numpy is required even for setup

[jbarnoud]

MDAnalysis cannot be installed with pip if numpy is not installed.

This means that pip install MDAnalysis does not work if numpy is not installed.

This also makes it impossible to install a virtualenv with MDAnalysis uning only a requirement.txt file. To do so, numpy must already be installed when pip is called with the requirement file, or MDAnalysis must be installed after pip is called with the requirement file.

This is inconvenient when writing a program that depends on MDAnalysis. It makes setting up Travis CI, read the docs, or similar services more difficult to setup.

Ideally, MDAnalysis should tell pip it requires numpy and not crash. This should be possible as other packages such as scipy do it.

[orbeckst]

Would be nice to have.

Just fyi: We gave up on RTD --- it seems impossible to have C-extensions (which contain docs) to get to build there. There are a bunch of issues on this in the tracker.

[kain88-de]

I'm sorry I don't see an easy way for this to work. Also I couldn't find any project besides scipy that is able to do that. And even with scipy I'm sure they con only pull that off right now because they are shipping a wheel (pre build binary).

The thing is our configuration that is run before setup(...) just needs to know the numpy include directories. There is no easy way around that.

[tylerjereddy]

@orbeckst Although note that if we start using conda then RTD may become a possibility pending resolution of this popular issue: readthedocs/readthedocs.org#857

[kain88-de]

I just tried to install scipy from source in a virtualenv without numpy. That fails because scipy can't import the appropriate numpy module to configure the build environment. This means numpy has to be installed if MDAnalysis is installed from source.

But scipy can be installed via pip without having numpy installed because they uploaded wheels to PyPi instead of eggs. So the only chance I see to fix this is converting our PyPi-package to a wheel.
But I suggest we wait with that until we refactored all C-extensions to be Py3 compatible. Then they should all hopefully be cleaned up making the process of building wheels easier.

Solving #368 here might actually help us for the process of distributing wheels if we use travis to build and upload them to PyPi

[orbeckst]

It seems there really isn't any sensible way to solve this issue at this moment so I am going to close it; we can revive it at a later date.

[mnmelo]

So, I dug a bit around and found that setup.py only uses numpy to find its include path.
By subclassing setuptools.Extension I could make a smart class that only really attempts to read numpy´s include dir at install time.
This seems to fix our problems but I'd appreciate it if someone else would test it in their virtualenvs. (PR #498)

On my tests the only issue was that biopython seems to suffer the same problem as we do, and defaults to a crippled install if numpy isn't found. Namely, as per their setup.py, the Bio.Affy, Bio.Cluster, Bio.KDTree, Bio.Motif, and Bio.PDB.QCPSuperimposer packages/extensions become unavailable. Even if numpy is later installed, biopython must be reinstalled/reconfigured to recognize it.

Do we use any of these functionalities?

[orbeckst]

Am Oct 19, 2015 um 17:58 schrieb mnmelo notifications@github.com:

Do we use any of these functionalities?

Yes, we absolutely rely on Bio.KDTree.

[mnmelo]

Well, then I might as well forget about it... My workaround doesn't seem to be working anyway!