On Mon, Jul 25, 2016 at 12:05 PM, Andreas Lobinger <lobing...@gmail.com> wrote:
> Hello colleagues, > > On Monday, July 25, 2016 at 4:40:10 PM UTC+2, Stefan Karpinski wrote: >> >> The difference between the Julia package ecosystem and DLL hell is that >> DLLs expose static interfaces and cannot adapt to their environment. If >> libA.dll expects on interface from libC.dll and libB.dll expects another, >> you're just stuck – you can use libA.dll or libB.dll but not both (unless >> you do some crazy stuff with loading multiple copies of libC.dll). Julia >> packages are quite a bit more flexible: upon loading, A.jl can introspect >> the version and interface of C.jl and adapt to it, so that it works with a >> newer or older version correctly. If you want to be able to use A and B at >> the same time with a newer versions of C in the scenario you describe, the >> solution is to update A so that it works with the newer version of C, and >> ideally continues to work the the older version as well. >> > > Can you give us a good example of situation A.jl and C.jl you described? > Compat.jl does this extensively with respect to Julia itself. While i value Stefan's optimism > >> and ideally continues to work > > and the experimental/scientific approach > >> A.jl can introspect the version and interface of C.jl and adapt to it, > > still i could raise the question: And what if not? > The "ideally" applies to developer actions, not whether this is possible or not – it is. > And why is it so easy to publish a package creating dependency problems? > Why are dependency problems not detected by automatic testing? And what is > the recommended way if a package author isn't available? etc. > Dependency problems don't seem particularly common and they seem to be resolved quickly when they come up. Automatic testing is the ideal way to handle this, but someone has to build, maintain and host that testing infrastructure. The problem of automatically detecting version interactions is exponential in the number of packages and (and polynomial in the number of versions with a very high degree), so some significant cleverness would need to be applied in order to make this computationally practical.
