Re: [igraph] Version 0.7?

[Gábor Csárdi]

Hi Diego, thanks for the answer.

On Tue, Jan 28, 2014 at 9:36 PM, Diego Diez <address@hidden> wrote:

Hi Gabor,

[...]

 

If I understand correctly from your email, the problem is that many
packages depend on igraph, and so pushing changes to CRAN that disrupt
the API may make a lot of packages to fail to build, potentially
creating a lot of disruption.

The main problem is not that I am breaking other packages, usually I am not. (The 0 -> 1 indexing case was an exception.) The problem is that I need to check 150 packages. Most of these are easy, but some (~30) are not and I spent my whole day with this. This is a burden, and makes me delay igraph releases all the time. This is not productive at all. I would rather work on igraph.

Now there will be some changes again in igraph before the release and I'll need to do this again. I need to install all these packages, together with their dependencies, altogether about 700 packages, more than 5 GBytes. Many of them require system libraries, like MPI, GTK, ggobi, etc.

I think the current CRAN organization is unsustainable, and makes maintainers with popular packages work a lot. This should be avoided, and my problem is that I don't see any improvement or developments towards this.

My opinion is that yes, usually changing the API or making other big
changes will cause problems. A potential solution is to add a
dependency on a particular igraph version (e.g. igraph <= 0.6). But
this will not solve the problem if the user has other packages that
were updated to use igraph 0.7,  as it is not straightforward to
have/maintain both versions of the package in a single installation
(AFAIK).

Practically impossible, I would say.

This problem will still exists if you remove igraph from CRAN, as
different maintainers will adopt igraph 0.7 or future versions at
different paces. It will only make slightly more painful to install
packages that depend on future versions of igraph.

Yes, that is exactly the problem. I am thinking about working around this, e.g. by having an igraph_installer package on CRAN, that would be able to install and load multiple versions of igraph. This way people could depend on exact versions. But I still need to work this out fully, in a way that it potentially acceptable for the CRAN maintainers, and convenient for people who use igraph.

 

A better solution is actually what you did when igraph 0.6 came out
and changed the index system in R. You released igraph0, which was
actually 0.5, to support legacy code. This required a small change in
the dependencies that most people were probably willing to do at any
time.

I am glad that this worked for you, this was exactly my motivation for doing it. But this is now not allowed on CRAN, because Prof Ripley didn't like it. Don't ask me why. He actually went after people, and made them upgrade to igraph from igraph0, and then deleted igraph0 from CRAN, even if I wanted to keep it there.

 

[...]

The main problem with CRAN is that, unlike Bioconductor, there is no
such a concept of "CRAN releases". In CRAN the latest version of a
package is available irrespective of whether that breaks all the
dependent packages. In Bioc, each release ensures that a particular
version of a package works with a particular version of a dependency.
Of course, this only applies to packages in the Bioconductor
environment.

Exactly. One solution would be having releases. As R also has scheduled releases, this would be rather easy and logical. But it does not seem that it will ever happen. Another solution would be exact dependencies and the ability to install and load multiple versions of the same package (at the same time, obviously). Jerome Ooms has a nice writeup in the R journal about this: http://journal.r-project.org/archive/2013-1/ooms.pdf

[...]

In summary, I do not think moving igraph from CRAN will solve the
problem. If the new code disrupts a considerable amount of packages
then using the igraph0 approach may be a much useful way to make the
transition smoothly for package maintainers.

Actually, you gave me a good idea by mentioning BioConductor. I could just move it to BioConductor from CRAN, that could work. There are some things I don't like in BioConductor, though, e.g. the way they do versioning instead of you, etc.

Alternatively, I could work out the igraph_installer package idea, if it is possible at all.

Best,

Gabor