Control: affects -1 caffe On Sat, 2016-09-24 at 22:51 +0200, Andreas Beckmann wrote: > This is Bug#838789, but for some reason the BTS considers this as an > unknown package, so you didn't get this forwarded ...
Thank you for forwarding. :-) > Oh, you did a source-only upload. That does not work in this case, since > you have Build-Depends in non-free. You'll always have to do > source+binary uploads (for *all* architectures you want). And having a > source package without binaries for some time seems to have made the > package "partially disappear" from the archive. Several weeks ago I posted a question on mentors, asking why buildd doesn't build caffe-contrib and saying "cuda" is missing. Now I know buildd won't pull non-free stuff and source-only upload may cause "partial missing" package. > caffe-contrib FTBFS in experimental: > ../../../lib/libcaffe.so.1.0.0-rc3: undefined reference to > `H5LTread_dataset_string' I noticed this FTBTS some time ago in package caffe (another version, free of cuda). However I didn't look into the reason of build failure because at that time there was an hdf5 transition. ╰─>$ find /var/lib/dpkg/info/ -name '*.symbols' | grep hdf5 | xargs ack H5LTread_dataset_string /var/lib/dpkg/info/libhdf5-openmpi-10:amd64.symbols 3734: H5LTread_dataset_string@HDF5_MPI_1.8.7 1.8.13 /var/lib/dpkg/info/libhdf5-10:amd64.symbols 3707: H5LTread_dataset_string@HDF5_SERIAL_1.8.7 1.8.13 ╰─>$ apt list libhdf5-10 libhdf5-openmpi-10 Listing... Done libhdf5-10/unstable,now 1.8.16+docs-8+b1 amd64 [installed,automatic] libhdf5-openmpi-10/unstable,now 1.8.16+docs-8+b1 amd64 [installed] Well, confusing. > I'm *not* attaching the 160 MB build log. But I could look for details > if needed. > > In case it does matter: > This was built against hdf5 1.8.16 from sid, not against 1.10 from > experimental. I used to test it against hdf5/sid. caffe* were uploaded to experimental because runtime-dep python3-protobuf is still missing. > IIRC some time ago apt changed how Build-Depends are resolved for > packages in experimental, preferring packages from unstable unless > a versioned constraint actually requires something from experimental. But this reminds to to test whether it compiles against hdf5/experimental. > I remember vaguely to have seen symbol changes during previous rebuilds, > but I don't have the old longs any longer. That's ok, I'll handle the changes if any, as long as I can pass the build. > > Andreas