Setting R_LIBS_USER to nonexistant file breaks R CMD build in R <= 3.3 about callr

[gaborcsardi]

I believe callr sets the library path to the same as in the calling R process. Can you try to reproduce with callr?

[krlmlr]

callr seems to set the R_LIBS_* variables to a nonexistent dummy file, like in this example. Still working on replicating it locally.

[gaborcsardi]

Yes, but it also sets the library path from the R profile.... maybe the profile is sometimes ignored?

[krlmlr]

I've got a reprex that involves callr, it fails with R 3.3.2 but succeeds with R 3.4.

To reproduce, pick any package that uses knitr as a vignette builder, adapt the path in the first line, and run the script. The knitr package must not be installed in the local library, this is checked by the reprex.

# Adapt this path
pkg_to_build <- "~/git/R/styler"

stopifnot(!("knitr" %in% rownames(installed.packages())))

lib <- tempfile()
dir.create(lib)
install.packages("knitr", lib)

stopifnot(!("knitr" %in% rownames(installed.packages())))

Sys.setenv(R_LIBS_USER = paste(c(lib, .libPaths()), collapse = ":"))

system(paste0(
  R.home("bin/R"), " ",
  "-q -e ",
  "'callr::rcmd(\"build\", \"", pkg_to_build, "\")'"
))

Output with R 3.3.2, consistent with what we see on tic-enabled repos on Travis CI:

...
> callr::rcmd("build", "~/git/R/styler")
$status
[1] 1

$stdout
[1] "* checking for file ‘~/git/R/styler/DESCRIPTION’ ... OK\n* preparing ‘styler’:\n* checking DESCRIPTION meta-information ... OK\n* installing the package to build vignettes\n* creating vignettes ... ERROR\nError in loadVignetteBuilder(vigns$pkgdir) : \n  vignette builder 'knitr' not found\nCalls: <Anonymous> -> loadVignetteBuilder\nExecution halted\n"

$stderr
[1] ""

$timeout
[1] FALSE

$command
...

I'm very sorry...

---

(The original submission had missing parens after installed.packages(), I just double-checked, it still fails.)

[gaborcsardi]

I'm very sorry...

No, thanks for creating the reprex!

I guess for building a vignette, R starts another R process, and that does not inherit the library path properly. We can just set R_LIBS_USER as well, and also create a fake .Renviron file.

It is relatively easy to call an R process with the same config, but to make sure that the other R processes started from that (not via callr) have the same config, now that is quite challanging....

[krlmlr]

Thanks for looking into it. Sorry for throwing these meta³ problems on you, unless of course you enjoy solving them ;-)

[gaborcsardi]

Can you pls try with the latest master?

[krlmlr]

Test-driving now, will comment if anything suspicious pops up.