When n != nfft, getindex(::ArraySplit, i) fills the fft buffer with n elements rather than nfft. This can result in a lot of trailing zeros in the buffer, so periodograms on irregular length data are likely to be contaminated with artifacts from the frequency response of a boxcar window.

On a related note, it would be great if we could change the default windowing function to something better. I'm aware that there can be subtle tradeoffs when choosing a window, but just about any choice will be a better default than what we have now: currently it's pretty much mandatory to set the window by hand, or the results will contain a lot of spurious ringing. I don't know a lot about window choice, but matlab seems to default to the Hamming window for pwelch().