Numpy and Scipy slow on completion.

[hifall]

I use Palantir's Python Language Server (PLS) for Python code completion. PLS in turn uses Jedi as the completion engine.

I found that triggering the completion list is very slow for Numpy and Scipy on a VPS. It takes 15 seconds for Numpy and 10+ seconds for Scipy respectively.

There are already some performance issues raised about Jedi in the community. But none of them is about Numpy or Scipy.

I wonder if this is a known issue?

[davidhalter]

It's a known issue that Jedi is slow for numpy, but it's definitely not that slow. Sometimes the first time you try to complete numpy it takes a few seconds (maybe 5), but after that completions should be pretty fast.

TLDR; I checked and your VPS sucks ass, because the https://cloudscale.ch servers are way faster.

---

I just checked with the cloud servers of my company (cloudscale.ch), and these are the results:

>>> from timeit import timeit
>>> import jedi
>>> timeit("jedi.Script('import numpy; numpy.').completions()", globals=globals(), number=1)
3.0075249230000054
>>> timeit("jedi.Script('import numpy; numpy.').completions()", globals=globals(), number=1)
0.23607791599999928

This means that the first completions takes 3 second and everyone following it 230 ms. It's a bit slower on my computer, it takes about 420 ms. I'm not saying that's good, but it's way better than 15s. I'm also aware of exactly that half a second for a completion problem, but that's known and is probably going to be solved, once we have an index (see #1059).