S3File without an S3FileSystem

[TomAugspurger]

This could be at odds with your goals, so feel free to just shoot it down!

Pandas' use-case for s3fs is at the key / object-level, and S3File taking an instance of S3FileSystem I think just causes headaches we'd have to work around (mainly permission level stuff).
Part of that is being worked on here.
A complementary approach would be to have S3File take a boto3 s3 connection instead of a S3FileSystem. A quick glance through the code showed that every use of S3File.s3 was only to get at the connection, except for S3File.close, which invalidated the bucket cache.

I'm going to play with a couple approaches today and see what breaks. Do you see any hope for this happening?

Also the docstring for S3File says that the first arg should be a boto3 connection. I'm hoping we can make that true :)

[martindurant]

Sure, be my guest.
One additional reference to S3FileSystem (see #46) is S3File.info(), which gets the cached entry if it exists, or does a list_objects or then head_object as fallback.