issue_comments: 153258925

A useful point of reference here is to compare xray's performance to pandas:

``` In [20]: %%timeit t = pd.DataFrame({'x': np.arange(10000)}) for _ in t.iterrows(): pass ....: 1 loops, best of 3: 558 ms per loop

In [21]: %%timeit t = DataArray(np.arange(10000)) for _ in t: pass ....: 1 loops, best of 3: 1.49 s per loop ```

So we're about 2.5-3x slower than pandas. We might be able to catch up there with some careful optimization, but I doubt we would be able to do much better.

Basically, the issue is that DataArray (like pandas's Series) is not an extension type (written in C/Cython), which means that it is much slower to construct -- iterating over it entails a lot of Python logic and a lot of memory allocation (e.g., a new dictionary for every instance).

We could imagine converting xray's core data objects to Cython extension types, but I doubt the tradeoff in easy of use would be worth it. I prefer to encourage users to use xray's easy interface to NumPy when entering into something performance critical like a loop.