[R-package] 32-bit library support

[jameslamb]

Summary

As I've investigated #2960 as part of #629 , I've learned that packages on CRAN have to be able to build on 32-bit Windows.

I believe that we need to add support for a 32-bit build LightGBM to get the R package to CRAN.

Evidence that this is required

I have not yet found any official statement that says "all R packages have to build on 32-bit Windows", but I see strong evidence that that is true.

You can find the current list of check environments at https://cran.r-project.org/web/checks/check_flavors.html. None of those seem to be 32-bit

but in "Writing R Extensions" I see the following:

On systems which support multiple sub-architectures (principally Windows), R CMD check will install and check a package which contains compiled code under all available sub-architectures.

And in the "R for Windows FAQs":

Packages without compiled code nor a configure.win script will run on both 32- and 64-bit R.

Packages with compiled code but no configure.win nor src/Makefile.win file will be built for both when running on a 64-bit version of Windows if both versions of R are installed.

An empty configure.win is treated in the same way as if none existed. Also, there is a list of packages known to have an architecture-independent configure.win hardcoded into R CMD INSTALL, and for these packages, both architectures will be built under the above conditions. Other packages can be installed with configure.win run for just the first architecture by using option --force-biarch.

Any package can be installed for first one architecture and then the other with option --merge-multiarch, but the package source must be a tarball (and as before, running on a 64-bit version of Windows with both versions of R installed).

Finally, a package without a src/Makefile.win file and no or empty or architecture-independent configure.win file can be installed for both architectures from 32-bit Windows if the 64-bit components were selected when R was installed and option --compile-both is given. Obviously, only the 32-bit installation can be tested.

This is why you see pull requests this in other projects: dmlc/xgboost#5777

I've tried looking in the source for R CMD check but it is very difficult to follow.

Motivation

I believe that we need to add support for a 32-bit build LightGBM to get the R package to CRAN.

Description

To check if the R package works with 32-bit R, try the following:

sh build-cran-package.sh
"C:\Program Files\R\R-4.0.2\bin\i386\R.exe" CMD check lightgbm_3.0.0-1.tar.gz --as-cran

References

There was some discussion about this in #1129 , where it was decided that supporting a 32-bit is possible but that the maintenance burden would be too much. Given that it is needed to get LightGBM to CRAN, I think it should be reconsidered.

[jameslamb]

I've updated #2302 , but I'd like to leave this open for a bit to discuss.

I believe it is worth calling this issue [R-package] because this feature specifically serves the interests of the R package (getting onto CRAN) and R-specific solutions (for example gross hacks with #ifdef LGB_R_BUILD) could be options for addressing this.

[StrikerRUS]

I remember we enabled optional building for 32bit in Python-package some time ago. The only change that was needed cmake ... -A Win32.
https://github.com/microsoft/LightGBM/pull/2188/files
BTW, I'm not sure that it is still working as we have no CI for that type of build.

[jameslamb]

oh cool! Hey that's at least something simple I can try, thank you!

[jameslamb]

It was kind of hard to figure out how to replicate the 32-bit R builds. The R CMD check flags like --extra-arch, --multiarch, and --force-multiarch have slightly confusing effects.

Just noting for myself here that this is the command to use to test if the package works on 32-bit Windows:

sh build-cran-package.sh
"C:\Program Files\R\R-4.0.2\bin\i386\R.exe" CMD check lightgbm_3.0.0-1.tar.gz --as-cran

You'll know this works if you see a message in the logs saying the library is being built on a 32-bit platform, like this

* using log directory 'C:/Users/James/repos/LightGBM/lightgbm.Rcheck'                                                                                        
* using R version 4.0.2 (2020-06-22)                                                                                                                                
* using platform: i386-w64-mingw32 (32-bit)                                                                                                                         
* using session charset: ISO8859-1                                                                                                                                  
* using option '--as-cran'    

[jameslamb]

