I didn't see any explicit nan handling. Are missing values allowed
e.g. in the constructor?
No, this is a valid point. We don't handle this as explicitly as we should. Are you mostly talking about nan handling in loading from delimited text files? (Or are you talking about something more general, like integration of masked arrays?) In loading from delimited text files, you can use the "linefixer" and "valuefixer" arguments, which are for more general purposes, and which will get the job done, but slowly. We should do something more specialized for missing values that would be faster.
Are these function supposed to work with arbitrary structured arrays?
Well, they're only really tested for working with strings, floats, and ints (tho only the int tests are included in the test module, we should expand that). I imagine it's possible they'd work with more sophisticated things but I'm not sure.
>>> arr = np.array([6,1,2,1e-13,0.5*1e-14,1,2e25,3,0,7]).view([('',float)]*2)
>>> arr
array([(6.0, 1.0), (2.0, 1e-013), (5e-015, 1.0),
(2.0000000000000002e+025, 3.0), (0.0, 7.0)],
dtype=[('f0', '<f8'), ('f1', '<f8')])
>>> np.sort([str(l) for l in arr])
array(['(0.0, 7.0)', '(2.0, 1e-013)', '(2.0000000000000002e+025, 3.0)',
'(5e-015, 1.0)', '(6.0, 1.0)'],
dtype='|S30')
Well on this example (as in tests that we did), fast.recarrayisin performed as spec'd. ... But definitely write back again if you think it's failing somewhere.
In general, extending a number of the thigns in Tabular (e.g. the loadSV and saveSV) to arbitrary structured dtypes as opposed to more basic types would be great.