convolve_fft defaults and documentation

[astromark]

I've just been trying to run some code that I haven't used for a year or two, when I had an older version of astropy installed (I'm now running 2.0.10). It was giving very strange results and I eventually managed to track this done to convolve_fft. I found that if I change my convolve_fft(image,psf) to convolve_fft(image,psf,nan_treatment='fill',normalize_kernel=False) I could then reproduce the results I was getting previously. I take it the defaults were changed at some point, presumably related to #926. Can I suggest that it would make debugging old code much easier if changes that break backwards compatibility, such as this, are noted in the documentation? I'm thinking of something along the lines of http://docs.astropy.org/en/v2.0.x/io/fits/api/files.html#writeto where it clearly states 'Changed in version 1.3: overwrite replaces the deprecated clobber argument.'

The examples also need to be completely rewritten as they were clearly written using the old defaults. For instance, the second one convolve_fft([1, np.nan, 3], [1, 1, 1]) now gives the result array([0.5, 2. , 1.5]).

[bsipocz]

Were do you find that example that is giving a different result than listed for the example? We normally test those automatically, so if they are wrong then something must have got broken along the way.

[astromark]

Here for example:
http://docs.astropy.org/en/stable/api/astropy.convolution.convolve_fft.html

[bsipocz]

You're right. We somehow have a __doctest_skip__ = ['*'] in that file, and that shouldn't be the case. Thanks for reporting.

[keflavich]

Right, we had major backward compatibility breaking changes in the v2.0 release, I believe. Those were (should have been?) documented in the changelog, but not directly in the docs.

[astromark]

Another place that I've noticed an issue is the 'What's new in Astropy 2.0?' document, which specifically states "convolve_fft’s behavior remains unchanged".
http://docs.astropy.org/en/stable/whatsnew/2.0.html#improvements-to-astropy-convolution

[bsipocz]

There were significant convolution changes in 3.1, too, they might have been internal only though.