I find the default values work for me; but sometimes if I have a big signal with lots of missing values, limiting the length of the regression is helpful if all I want is a quick answer. Ideally the length would be just enough to cover the portion of the signal that has a constant AR model. If I'm missing a portion of a slowly varying oscillation, I use a larger value to cover a few cycles of it. If instead I'm missing just a portion of a decaying signal, then I keep the length to be within the decaying portion.
As for order, that can vary. I often start low at first, then increase. Most of the audio clips I tested tended to top out around an order of 200.
This example tries to explain what's going on under the hood. If you're looking at something that can just interpolate a couple of points that aren't oscillating much (if at all), then take a look at fillmissing.
Let us know what you tried and share your data if you can.