[Rd] R 3.5.0 fails its regression test suite on Linux/x86_64

[Ricardo Wurmus]

Hi Dirk,

> On 23 April 2018 at 12:33, Peter Simons wrote:
> | Hi,
> |
> | I just tried to upgrade Nixpkgs to R 3.5.0, but unfortunately the new
> | version fails its regression test suite. We configure the build using
> | the flags "--without-recommended-packages", in case that's relevant. You
> | can see a complete build log with all relevant information at [1].
> | Anyway, the test failures look like this:
>
> The flag instructs R not to build new 'recommended packages'; however the
> tests you run later need them.

Is there a way to disable only those tests that depend on the
recommended packages?  Would it make sense to disable these tests by
default when “--without-recommended-packages” is passed?

> This release has a binary change so you
> actually need to rebuild dependent packages -- in other words I think you can
> expect (some of) these tests to fail until these recommended packages are
> rebuilt.

I’ve hit the same problem as Peter in upgrading R for Guix.  We build
the recommended packages as separate packages later, so that we can
offer a minimal R package.  At the time when R itself is built the
environment does not contain any R packages, so there is nothing to be
rebuilt.

Another problem with building the recommended packages as part of the R
build process is that it’s more difficult to make them build
bit-reproducibly.  Building them as separate packages simplifies that
and enables us to prevent e.g. the embedding of timestamps.

--
Ricardo