[R-pkg-devel] Registering methods conditionally

[Lluís Revilla]

Hi Ivan,

Thanks for the careful investigation. I didn't notice that tidy was masked!

I expected that a delayed registration could prevent the error. But I think
I made a bad decision when I defined tidy and I will probably change the
method name.

But I'll talk with Bioconductor people, for advice to see what can be done.
I don't think when BiocGenerics depended on generics they accounted for
this kind of breakage or informed their reverse dependencies.

Many thanks for your help.

On Sat, 8 Feb 2025 at 21:07, Ivan Krylov <ikrylov using disroot.org> wrote:

> В Sat, 8 Feb 2025 19:18:59 +0100
> Lluís Revilla <lluis.revilla using gmail.com> пишет:
>
> > Error in UseMethod("tidy") : no applicable method for 'tidy' applied
> > to an object of class "GeneSet"
>
> This seems to be a distant consequence of
> <https://github.com/Bioconductor/BiocGenerics/issues/20>.
>
> The example for BaseSet::tidy starts with
>
> >> if (requireNamespace("GSEABase", quietly = TRUE)) {
>
> which produces the following messages:
>
> Loading required package: BiocGenerics
> Loading required package: generics
> Attaching package: ‘generics’
> The following objects are masked from ‘package:BaseSet’:
>    tidy, union                    # <-- BaseSet::tidy now shadowed
>
> Indeed, at the point of the error, 'tidy' is
>
> Browse[1]> tidy
> function (x, ...)
> {
>     UseMethod("tidy")
> }
> <bytecode: 0x5615feed95c0>
> <environment: namespace:generics>
>
> ...which is different from your own 'tidy' generic:
>
> Browse[1]> BaseSet::tidy
> function (object)
> {
>     UseMethod("tidy")
> }
> <bytecode: 0x5615eed320c8>
> <environment: namespace:BaseSet>
>
> ...and the method is registered for the BaseSet::tidy generic, not the
> generics::tidy generic:
>
> Browse[1]> 'tidy.GeneSet' %in% ls(BaseSet:::.__S3MethodsTable__.)
> [1] TRUE
> Browse[1]> 'tidy.GeneSet' %in% ls(generics:::.__S3MethodsTable__.)
> [1] FALSE
>
> The problem only appears on R-devel because only the development
> version of 'BiocGenerics' has the Depends: relationship with 'generics'.
>
> This one is easier to diagnose than to cure. The example seems to work
> fine if the methods are registered for the other generic at runtime:
>
> library(BaseSet)
> # these are undocumented, use .S3method() in more normal circumstances
> registerS3method("tidy", "GeneSet", BaseSet:::tidy.GeneSet,
>                  loadNamespace("generics"))
> registerS3method("tidy", "GeneSetCollection",
>                  BaseSet:::tidy.GeneSetCollection,
>                  loadNamespace("generics"))
> example("tidy") # doesn't crash
>
> Since BaseSet depends on a new enough R version, it could in theory use
> delayed S3 method registration in its NAMESPACE file [1] in addition to
> registering a method for BaseSet::tidy, but I think that you're now
> expected to importFrom(generics, tidy) and export(tidy) to register S3
> methods on it. It's unfortunate that generics::tidy has a different
> documented purpose. Perhaps the people better-versed in Bioconductor
> have a better suggestion?
>
> Not sure where S4 method dispatch comes into play in your example, but
> conditional import of S4 classes may be hard to implement. NAMESPACE
> files use R syntax but different semantics. Testing for getRversion()
> works because that doesn't usually change for the lifetime of the
> package library. Testing for requireNamespace() wouldn't have worked
> because the results are normally cached at installation time. Maybe
> there's a way using load hooks, but we'd need an S4 wizard to implement
> that correctly. Thankfully, 'BaseSet' seems to only call setMethod()
> for its own classes, so there are no copies of foreign classes creating
> binary dependencies in its namespace.
>
> --
> Best regards,
> Ivan
>
> [1]
> https://cran.r-project.org/doc/manuals/R-exts.html#Registering-S3-methods
>

	[[alternative HTML version deleted]]