[R] make methods work in lapply - remove lapply's environment

[Duncan Murdoch]

On 08/09/2008 9:37 PM, Tim Hesterberg wrote:
> I've defined my own version of summary.default,
> that gives a better summary for highly skewed vectors.
> 
> If I call
>   summary(x)
> the method is used.
> 
> If I call
>   summary(data.frame(x))
> the method is not used.
> 
> I've traced this to lapply; this uses the new method:
>   lapply(list(x), function(x) summary(x))
> and this does not:
>   lapply(list(x), summary)
> 
> If I make a copy of lapply, WITHOUT the environment,
> then the method is used.
> 
> lapply <- function (X, FUN, ...) {
>     FUN <- match.fun(FUN)
>     if (!is.vector(X) || is.object(X))
>         X <- as.list(X)
>     .Internal(lapply(X, FUN))
> }
> 
> I'm curious to hear reactions to this.
> There is a March 2006 thread
>     object size vs. file size
> in which Duncan Murdoch wrote:
>> Functions in R consist of 3 parts: the formals, the body, and the
>> environment. You can't remove any part, but you can change it.
> That is exactly what I want to do, remove the environment, so that
> when I define a better version of some function that the better
> version is used.

But that's not removing the environment, that's changing it.  Your 
function has globalenv() as its environment.

> 
> Here's a function to automate the process:
> copyFunction <- function(Name){
>   # Copy a function, without its environment.
>   # Name should be quoted
>   # Return the copy
>   file <- tempfile()
>   on.exit(unlink(file))
>   dput(get(Name), file = file)
>   f <- source(file)$value
>   f
> }
> lapply <- copyFunction("lapply")
> 

A shorter version is

copyFunction <- function(fn){
   environment(fn) <- globalenv()
   fn
}

(which doesn't require quoting the function name).

But getting back to your original question:  the real problem is with S3 
method dispatch and its interaction with lapply.  In the bad case, 
lapply calls the generic, which calls the dataframe method, which calls 
methods (probably via the generic again, but I didn't look) for each of 
the columns.  This is an ambiguous case:  should the dataframe method 
act the way its author expected, and call the standard method, or should 
it follow the search order you want?

I'd tend to think authors of functions should be able to depend on their 
behaviour not changing based on what's happening in the global 
environment.  That's not an absolute rule:  you should be able to define 
a new class and a method for it and have things work, but I don't think 
the fact that you have a summary.default should affect functions in 
namespaces that were written for the standard summary.default.

Duncan Murdoch