On 21 September 2015 at 13:06, Simon McVittie wrote: | On 21/09/15 12:51, Dirk Eddelbuettel wrote: | > | Yes, you should rename to libquantlib0v5 and ask the release team to | > | > I'd prefer libquantlib1. Any reason not to? | | Changing the SONAME for this transition would be misleading unless the | SONAME is already Debian-specific (in which case it would be | libfoo.so.0debian1 or something), because the ABI depends on whether it | was compiled with g++-4 or g++-5, and not on anything about the actual | source code. | | Your current source code compiled with g++-5 has the old SONAME but the | new ABI; source code with a new SONAME compiled with g++-4 would have | the new SONAME but the old ABI. | | The solution is to change something that is part of the Debian | packaging: the package name.
So is the trailing '1'. All that matters is a string comparison making it different from libquantlib0. As you're subtly hinting that I should rather use libquantlib0v5 I may do so. It is just 'uglier' but common ... | > | binNMU the bindings against it, unless you have another way to ensure | > | > I fail to understand. The two packages both have sufficient Build-Depends as | > eg here for quantlib-python | > | > Build-Depends: debhelper (>= 7.0.0), python, python-all-dev (>= 2.6.6-3), \ | > libquantlib0-dev (>= 1.6.2), gcc (>= 4:5.2), g++ (>= 4:5.2), \ | > ^^^^^^^^^^^^^^^^^^^^^^^^^^^ | > libboost-dev (>= 1.34.0), libboost-test-dev (>= 1.34.0) | > | > resulting in corresponding versioned Depends. | | That is *normally* sufficient to do the right thing, but is not | sufficient to prevent breakage in this ABI transition, because the | libstdc++ change makes newer versions incompatible with older versions | (if it didn't, we wouldn't have needed hundreds of transitions involving | thousands of packages). | | Specifically, the versions of the bindings currently in jessie depend on | a version that is satisfied by the version of quantlib in stretch - but | the quantlib binary in stretch is not compatible with those binding | binaries, because compilation with a newer g++ caused it to break its ABI. I may get myself confused by the names (I find '(old)stable', 'testing' and 'unstable' simpler to use). And both testing and unstable have the correct pair: quantlin 1.6.2-1 and quantlib-swig 1.6.1-1 (upstream needed a QL only fix, hence 1.6.2 there). Neither one will make it to (old)?stable. Only forcing by the user would. And then they get what they asked for... | > | that broken combinations of packages cannot be installed (including | > | during the upgrade from jessie-as-oldstable to stretch-as-stable after | > | stretch is released). | > | > I don't see how using the package manager. | | If you do the v5 rename as suggested, it prevents the installation of a | broken combination of packages; libquantlib0v5 should conflict/replace Yes, 'Conflicts:' and 'Replaces:'. Or 'Breaks:' and 'Replaces:' ? I have an old Breaks from a decade ago: Breaks: libquantlib-1.1, libquantlib-1.0.0 Replaces: libquantlib-1.2, libquantlib-1.1, libquantlib-1.0.0 Provides: libquantlib-1.2, libquantlib-1.1, libquantlib-1.0.0 | libquantlib0 (because they conflict over the path libquantlib.so.0), old | bindings depend on libquantlib0, and new bindings depend on | libquantlib0v5. The combination of those forces quantlib and all the | installed bindings to be from the same "world". Right. | The wider transition of which this is a part has been going on for | nearly 2 months, and is quite well-understood. I'm sorry it took so long | for the mass-bug-filing to reach the last few packages with an ABI break | and known rdeps (such as quantlib), but everyone involved is getting | quite burned-out at this point. Understood. Dirk -- http://dirk.eddelbuettel.com | @eddelbuettel | e...@debian.org
