Better code reuse here seems good, and enables memoization etc tricks

Also, I think somewhere in the code I realized a useful trick was, when just a check is needed, instead of trying via an import (slow), do a check for the module in the path, which skips the costs of actually importing (which runs all sorts of code in the dep's top-level)

is it cheating to use another library to help?

import importlib.util, sys
def check_umap():
  umap_spec = importlib.util.find_spec("umap")
  has_umap = umap_spec is not None
  if has_umap and "umap" not in sys.modules:
     import umap
     return umap, has_umap
  else:
     return None, has_umap

i see the way we do it now is better for >3.6, but since we want to check and not import every time perhaps if > try?

Hmm, these calls do different things;

• importlib will see it is resolvable: This has false positives - still needs an 'import' w try/catch to truly validate - you can DOWNLOAD cudf on a sys without a GPU, just not import it. A downside is it might be a tad slower as a full search vs kv lookup.

• sys.modules will check if it is already imported, but has false negatives in that it will miss existence when cudf is downloaded but the user hasn't imported elsewhere yet (I think?) . That is arguably a feature, but is not obvious to work with, so maybe a bug?

So yeah, maybe:

• encapsulate
• the True/False is unnecessary , return Tuple[Optional[package], Optional[exn]] ?
• Hopefully mypy is ok with a pattern here..
• find_spec. Changes behavior, but likely to match user intent better.
• with caching decorator so won't search every single time