[r-package] cut CI-time dependency on craigcitro/r-travis (fixes #4348)

[jameslamb]

Currently, this project's R stage on Travis depends on craigcitro/r-travis, a project that was officially deprecated in July 2016. In this PR, I propose a way to factor it out.

See #4348 and other linked issues there for more details.

Thanks for considering it!

[trivialfis]

@jameslamb Out of curiosity, I planed to learn R at some point for its abundant statistic libraries. But I found there's no implementation of R semantic tools for code navigation. Closest one I can find would be https://github.com/REditorSupport/languageserver , but their TODO list seems to be too long. Do you use any of these stuffs?

[hcho3]

@jameslamb I changed Travis CI script to install R.

[jameslamb]

Thanks @hcho3 . I see now that the R side of Travis is passing!

Is the failing Python task related to this? Or is it something transient that you've seen before?

[jameslamb]

@trivialfis Thanks for the link, I've never seen this before!

I'm not familiar with language servers, but is the general idea here that you have a particular IDE you like and you want to be able to develop R code in it and get features like autocompletion, syntax highlighting, error flags, etc.?

RStudio is a wildly popular free, open source IDE amongst R developers. The source code is at rstudio/rstudio if you want to look around for how they do things like autocompletion of R code.

[jameslamb]

CI restarted because I just rebased to current master and re-pushed (in case reviewers are wondering)

[jameslamb]

Oh wait @hcho3 I still see a failure of the R test (https://travis-ci.org/dmlc/xgboost/jobs/518574498)

I think the bit of .travis.yml you edited will only affect Linux builds, but currently the xgboost R-package only gets tested in Mac OS. I'll go back to r-travis, I'm sure I just missed some step hidden in the many layers of indirection there.

[jameslamb]

I am going to try using conda to install R on Mac. The things being done in r-travis required sudo which I feel weird about.

The conda approach has worked for me before, e.g. https://github.com/jameslamb/doppel-cli/blob/master/.ci/setup.sh#L19

[jameslamb]

oh @hcho3 I see you just pushed a commit to update the homebrew deps. I did not notice that in .travis.yml! Let's see if it works.

If not, I'm going to try this in setup.sh

if [ ${TASK} == "r_test" ]; then

    conda install \
        -y \
        -c r \
        r-essentials

fi

[jameslamb]

@hcho3 for my own knowledge, why was it necessary to change to cloud.r-project.org from the RStudio mirror?

[hcho3]

From this page: "Automatic redirection to servers worldwide, currently sponsored by Rstudio"

[jameslamb]

thank you

[jameslamb]

pushed a commit to resolve this error:

* checking package dependencies ... ERROR
Packages suggested but not available:
  ‘knitr’ ‘rmarkdown’ ‘testthat’ ‘lintr’

@hcho3 what is xgboost's approach to dev version of the R package on Github? I see this warning:

* checking CRAN incoming feasibility ... WARNING
Maintainer: ‘Tong He <[email protected]>’
Insufficient package version (submitted: 0.82.0.1, existing: 0.82.1)

Is it just a mistake in the DESCRIPTION? Should current version on GitHub actually be 0.82.1.9000? Many R projects follow convention <cran_version>.9000

[jameslamb]

Good news: green build on the R test after my most recent commit!
Bad news: R test took 36 minutes installing a million dependencies

In f843af3 I removed rmarkdown and knitr (should not be needed since we use --no-vignettes) and added the environment variable to suppress R CMD CHECK failing a build because of packages in Suggests not being present in the environment.

[codecov-io]

Codecov Report

Merging #4353 into master will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##           master    #4353   +/-   ##
=======================================
  Coverage   67.96%   67.96%           
=======================================
  Files         132      132           
  Lines       12251    12251           
=======================================
  Hits         8326     8326           
  Misses       3925     3925

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 3078b59...f843af3. Read the comment docs.

[jameslamb]

Alright, R tests are working! https://travis-ci.org/dmlc/xgboost/builds/519083803

They still take 35 minutes though, not sure how to work around that. DiagrammeR is just a heavy dependency with a lot of recursive dependencies.

I see that the python test is failing on the build. Do you think it is related to the current state of this branch, or just transient?

[hcho3]

It’s transient failure. I restarted it. As for the long setup time, we can use Docker to cache builds in Jenkins (when R tests get migrated there)

[hcho3]

Merging for now.