[R-SIG-Mac] packages, R-patched and RC [Was: Importing Excel files]

[Prof Brian Ripley]

On 20/06/2012 11:20, peter dalgaard wrote:
>
> On Jun 19, 2012, at 23:29 , Prof Brian Ripley wrote:
>
>> On 19/06/2012 17:35, Simon Urbanek wrote:
>>> On Jun 19, 2012, at 5:36 AM, peter dalgaard wrote:
>>>
>>>>
>>>> On Jun 19, 2012, at 01:16 , Colstat wrote:
>>>>
>>>>> I think the error says it "package Œgdata‚ was built under R
>>>>> version 2.15.1" and you have R 2.5.10.
>>>>>
>>>>> Update your R first, let me know if it doesn't work.
>>>>
>>>> 2.15.1 is announced for Friday... I think it's a bit of a
>>>> glitch that CRAN is already automatically providing packages
>>>> for it, but you are of course more than welcome to test the
>>>> prereleases (from http://R.research.att.com/).
>>>>
>>>> I suppose that this comes about from building packages with
>>>> R-patched, which transitions directly into the prereleases for
>>>> the next version.
>>>>
>>>
>>> Yes, this is a side-effect of that. I wonder what we can do about
>>> it - I used to build packages with released versions only, but
>>> then people complained that patched had fixes for some things
>>> they needed...
>>
>> We could simply not update the version of R-patched used during the
>> beta/RC periods: they are after all only about 10 days.
>
> Yes, that would handle the formal issue of the version number
> conflict.
>
> However, there's a deeper issue which got highlighted by the
> identical() debacle in April-June: It seems that we can't guarantee
> ABI compatibility throughout the R-branch series (e.g., everything
> labeled 2.15.x). We try to ensure that the modified versions of R
> will run packages built with a previous version, but the other way
> around might not work. If that sort of thing happens in R-patched,
> midway between releases, then users could find themselves blocked
> from installing packages until they update to R-patched, which could
> be undesirable in corporate or educational settings.

Note this only applies to binary packages: people can always install 
from the sources (trivially for 'gdata' since it has no compiled code).

My understanding is that the Windows' policy is to stick with the 
current 2.15.x release: 2.15.0 is currently used but once 2.15.1 is out 
packages will be built with 2.15.1 (and re-built if their dependencies 
change).

The 'debacle' was Bioconductor's lack of understanding of the 
assumptions of their own distribution model.  As we said at the time, if 
you want to provide binary packages that work under all of 2.15.x, you 
need to prepare them with 2.15.0.  But as Simon points out, that may 
stop binary packages being provided for some recent source packages.

Note that there are other issues here: as we have seen in the last week, 
updates and downdates to a recommended package such as Matrix have 
required other packages to be re-installed. It is not just the version 
of R that is relevant but also what other packages were installed at the 
time the binary package was prepared.

We may need to work harder to get people to understand that binary 
packages have 'use at your own risk' status.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595