I use GSL in some radar signal processing code. I use the array containers and routines for linear algebra, zero finding, optimization, and B-splines. The B-splines are only practical for my problems if I use a sparse solver; currently I rely on csparse to solve the B-spline fits, so it'd be great to have that functionality in GSL.
I second the desire for C99 complex types; it's much more convenient to write arithmetic than a bunch of function calls or type casts. Making the array containers more flexible might be worthwhile, too. Numpy allows arbitrary byte strides on each axis, which lets you do cool things like get a vector view of structured data (e.g., the real part of a complex vector). However, I understand there might be performance concerns. I much prefer regular releases to relying on a git tree of unknown maturity. Also, even though GSL is pretty easy to compile, it's still simpler to use a package manager, especially when deploying virtual machines.
