[R] could XGBoosterPredict_R be removed?

[jameslamb]

#6819 introduced a new R-to-C interface for predicting, replacing XGBoosterPredict_R().

As of that PR, XGBoosterPredict_R() is no longer used.

• no R code in the R package calls its
• none of the R package's tests call it

Would you support a PR to remove it?

Why I think it should be removed

I don't think that would be a user-facing breaking change, unless this project supports other R packages linking against the lib_xgboost.{so,dll,dylib} built for the R package, or supports invoking those R-to-C entrypoints directly from R code with :::.

And removing it would have the following benefits:

• smaller built library size
• faster compilation
• reduced risk of spending time fixing warnings from static analyzers (esp. checks run by CRAN) about code that isn't actually used in the package

How I found this

used covr to find bits of code not covered by tests (click me)

Ran the following to generate a coverage report.

cd R-package/
Rscript -e "install.packages(c('covr', 'DT'), repos = 'https://cran.r-project.org')"
Rscript -e " \
    coverage <- covr::package_coverage(type = 'tests', quiet = FALSE);
    print(coverage);
    covr::report(coverage, file = file.path(getwd(), 'coverage.html'), browse = TRUE);
    "

That showed the following coverage:

xgboost Coverage: 84.73%
R/xgb.create.features.R: 0.00%
R/xgb.load.R: 40.00%
R/xgb.cv.R: 63.89%
R/xgb.save.R: 66.67%
R/xgb.DMatrix.save.R: 71.43%
R/xgb.DMatrix.R: 72.51%
R/xgb.plot.importance.R: 75.00%
R/xgb.plot.shap.R: 76.24%
R/xgb.ggplot.R: 77.05%
R/xgb.plot.deepness.R: 80.39%
R/utils.R: 82.24%
R/callbacks.R: 84.38%
R/xgb.dump.R: 85.71%
src/xgboost_custom.cc: 85.71%
R/xgb.Booster.R: 89.51%
R/xgb.load.raw.R: 90.00%
R/xgb.plot.tree.R: 91.67%
R/xgb.model.dt.tree.R: 92.11%
R/xgb.train.R: 92.21%
src/xgboost_R.cc: 92.29%
R/xgb.unserialize.R: 92.86%
R/xgb.importance.R: 93.33%
R/xgb.plot.multi.trees.R: 97.40%
R/xgb.config.R: 100.00%
R/xgb.save.raw.R: 100.00%
R/xgb.serialize.R: 100.00%
R/xgboost.R: 100.00%
src/init.c: 100.00%

And I saw in the clickable report that XGBoosterPredict_R is never called by any test code.

[trivialfis]

Thank you for checking the code! Please remove the function.

At the time of writing the new function, I talked with others and was told that the C package is considered a part of the interface. I guess it's not really an issue for XGBoost since we never publish any formal document about the R-C packagae.

[jameslamb]

Ok great, thanks very much! I'll put up a PR shortly.

since we never publish any formal document about the R-C package.

I agree with treating this as a private interface. I'd be surprised to learn that there are many use cases for others to link to the R-to-C library, instead of either using the C API directly or the R package directly.