[SciPy-dev] issues with scipy and eggs

[Andrew Straw]

Ravikiran Rajagopal wrote:

>On Monday 12 December 2005 15:53, Fernando Perez wrote:
>  
>
>>Robert Kern wrote:
>>    
>>
>>>Well, there's the problem with the __init__.py's from full scipy
>>>overwriting scipy_core's during a regular install. As Andrew suggested,
>>>we can simply have synced copies in both packages. To do that *right*, we
>>>need some SVN magic (svn:external?).
>>>      
>>>
>>Mmh, I hadn't noticed this because I've been building from source so far.
>> Is this going to be problematic for package managers?  From what I've
>>seen, rpm does not like packages which overwrite files in other packages
>>(though there may be directives to tweak that), I don't know what apt-get
>>does in this case.
>>    
>>
>
>  
>
I agree that the issue of package managers not wanting to overwrite 
files seems a serious and understandable one, and I agree this is best 
avoided. Hopefully we can find an alternative.

>Cheapo solution: the proverbial extra level of indirection. Make __init__.py 
>only in scipy_core which checks for existence of scipy_full_init.py at run 
>time and sources it if necessary. As I understand it, we need scipy_core to 
>install scipy_full; so we can remove __init__.py from scipy_full.
>  
>
Unfortunately, I don't think this is a solution -- I think it's 
essentially what we have now.  Basically, we have two conflicting cases:

1) A non-setuptools install where site-packages/scipy/__init__.py is 
installed by scipy_core but is smart enough to figure out full scipy is 
installed
2) A setuptools install where site-packages/scipy_core.egg and scipy.egg 
exist and both need a scipy/__init__.py file to tell pkg_resources to 
keep looking for the other scipy package.

I don't know enough about the setuptools innards to know if approach 2 
is the only way to go. (Heck, I was just blindly following Robert's lead 
to get as far as I did in this...) Robert (or anyone): what if we used 
"declare_namespace" in full scipy's setup() call -- would that perform 
the necessary setuptools magic without needing an __init__.py file that 
would overwrite scipy_core's?

Perhaps a question to the distutils list (the de facto setuptools list) 
will help resolve the issue. Perhaps the issue is even worth a feature 
request in setuptools if one is needed to support our use cases.