I can confirm that the package is installable with 32-bit R on Windows

* checking for portable file names ... OK
* checking whether package 'lightgbm' can be installed ... OK
* checking installed package size ... OK
NB: this package is only installed for sub-architecture 'i386'
* checking package directory ... OK
* checking for future file timestamps ... OK
* checking DESCRIPTION meta-information ... OK
* checking top-level files ... OK
* checking for left-over files ... OK
* checking index information ... OK
* checking package subdirectories ... OK
* checking R files for non-ASCII characters ... OK
* checking R files for syntax errors ... OK
* checking whether the package can be loaded ... OK
* checking whether the package can be loaded with stated dependencies ... OK
* checking whether the package can be unloaded cleanly ... OK
* checking whether the namespace can be loaded with stated dependencies ... OK
* checking whether the namespace can be unloaded cleanly ... OK

However, the tests are failing

[jameslamb]

here are the full logs from the tests. All failed with this cannot create Dataset handle

error log

R version 4.0.2 (2020-06-22) -- "Taking Off Again"
Copyright (C) 2020 The R Foundation for Statistical Computing
Platform: i386-w64-mingw32/i386 (32-bit)

R is free software and comes with ABSOLUTELY NO WARRANTY.
You are welcome to redistribute it under certain conditions.
Type 'license()' or 'licence()' for distribution details.

R is a collaborative project with many contributors.
Type 'contributors()' for more information and
'citation()' on how to cite R or R packages in publications.

Type 'demo()' for some demos, 'help()' for on-line help, or
'help.start()' for an HTML browser interface to help.
Type 'q()' to quit R.

> library(testthat)
> library(lightgbm)
Loading required package: R6
> 
> test_check(
+     package = "lightgbm"
+     , stop_on_failure = TRUE
+     , stop_on_warning = FALSE
+ )
-- 1. Error: predictions do not fail for integer input (@test_Predictor.R#7)  --
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 2. Error: start_iteration works correctly (@test_Predictor.R#36)  -----------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 3. Error: train and predict binary classification (@test_basic.R#12)  -------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 4. Error: train and predict softmax (@test_basic.R#40)  ---------------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 5. Error: use of multiple eval metrics works (@test_basic.R#65)  ------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 6. Error: lgb.Booster.upper_bound() and lgb.Booster.lower_bound() work as exp
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 7. Error: lgb.Booster.upper_bound() and lgb.Booster.lower_bound() work as exp
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 8. Error: lightgbm() performs evaluation on validation sets if they are provi
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 9. Error: cv works (@test_basic.R#215)  -------------------------------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.cv(...)
 2. data$construct()

-- 10. Error: lightgbm.cv() gives the correct best_score and best_iter for a met
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.cv(...)
 2. data$construct()

-- 11. Error: lgb.train() works as expected with multiple eval metrics (@test_ba
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 12. Error: lgb.train() works with force_col_wise and force_row_wise (@test_ba
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(params = params, data = dtrain, nrounds = nrounds)
 2. data$construct()

-- 13. Error: lgb.train() works as expected with sparse features (@test_basic.R#
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 14. Error: lgb.train() works with early stopping for classification (@test_ba
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 15. Error: lgb.train() treats early_stopping_rounds<=0 as disabling early sto
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 16. Error: lgb.train() works with early stopping for classification with a me
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 17. Error: lgb.train() works with early stopping for regression (@test_basic.
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 18. Error: lgb.train() works with early stopping for regression with a metric
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 19. Error: when early stopping is not activated, best_iter and best_score com
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 20. Error: lightgbm.train() gives the correct best_score and best_iter for a 
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 21. Error: using lightgbm() without early stopping, best_iter and best_score 
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 22. Error: lgb.train() gives same result when interaction_constraints is spec
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(data = dtrain, params = params, nrounds = 2L)
 4. data$construct()

