[Rd] Dependencies of Imports not attached?

[Simon Urbanek]

On May 8, 2013, at 1:52 PM, Martin Morgan wrote:

> On 05/08/2013 10:25 AM, Simon Urbanek wrote:
>> 
>> On May 7, 2013, at 11:35 PM, Peter Meilstrup wrote:
>> 
>>> Encountered an error in scripting, which can be reproduced using Rscript as
>>> follows:
>>> 
>>> $ Rscript -e "library(httr); handle('http://cran.r-project.org')"
>>> 
>>> Error in getCurlHandle(cookiefile = cookie_path, .defaults = list()) :
>>>  could not find function "getClass"
>>> Calls: handle -> getCurlHandle
>>> 
>>> or by starting R without the methods package attached:
>>> 
>>> $ R_DEFAULT_PACKAGES=base R
>>> [snip]
>>>> library(httr)
>>>> handle('http://cran.fhcrc.org/')
>>> Error in getCurlHandle(cookiefile = cookie_path, .defaults = list()) :
>>>  could not find function "getClass"
>>> 
>>> As far as I can tell the error occurs when getCurlHandle .Calls a C
>>> function which then calls SET_CLASS, which (I guess) requires
>>> methods::setClass to be in the search path.
>>> 
>>> Now 'httr' Imports 'RCurl' which Depends on 'methods'. So I think
>>> `library(httr)` should end up attaching 'methods' to the search path, but
>>> it seems 'methods' is just imported to RCurl's namespace.
>>> 
>>> I think this is a problem since the Depends line is indicating that
>>> 'methods' must be attached for RCurl to work, whether or not RCurl itself
>>> is being attached.
>>> 
>> 
>> For the record, I see the same problem with other packages that use S4 - very often it trips packages that don't use S4 but import a package that does. The analysis is correct - if a package B just imports a function from another package A which in turn relies on something in Depends, it breaks, because A is not on the search path and thus A doesn't have access to the dependencies it needs. I don't know that was the intended design. I see two way to fix this
>> 1) make sure Depends: are always put on the search path even if the package is not attached
>> 2) automatically generate imports for all packages in Depends:
>> 
>> The main problem is that B is helpless - only a change in A can make it work. Fix for A is to explicitly add import(methods) even though it's already in Depends:. Note that R CMD check doesn't detect this.
> 
> As described, this is a 'package maintainer' problem -- fix package A. Also, this isn't unique to methods, other packages routinely Depend: on something that should instead be Import:'ed.
> 

By that argument Depends: is unusable. If you specify Depends: you are expressing that your package will work if those packages are on the search path. The fact that R will happily import functions from the package despite that requirement not being fulfilled seems like a bug to me.


> But I think Peter's case is different, because the C implementations of
> 
>    SEXP R_do_MAKE_CLASS(const char *what);
>    SEXP R_getClassDef  (const char *what);
> 
> need to be called from inside the package environment, but there is no way to pass the package environment through the interface above.
> 

Yes, it's the same issue, but a case where adding imports doesn't help because they are not consulted. So in the current setup only 1) is a solution. With your patch, 2) would work as well.


> Maybe the partial patch (attached) points to a solution? It changes the interface (though would only break a couple of Bioconductor packages in a minor way) (I think there's a missing PROTECT in there too.)
> 

At least currently the PROTECT is not needed because setAttrib() protects its arguments (although there are some dangerous comments in installAttrib ;)).

Cheers,
Simon



> 
>> 
>> Cheers,
>> Simon
>> 
>> ______________________________________________
>> R-devel at r-project.org mailing list
>> https://stat.ethz.ch/mailman/listinfo/r-devel
>> 
> 
> 
> -- 
> Computational Biology / Fred Hutchinson Cancer Research Center
> 1100 Fairview Ave. N.
> PO Box 19024 Seattle, WA 98109
> 
> Location: Arnold Building M1 B861
> Phone: (206) 667-2793
> <object-env.diff>