dvc get/artifacts get: repo clone is slow

[dberenbaum]

Bug Report

Description

dvc get is slow to clone the git repo.

This is shown in the DVC benchmarks:

Reproduce

I can also see it by profiling the example command dvc get -f git@github.com:iterative/example-get-started-experiments.git models/model.pkl --viztracer --viztracer-depth=32 (CLI git clone for this repo takes <1 second):

viztracer.tar.gz

Expected

The operation should take a few seconds at most instead of almost a minute.

[skshetry]

@dberenbaum can you try with more depth?

[dberenbaum]

@skshetry More viztracer-depth? This is already at depth of 32 and looks like the slowdown is in dulwich fetch. Would it help you to have depth 64?

Edit: Happy to provide it but wasn't sure if maybe you meant git depth or something. Also would like to see if you can reproduce it independently.

[skshetry]

@dberenbaum, yes, fetch is slow, but more information about where it's getting blocked would be helpful.

For me, it's not working at all. It's just getting stuck forever on client.fetch_pack(). It could be that my internet is slow. But this has likely to do something with asyncssh client.

[skshetry]

Also, could you please try with --viztracer-async? There should be some tasks from asyncssh that should be visible, but above log does not have anything like that.

[dberenbaum]

Depth 64:

viztracer.tar.gz

Depth 64 with async:

viztracer.tar.gz

[pmrowla]

Also, do you get different clone performance with https github URLs vs ssh?

I'm not sure this is related to the asyncssh client given that the dvc-bench test is only cloning local paths