R Package {pax}

Yves Deville deville.yves@alpestat.com

Goal

{pax} eases the management of a suite of R packages by analysing possible sources of conflicts. Examples of conflicts are

• Several packages exporting an object with the same name. Although this can most of time be solved by using ::, this may be avoided.

• S3 generic functions defined several times across the suite. This can make the packages impossible to use simultaneously. Assume that a generic say fun is defined in packages pkg1 and pkg2 which are attached in this order. When pkg2 is attached, the generic function is re-defined masking the existing function. This re-initialises the register of methods for the generic hence the methods defined in pkg1 can no longer be used.

Analysing a suite: example

Let us call package suite a list of “packages” that are likely to be used in a same session. {pax} analyses the DESCRIPTION and the NAMESPACE files to detect some possible problems. The requirement is for now that a “package” is a directory containing these two files. So this can be: a source package or an installed package in one of the directory known in .libPaths()

library(pax)

## Le chargement a nécessité le package : data.table

## Use (assumedly) installed packages 
pkgs <- c("evd", "ismev", "extRemes", "mev", "POT")
L <- pkgList(pkgs)

## package:  evd   Using `.libPaths()[1]`
## package:  ismev   Using `.libPaths()[1]`
## package:  extRemes   Using `.libPaths()[1]`
## package:  mev   Using `.libPaths()[1]`
## package:  POT   Using `.libPaths()[1]`

res <- newS3generics(L$NS)

## Warning in newS3generics(L$NS): package 'ismev' has no exports. The analysis
## may be misleading

## Warning in newS3generics(L$NS): package 'extRemes' has no exports. The analysis
## may be misleading

## Some new generics are defined in more than one package
##       Multi Generic      Pkg     
##  [1,]       blockmaxxer  extRemes
##  [2,]       bvcpp        evd     
##  [3,]       bvdens       evd     
##  [4,]       bvdp         evd     
##  [5,]       bvh          evd     
##  [6,]       bvqc         evd     
##  [7,]       ci           extRemes
##  [8,]       convassess   POT     
##  [9,]       datagrabber  extRemes
## [10,]       decluster    extRemes
## [11,] X     dens         evd, POT
## [12,]       distill      extRemes
## [13,]       findpars     extRemes
## [14,]       fpois        extRemes
## [15,]       initializer  extRemes
## [16,]       pextRemes    extRemes
## [17,]       postmode     extRemes
## [18,] X     pp           evd, POT
## [19,]       profile2d    evd     
## [20,] X     qq           evd, POT
## [21,]       retlev       POT     
## [22,]       return.level extRemes
## [23,]       rextRemes    extRemes
## [24,]       rl           evd     
## [25,]       setup.design extRemes
## [26,]       std.errors   evd     
## [27,]       strip        extRemes
## [28,]       trans        extRemes

So we see that some generics such as pp are defined several times

library(evd)
uvdata <- rgev(100, loc = 0.13, scale = 1.1, shape = 0.2)

fit <- fgev(uvdata)
pp(fit)

head(getS3method("pp", class = "gev"))

##                                                                           
## 1 function (x, ci = TRUE, cilwd = 1, a = 0, main = "Probability Plot",    
## 2     xlab = "Empirical", ylab = "Model", ...)                            
## 3 {                                                                       
## 4     ppx <- ppoints(x$n, a = a)                                          
## 5     probs <- pgev(sort(x$tdata), loc = x$loc, scale = x$param["scale"], 
## 6         shape = x$param["shape"])

library(POT)

## Registered S3 methods overwritten by 'POT':
##   method      from
##   print.bvpot evd 
##   plot.bvpot  evd

## 
## Attachement du package : 'POT'

## Les objets suivants sont masqués depuis 'package:evd':
## 
##     dens, dgpd, exiplot, mrlplot, pgpd, pp, qgpd, qq, rgpd, tcplot

pp(fit)

## 
## Call: fgev(x = uvdata) 
## Deviance: 351.2626 
## 
## Estimates
##    loc   scale   shape  
## 0.1363  1.0238  0.2683  
## 
## Standard Errors
##     loc    scale    shape  
## 0.11704  0.09693  0.08897  
## 
## Optimization Information
##   Convergence: successful 
##   Function Evaluations: 40 
##   Gradient Evaluations: 13

try(getS3method("pp", class = "gev"))

## Error in getS3method("pp", class = "gev") : 
##   Méthode S3 'pp.gev' introuvable

The problem is that when a generic is re-defined, the corresponding S3 methods are lost because the register of S3 methods is cleared or re-initialised. So once {POT} has been attached, the `pp S3 method no longer exists for the class "gev" and pp(fit) calls the default method as implemented by the pp.default function.

POT:::pp.default

## function (object, ...) 
## return(object)
## <bytecode: 0x5c2506765450>
## <environment: namespace:POT>

We see that this method simply returns a copy of the object.

So, it is not possible to use simultaneously all the functionalities of the two packages. One possibility to avoid this problem would have been to make {POT}* import the function pp from {evd}. Anyway, new S3 generic functions should only defined with parsimony and care.