[Rd] Outdated dependencies and install.packages()

[Martin Maechler]

>>>>> "PD" == Peter Dalgaard <P.Dalgaard at biostat.ku.dk>
>>>>>     on Wed, 20 Aug 2008 17:06:36 +0200 writes:

    PD> Prof Brian Ripley wrote:
    >> On Wed, 20 Aug 2008, Peter Dalgaard wrote:
    >> 
    >>> This looks like a buglet:
    >>> 
    >>>> install.packages("lme4", depend=TRUE, lib="~/Rlibrary/")
    >>> --- Please select a CRAN mirror for use in this session ---
    >>> [...snip...]
    >>> ** preparing package for lazy loading
    >>> Error: package 'Matrix' 0.999375-4 was found, but >= 0.999375.11 is
    >>> required by 'lme4'
    >>> Execution halted
    >>> ERROR: lazy loading failed for package 'lme4'
    >>> ** Removing '/home/bs/pd/Rlibrary/lme4'
    >>> ** Restoring previous '/home/bs/pd/Rlibrary/lme4'
    >>> 
    >>> and the DESCRIPTION file for lme 4 does have
    >>> 
    >>> Depends: methods, R(≥ 2.7.0), Matrix(≥ 0.999375-11), lattice
    >>> <http://cran.r-project.org/web/packages/lattice/index.html>
    >>> 
    >>> 
    >>> Looks like we're either not checking the version requirement of
    >>> dependencies, or something is making 4 > 11... (from my reading of the
    >>> code, it is the former, but I could have missed something).
    >> 
    >> We don't check version requirements of dependencies in
    >> install.packages(). That's a minefield related to having multiple
    >> versions of a package on a system and which gets loaded/namespace
    >> loaded, so you can easily have the needed version but pick up a
    >> different one.
    >> 
    >> Depending on whether lazy-loading is in place (and various other
    >> issues) you may or may not need the dependencies to install the
    >> package.  It's also a problem that you can update dependencies to
    >> incompatible ones, so the versions available at installation are only
    >> part of the story.
    >> 
    >> I don't see why the message reports .11 not -11, but that may be
    >> related to "package_version" classes.
    >> 
    >>> We should at least document the behaviour.
    >> 
    >> I am not sure where you think it should be documented and is not.
    >> But for install.packages():
    >> 
    >> dependencies: logical indicating to also install uninstalled packages
    >> on which these packages depend/suggest/import (and so on
    >> recursively).  Not used if 'repos = NULL'. Can also be a
    >> character vector, a subset of 'c("Depends", "Imports",
    >> "Suggests")'.
    >> 
    >> 'uninstalled' seems pretty clear to me.
    >> 
    PD> It isn't. Matrix 0.999375-11 was 'uninstalled'. The above could just
    PD> mean that if you installed it, it wouldn't be installed again.

    PD> Something like "Notice that version dependencies are not checked." would
    PD> help. (Assuming that people read the help page.)

Or even more:

"
Notice that version dependencies are not checked such that
 outdated versions of installed packages are \emph{not} updated.
 Consider using  \code{\link{update.packages}()} before 
 \code{install.packages()}.
"

Note that --- thanks to Kate Mullen's nice help.request() proposal ---
one of my top "TODOs" for R-devel is
to finish a generalization of the current
  update.packages()
function (that I have not quite finished)
which will allow to only look for updates of specified set of
packages rather than all packages in a given lib.loc
{something that should *not* be triggered by a help.request()
 for a user (as ours) who has ~ 2000 R packages installed}.

Martin