[Zope-dev] setup.py "extra" dependencies

[Gary Poster]

On Mar 5, 2009, at 1:43 PM, Martijn Faassen wrote:

> Hi there,
>
> I know opinions are divergent about 'extra' dependencies in setup.py.
> These ar dependencies that effectively make a single project with a
> single dependency structure into a number of "virtual" packages that
> each can have a separate list of dependencies. Such a virtual package
> that we're currently getting rid of is zope.component[zcml].

...

> Opinions?

I disagree with the blanket statement.

I do lean towards not having the extras for the test package only.   
I'm fine with the policy "If you want zope.testing for your tests,  
then keep it as a dependency for the package".

But I like to have the option of extras for testing additional setups.

zc.async uses extras so that the main tests show the functionality as  
a Python library; another level adds more Zope dependencies, with  
associated tests; and the last level adds the most.  I could have  
divided these up into multiple teensy-weensy packages but I didn't  
really want to.  It seemed like overkill.

I've also done this in other circumstances in which code could use  
zope.proxy/zope.security, but I really didn't want to add the hard  
dependency.  The tests to show that proxies were handled correctly  
were only part of the "extras".  Dividing this package also would have  
made no sense--it was already just a few small classes.

For a package as central as zope.component, I think the pattern Tres  
is pursuing--dividing everything up--makes sense.

For most other packages, I personally feel that there are  
circumstances in which extras are a useful tool.

I do wonder if there's a ``setup.py test`` spelling for testing these  
extras though.  If there were not, I'd find that a good argument  
against the "extras" pattern, at least for core Zope packages.

Gary