Transitive deps

[Neil Bartlett]

On Thu, Sep 5, 2013 at 12:34 PM, Jaroslav Tulach
<jaroslav.tulach at oracle.com> wrote:
>> [...]
> As splitted packages are rare (and maybe even unsupported in some versions of
> OSGI), then a package import basically means non-transitive dependency.
>
> You may realize that your belief is a specialization of my general
> observation. As such it is of no surprise, that if non-transitive JAR
> (compilation) dependencies work sufficiently, import package works for you in
> OSGi.
>
> Import-Package seems to be a solution of the transtivity issues, but a bit too
> narrow and certainly not the only one.
> -jt


Stepping back a bit, I feel that the underlying failure in Maven is to
distinguish between build and runtime dependencies. They require
different approaches.

I think we are in broad agreement about how to manage build
dependencies. They are not required to be transitive -- to compile
each module we only need visibility of the modules it directly
imports. They should not use ranges because this results in
non-repeatable builds; for my customers it is important to be able to
reproduce a build precisely, even many years into the future. In the
bnd build system we can go beyond fixed module identity+version and
actually depend on the SHA hash of a module. Admittedly we also made
the mistake of using dependency ranges at build time but now strongly
discourage it.

Runtime dependencies are unavoidably transitive. The key is therefore
to find ways to minimise the transitive dependency graph, which is the
purpose of the OSGi Import-Package instruction, since it creates
opportunities to refactor module contents and/or find alternative
providers.

Neil

>
>
>