auto complete suggestion takes almost 6 seconds [Performance]

[killown]

Auto complete suggestion takes almost 6 seconds to appear in a file com 900 lines

Environment data

VS Code version: 1.19.1
Python Extension version: 0.9.1
Python Version: 3.5
OS and version: Manjaro "updated"

Actual behavior

auto complete takes 6 seconds to appear

Expected behavior

appear instantaneously

Steps to reproduce:

Logs

Output from Python output panel

Output from Console window (Help->Developer Tools menu)

workbench.main.js:140176 [Extension Host] debugger inspector at %cDebugger listening on port 9334.
Warning: This is an experimental feature and could change at any time.
To start debugging, open the following URL in Chrome:
    chrome-devtools://devtools/bundled/inspector.html?experiments=true&v8only=true&ws=127.0.0.1:9334/d2af9450-6983-4e59-8a41-0ac4a7f5dbd8
workbench.main.js:18913 [Extension Host] (node:16144) DeprecationWarning: os.tmpDir() is deprecated. Use os.tmpdir() instead.

[brettcannon]

Can you provide us with sample code that triggers the 6 second delay?

[MikhailArkhipov]

It has been a while. We need sample code to see whether the issue is in the jedi or elsewhere. Feel free to reactivate.

[balta2ar]

I'd like to share my experience. I also noticed really annoying delays. Sample code is very simple:

from tensorflow import tf
tf.

I'm using tensorflow 1.5, python 3.6 in a conda environment (someone in the comments noticed that jedi works faster with python3 than with python2).

Actually I mentioned this issue here: davidhalter/jedi#937 (comment)
I'm not 100% sure who's causing these delays: VSCode or solely jedi. I said in the comment that delay is about 2 sec. Now it feels more closer to 1 sec but still if you type tf. a lot, the delay really gets on your nerves because it's far from instant. And it's definitely larger than 300ms which we can see in "synthetic" jedi test mentioned in the comment (I'd be glad to have only 300ms!). Could it be that VSCode is discarding some data structures (maybe along with jedi itself) that could be kept and reused?

I'm not sure how to track and it down and measure exactly. Besides, VSCode uses not the latest version of jedi, which might be different in terms of performance (but that's not certain). At least it could be checked whether VSCode adds up to that delay, right?

[DonJayamanne]

@balta2ar
Please could you download our Insiders Build of the extension and give this a go.
We're using the latest version of Jedi and this improves the performance significantly.
At my end the autocompletion list popups up in under 1 second.

Please try the following code:

from tensorflow import <>

Or

import tensorflow
tensorflow.<>

The following code doesn't provide any autocompletion (Jedi issue):

from tensorflow import tf
tf.

[balta2ar]

I've installed from VSIX and now the version says "2018-2.0-beta", so hopefully I'm trying the right version.

However, the delay is still there for me. I agree it's around one second but it's still annoying... I can't say it got any faster after update, but that's just my subjective perception. I'd be glad to give actual numbers if you could guide me with instrumenting the code, I'm just not sure where and how to put measurements.

Though, as David said in another jedi issue, we may very well be approaching the limits of jedi in terms of performance and it's hard to squeeze any more without aggressive caching.

I think I mistyped in my comment and I wanted to write this:

import tensorflow as tf
tf.

but you got me right.

[MikhailArkhipov]

You may be interested in https://github.com/Microsoft/vscode-python/projects/7#card-8587079 then :-)

[MikhailArkhipov]

No delay with VS engine. Comparable to PyCharm.