dvc does not cache stream opened files

[skshetry]

Noticed during #8786 that we don't cache files that were streamed from remotes. So subsequent plots show always hits the remote and is thus slow.

[efiop]

Might be solved by #9037

[efiop]

@skshetry Could you try with the upstream, please?

[skshetry]

I don't think so. DVCFileSystem does not support caching, we just use fs.open() which only streams. AFAIR we had fetch=True that would fetch and then re-open from local cache in the past but we got rid of that in 2.0.

[efiop]

@skshetry Ah, I see. Maybe it is worth doing dvc fetch explicitly there instead.

[skshetry]

In plots, we return a dict with partial function to read file, and resolve later, so we can't do a fetch there.

dvc/dvc/repo/plots/__init__.py

Lines 169 to 179 in 4f49023

	res[fs_path].update(
	{
	"data_source": partial(
	parse,
	fs,
	fs_path,
	props=joined_props,
	onerror=onerror,
	)
	}
	)

Since we are reading the file completely in memory on fs.open(), an easier way would be to bring fetch=True in fs.open() that'd first cache locally and reopen from it. :)