Unexpectedly the computation of 'sqeuclidean' metric is slower than that of 'euclidean' metric, in scipy.spatial.cdist(). Given that, in principle, 'sqeuclidean' is the same as 'euclidean' but without the expensive sqrt(), then the computation should be much faster.
In : X = np.random.random((100,3))
In : Y = np.random.random((80,3))
In : %timeit cdist(X, Y, metric='euclidean')
10000 loops, best of 3: 127 us per loop
In : %timeit cdist(X, Y, metric='sqeuclidean')
10000 loops, best of 3: 152 us per loop
The reason of this issue is related the implementation of 'sqeuclidean' metric which computes the distance by first computing the 'euclidean' one and then computing the square (with "**2" !, see https://github.com/scipy/scipy/blob/master/scipy/spatial/distance.py#L1224 ). The main reason behind all this is that there is no C implementation of 'sqeuclidean' to wrap.
I may provide a pull request in near future to address this issue.
I think this has been fixed and merged.