-- 23. Error: lgb.train() gives same results when using interaction_constraints 
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(data = dtrain, params = params, nrounds = 2L)
 4. data$construct()

-- 24. Error: custom objective works (@test_custom_objective.R#43)  ------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(param, dtrain, num_round, watchlist, eval = evalerror)
 2. data$construct()

-- 25. Error: using a custom objective, custom eval, and no other metrics works 
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 26. Error: lgb.Dataset: basic construction, saving, loading (@test_dataset.R#
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.Dataset.save(dtest1, tmp_file)
 2. dataset$save_binary(fname)
 3. self$construct()

-- 27. Error: lgb.Dataset: getinfo & setinfo (@test_dataset.R#29)  -------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. dtest$construct()

-- 28. Error: lgb.Dataset: slice, dim (@test_dataset.R#44)  --------------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.Dataset.construct(dtest)
 2. dataset$construct()

-- 29. Error: lgb.Dataset: colnames (@test_dataset.R#55)  ----------------------
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.Dataset.construct(dtest)
 2. dataset$construct()

-- 30. Failure: lgb.Dataset: Dataset should be able to construct from matrix and
is.na(handle) isn't false.

-- 31. Error: lgb.Dataset$setinfo() should convert 'group' to integer (@test_dat
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. ds$construct()

-- 32. Error: learning-to-rank with lgb.train() works as expected (@test_learnin
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(params = params, data = dtrain, nrounds = 10L)
 2. data$construct()

-- 33. Error: learning-to-rank with lgb.cv() works as expected (@test_learning_t
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.cv(...)
 2. data$construct()

-- 34. Error: lgb.get.eval.result() should throw an informative error for incorr
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 35. Error: lgb.get.eval.result() should throw an informative error for incorr
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lgb.train(...)
 2. data$construct()

-- 36. Error: lgb.load() gives the expected error messages given different incor
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 37. Error: Loading a Booster from a file works (@test_lgb.Booster.R#144)  ---
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 38. Error: Loading a Booster from a string works (@test_lgb.Booster.R#177)  -
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

-- 39. Error: If a string and a file are both passed to lgb.load() the file is u
lgb.Dataset.construct: cannot create Dataset handle
Backtrace:
 1. lightgbm::lightgbm(...)
 4. data$construct()

Error in self$construct() : 
  lgb.Dataset.construct: cannot create Dataset handle

[StrikerRUS]

Hmmm, how is it possible that according to the R Hub we get only one NOTE for 32 bit Windows?

Windows Server 2008 R2 SP1, R-oldrel, 32/64 bit
1 NOTE, 0 WARNINGS, 0 ERRORs

[jameslamb]

Unfortunately the logs link from that R Hub build are no longer available, but if they were you'd see a line like this:

NB: Only installing package for architecture x86_64

It's a frustrating part of using R Hub, it doesn't exactly replicate CRAN as far as I can tell, even when you pass --as-cran. When I have tried with win-builder (https://win-builder.r-project.org/) before, it tried to build with both 32-bit R and 64-bit R and failed for the 32-bit checks.

Honestly, maybe at this point it makes sense to just try submitting to CRAN and see what they say? I think we are close enough that it's worth a first attempt. Then we'll know for sure what needs to be done.

@guolinke since you are the maintainer in R-package/DESCRIPTION, you would have to do the submission to CRAN. Could you try it?

You just have to upload the package here (https://cran.r-project.org/submit.html), and then they'll send you an email (usually within 48 hours), with some description of issues and links to logs from builds.

You can use this source distribution of the package that I just created from master:

lightgbm_3.0.0-1.tar.gz

[jameslamb]

in the meantime I'll submit to win-builder and share the full logs here once they're done, so we can see where we stand with that service.

sh build-cran-package.sh

devtools::check_win_release(
    pkg = "lightgbm_r/",
    args = NULL,
    manual = TRUE,
    email = "jaylamb20@gmail.com",
    quiet = FALSE,
)