Re: [IMP-dev] Policy for versions of IMP dependencies
To: , List for IMP development <>
Subject: Re: [IMP-dev] Policy for versions of IMP dependencies
From: Daniel Russel <>
Date: Tue, 31 Jul 2012 11:17:09 -0700
Reply-to: List for IMP development <>
It is really not clear to me how much backwards support is worth it.
For mac os, Apple doesn't do patches for versions older than the current -1, I believe. So no one should be running 10.6 at this point as they can't get security updates. So supporting that doesn't seem worthwhile.
For linux, everywhere I have been it is either upgrade within 6 months or so of CentOS/RHEL/Ubuntu being upgraded or upgrade to every other version of Fedora. So again, its not clear that we benefit anyone by supporting older versions here either.
For windows, I don't think anyone else will manage to build IMP :-) (I failed twice), so supporting old compilers there doesn't buy us much either.
So while it seems nice in theory, I don't see that there is much benefit in practice to go far back. Has someone been at an institution where older than the above was used?
On Jul 30, 2012, at 5:16 PM, Barak Raveh <">> wrote:
Good point (and Daniel said something similar in different words I think).
So perhaps as a policy, we can say: "we give XX (2-3) years backward compatibility, but for rare and true necessities (e.g., python multiprocessing), you must upgrade your dependencies in order to use IMP since it's too important and helpful ; And if possible and not too complicated, we will strive to provide partial functionality even without such upgrade (e.g., you will have IMP but without python multiprocessing)."
I had 2-3 years in mind :) quite an arbitrary figure though.
Right, this is how I chose the most recent versions of Boost to support originally. But it makes sense to agree on an "XX" as you suggest. I think 2 years is reasonable.
It's just that flawed backward compatibility is usually not due to
amazing technological breakthroughs we cannot live with out, but
probably due to some package changing the name of function X to function
Y, or a few #include statements that need to be altered...
True, I think we can live without some fancy CXX11 features. More annoying is the lack of some Boost classes and Python modules (only very very recent versions of Python ship with the multiprocessing module, for example).