`import sgkit` takes ~1.5s

[benjeffery]

For tools using the CLI this amount of delay feels excessive. Around 1s of this time is performing imports. Here's the import flame graph:
(0.1s on xarray.tutorial?!?)

I assume some of these could be imported only when they are needed - although given that many of the imports are referenced in typing specifications that might not be possible.

I'm not sure yet what the remaining 0.5s is - a cProfile callgraph is pretty useless on import but I think there is a way around that by profiling the individual sgkit files.

[tomwhite]

I agree that deferring some of the imports could help.

I think numba compilation is still a significant part of this, see #363, although we now have numba caching on so it's faster the second time. Not sure if we could defer this until it's needed too.

[benjeffery]

Thanks @tomwhite I'll check the numba compilation.
I also wonder if PEP484 forward references might help with delaying imports: https://legacy.python.org/dev/peps/pep-0484/#forward-references

[tomwhite]

I also wonder if PEP484 forward references might help with delaying imports: https://legacy.python.org/dev/peps/pep-0484/#forward-references

It would be great if we could take advantage of forward references.

Do you think xarray itself is doing unnecessary imports - like the tutorial?

[benjeffery]

Do you think xarray itself is doing unnecessary imports - like the tutorial?

Yes, planning on checking and raising a PR/issue upstream, unless it looks like a rabbit-hole!

[benjeffery]

Also see pydata/xarray#6726, seems they are aware of the issue.

[jeromekelleher]

If we're importing some things just for typing purposes, then I'd be +1 for making the typing less strict.

[tomwhite]

Pandas is not used very much in the codebase, so it might be possible to import it lazily.

Similarly, the distance API is pretty niche so making that lazier would be good too.