regression: dvc init is slow

[skshetry]

The 3 major issue seems to be:

• tpi (see slow import on dvc iterative/tpi#14) (~300ms)
• fsspec callback (~200ms)
• prompts import on experiments from rich (for dvc exp init) (~50ms) move exp init logic to dvc.repo.experiments.init #6738
• dvc.fs importing fsspec by default (~100ms)
  (There may not be much point in making fsspec lazy, as we'll need local filesystem anyway).

The following is a startup log for dvc init.

Following is the benchmark for dvc init.

$ hyperfine "dvc init" --warmup 3 -p "rm -rf .dvc"
Benchmark #1: dvc init
  Time (mean ± σ):      1.311 s ±  0.065 s    [User: 1.066 s, System: 0.203 s]
  Range (min … max):    1.230 s …  1.435 s    10 runs

dvc init crossing 1 sec is really bad, this was the result a month ago (before tpi and prompts, and with fsspec.Callback made lazy):

$ hyperfine "dvc init" --warmup 3 -p "rm -rf .dvc"
Benchmark #1: dvc init
  Time (mean ± σ):     506.8 ms ±  31.4 ms    [User: 369.2 ms, System: 103.7 ms]
  Range (min … max):   477.4 ms … 566.6 ms    10 runs

[skshetry]

This is why I prefer having benchmarks in the repo (or, at least run some tests in each PR). It's hard to figure out later.

[pmrowla]

TPI related issues should be fixed in current DVC master

[skshetry]

We have slowly regressed on init side (most likely due to fsspec imports) by 25% between 2.6 and main.

But nothing we can do at this time. Closing...