Error converted from warning prevents package install when dependency was built under newer R version

[psolymos]

I am having R version 3.6.0 in Windows, and when I try to use remotes::install_github I get this an error converted from a warning that package 'x' was built under R version 3.6.1. I was trying to override the behaviour at the function call level using dependencies=FALSE to no avail.

> remotes::install_github("psolymos/bSims")
Downloading GitHub repo psolymos/bSims@master
[...]
* installing *source* package 'bSims' ...
** using staged installation
** R
** inst
** byte-compile and prepare package for lazy loading
Error: (converted from warning) package 'mefa4' was built under R version 3.6.1
Execution halted
ERROR: lazy loading failed for package 'bSims'
* removing 'C:/R/R-3.6.0/library/bSims'
* restoring previous 'C:/R/R-3.6.0/library/bSims'
Error in i.p(...) : 
  (converted from warning) installation of package ‘C:/Users/PETERS~1/AppData/Local/Temp/RtmpSssGwE/file3d706900445a/bSims_0.1-2.tar.gz’ had non-zero exit status

> remotes::install_github("psolymos/bSims", dependencies=FALSE)
Downloading GitHub repo psolymos/bSims@master
[...]  
* installing *source* package 'bSims' ...
** using staged installation
** R
** inst
** byte-compile and prepare package for lazy loading
Error: (converted from warning) package 'mefa4' was built under R version 3.6.1
Execution halted
ERROR: lazy loading failed for package 'bSims'
* removing 'C:/R/R-3.6.0/library/bSims'
* restoring previous 'C:/R/R-3.6.0/library/bSims'
Error in i.p(...) : 
  (converted from warning) installation of package ‘C:/Users/PETERS~1/AppData/Local/Temp/RtmpSssGwE/file3d7038435b8a/bSims_0.1-2.tar.gz’ had non-zero exit status

Is this warning-to-error conversion really necessary in this case? Seems to me a bit overkill for same major R version.

[gaborcsardi]

See R_REMOTES_NO_ERRORS_FROM_WARNINGS at https://github.com/r-lib/remotes#environment-variables.

[psolymos]

Worked. Thanks.

[skranz]

Just a short comment in case you are considering to change this default setting of remotes.

The functionality to easily install from Github you provide is really great. But I found it extremely confusing that warnings are by default converted to errors. It took me a lot of time to figure out why the installations of some Github packages did not work. For me it would have been much easier if you had followed the default R convention to only throw an error if somebody set options(warn=2).

When I got such a warning error from devtools:install_github , I wrote several times options(warn=0) and really did not understand what is going on and why I can't install.

I am not sure whether your assessment is empirically true for Github packages:

Warnings usually mean installation errors, so by default remotes stops for a warning.

I think particular on Github there are packages that have been programmed quickly and don't follow all CRAN guidlines. (Several of my own packages are such bad examples). For non CRAN compliant packages there can be many relatively irrelevant warnings like

Warning:  codeUtils/man/codeUtils-package.Rd:32: All text must be in a section

or

Error: (converted from warning) package 'restorepoint' was built under R version 3.6.1

I personally see Github a bit as a land of freedom where I can write quickly packages without spending a lot of time to adhere to all CRAN conventions. But the new remotes default to throw an error if there is a warning basically enforces a lot of CRAN conventions in order to install a package without diving into how to change environment variables.

In short, this is just +1 for not converting warnings to errors by default.

[cdriveraus]

Just want to drop in to join the argument against this default. Failing a package build because of the warning that a different package was built in a slightly newer version seems a bit much... the user could at least be presented with an option to continue, if in interactive mode...

[gaborcsardi]

the user could at least be presented with an option to continue, if in interactive mode...

That is not possible, because the installation runs in a subprocess by install.packages().

[thomas-neitmann]

Even though I am late to this discussion I'd like to give my 2 cents on this. After a lot of searching I found out that the reason why people got an error trying to install my ggcharts package is because of this behavior to convert all warnings to errors.

Most R users on Windows do not install packages from source (due to not having installed Rtools) and thus end up with packages that have been built on newer R versions at one point. In 99.9% of cases this is not an issue and you can safely ignore the warning that package xyz was built under R version x.y.z. In think this is a great example that the statement

Warnings usually mean installation errors

is wrong. I would thus strongly suggest to change this behaviour. After all, why be stricter than install.packages()?

[TillF]

+1 for changing the default behaviour to ignore warnings

[hadley]

It looks like this behaviour is causing a lot of pain so I think we should reconsider. It seems like the other failure mode (apparent install success but actual install failure with warning) is less common.