winbuilder's note on namespaces

[tkonopka]

Hello reticulate -

I maintain a package that uses reticulate to offer an optional feature, here. This works great for me, but recently started giving a 'note' through cran's winbuilder service.

For background, my package uses .onLoad following your advice. Because the python feature is nonessential, my .onLoad first checks if a python module is available using reticulate::py_module_available and then imports it using reticulate::import.

On the winbuilder service at CRAN, the package now takes a very long time to install (several hours) and generates a 'note' about namespaces.

** checking whether the namespace can be loaded with stated dependencies ... NOTE
Warning in system2(command = python, args = paste0("\"", config_script,  :
  running command '"C:/Users/CRAN/AppData/Local/r-miniconda/envs/r-reticulate/python.exe" "d:/RCompile/CRANpkg/lib/3.6/reticulate/config/config.py"' had status 2

A namespace must be able to be loaded with just the base namespace
loaded: otherwise if the namespace gets loaded by a saved object, the
session will be unable to start.

I do not see this note on travis and on a linux system (R 3.6.1-3). Have you encountered this, and could you offer some tips please? Thanks.

[kevinushey]

The problem is that this line:

https://github.com/tkonopka/umap/blob/master/R/umap.R#L33

forces reticulate to initialize Python. You should only use reticulate::import() here.

[tkonopka]

Thank you!

Turned out a one-line onLoad still gave me the same note on winbuilder. But the explanation put me on the right track and after a few more tries, I found something similar that worked.

In the R directory (using 'scipy' in this example):

# python 'scipy' module
scipy <- NULL
.onLoad <- function(libname, pkgname) {
  try({
    scipy <<- import("scipy", delay_load = TRUE)
  }, silent = TRUE)
}

And in the test suite:

# determine if the system has 'python' as well as 'scipy'
has_scipy <- FALSE
try({
  has_scipy <- reticulate::py_module_available("scipy")
}, silent = TRUE)

# run test suite
if (has_scipy) {
  test_that("Things work as expected", {
    # test code here...
  })
  # more test cases here ...
}

Without the try in onLoad, the sequence document(); test(); test(); from devtools stops when the system has python but does not have 'scipy'.

The try in the test suite seems to have no consequence on a computer with python. But it could be helping to avoid a warning in winbuilder.

[tkonopka]

With a few more tests, it appears that winbuilder now tolerates cases without the try blocks, including versions that generated a note earlier. So this issue may have been a temporary glitch.

In any case, thank you very much for your help.