On 1/26/13 1:53 PM, Tom Lane wrote:
> [ pokes around... ] Hm, it appears that that does work on Linux,
> because for some reason we're specifying RTLD_GLOBAL to dlopen().
> TBH that seems like a truly horrid idea that we should reconsider.
> Aside from the danger of unexpected symbol collisions between
> independent loadable modules, I seriously doubt that it works like
> that on every platform we support --- so I'd be very strongly against
> accepting any code that depends on this working.
Well, that would kill a lot of potentially useful features, including
the transforms feature I've been working on and any kind of hook or
debugger or profiler on an existing module. (How do plpgsql plugins
work?) We also couldn't transparently move functionality out of the
postgres binary into a module.
I see the concern about symbol collisions. But you can normally work
around that by prefixing exported symbols.