[Bioc-devel] Pushing towards a better home for matrix generics

[McDavid, Andrew]

As a casual observer of this thread, my takeaway is that 1. the current situation is untenable (e.g. it means that Aaron would have to essentially reimplement S4 method dispatch) 2.  Given the long history, and extent of reverse dependencies to Matrix, there are good reasons to not mess with the signature of its implicit generic (though I don't see much utility in multiple dispatch here) and 3.  therefore the least-bad alternative may be to eliminate the call to setGeneric('colSums'), etc, in BiocGenerics, hopefully with some fixes to the help system to make it more tolerant to S4 method naming.  I appreciate that until these fixes are forthcoming it means more work maintaining the help aliases for some methods.  How often do we think the aliases would be break?

Andrew McDavid
Biostatistics and Computational Biology
Office: SRB 4.206 Ph: 585.275.5983

Message: 1
Date: Sun, 10 Feb 2019 13:36:43 +0000
From: Aaron Lun <infinite.monkeys.with.keyboards using gmail.com<mailto:infinite.monkeys.with.keyboards using gmail.com>>
To: "Pages, Herve" <hpages using fredhutch.org<mailto:hpages using fredhutch.org>>, Martin Maechler
<maechler using stat.math.ethz.ch<mailto:maechler using stat.math.ethz.ch>>
Cc: Michael Lawrence <lawrence.michael using gene.com<mailto:lawrence.michael using gene.com>>,
"bioc-devel using r-project.org<mailto:bioc-devel using r-project.org>" <bioc-devel using r-project.org<mailto:bioc-devel using r-project.org>>
Subject: Re: [Bioc-devel] Pushing towards a better home for matrix
generics
Message-ID: <1549805803.3935.11.camel using gmail.com<mailto:1549805803.3935.11.camel using gmail.com>>
Content-Type: text/plain; charset="utf-8"

Returning to this topic:

It's good to hear some of the rationale behind the current state of
affairs. That said, the set-up we have now is quite difficult to work
with; as mentioned before, I've had to hack around it like:

# Example from "BiocSingular", https://urldefense.proofpoint.com/v2/url?u=https-3A__github.com_LTLA_BiocSingular&d=DwIGaQ&c=4sF48jRmVAe_CH-k9mXYXEGfSnM3bY53YSKuLUQRxhA&r=YOujD4cGNGWN0ZskTj96dUHU2a4CkraklBd4BndS21sf3_emyYLIG1llnUarWCda&m=VBZgu2tHS9vBNQ2d8RSG4ijIcp3jyZkfCsRnA1mk7XI&s=FN4WdWx8gfnjFoLRQuBaIHaKrAjONVa9hsbAyIXSwGo&e=
.safe_colSums <- function(x) {
    if (is(x, "Matrix")) {
        Matrix::colSums(x)
    } else {
        colSums(x)
    }
}

... which is ugly, and even worse, still incorrect, e.g., for non-
Matrix classes that have methods for the implicit colSums generic. This
situation is not sustainable for further package development.

Is there a path forward that is palatable to everyone? Or perhaps these
conversations are already happening on R-devel?

-A

On Tue, 2019-01-29 at 18:46 +0000, Pages, Herve wrote:
Yes the help system could enforce the full signature for the aliases
but
that means the end user then will have to always do
?`colSums,SomeClass,ANY,ANY-method`, which feels unnecessary
complicated
and confusing in the case of a generic where dispatching on the 2nd
and
3rd arguments hardly makes sense.

Or are you saying that the help system should enforce an alias that
strictly matches the signature explicitly used in the setMethod
statement? Problem with this is that then there is no easy way for
the
end user to know a priori which form to use to access the man page.
Is
it ?`colSums,dgCMatrix,ANY,ANY-method` or is it
?`colSums,dgCMatrix-method`. Right now when you type colSums<ENTERN>
(after loading the Matrix package), you get this:

   > library(Matrix)
   > colSums
   standardGeneric for "colSums" defined from package "base"

   function (x, na.rm = FALSE, dims = 1, ...)
standardGeneric("colSums")
   <bytecode:0x591c7d0>
   <environment:0x591a408>
   Methods may be defined for arguments: x, na.rm, dims
   Use  showMethods("colSums")  for currently available ones.

This suggests that the correct form is ?`colSums,dgCMatrix,ANY,ANY-
method`.

All this confusion can be avoided by specifying signature="x" in the
definition of the implicit generic. It formalizes where dispatch
really
happens and sets expectations upfront. No loose ends.

Hope this makes sense,

H.


	[[alternative HTML version deleted]]