On Fri, Aug 12, 2022 at 11:49:06PM +0800, ZheNing Hu wrote:

quoted

...I agree with this. It is nice to put more power in the hands of the
clients, but we have to balance that with other issues like server
resource use. The approach so far has been to implement the simplest and
most efficient operations at the client-server level, and then have the
client build local features on top of that. So in this case, probably
requesting that _no_ trees are sent in the initial clone, and then
faulting them in as the client explores the tree using its own local
sparse definition. And I think that mostly works now.

Agree. But we have to fetch these blobs one by one after partial clone,
why not reduce some extra network overhead If we can get those blobs
that are *most* needed in the first partial clone, right?

Ideally we wouldn't be getting them one-by-one, but rather batching
them. I'm not sure of the current state of the code. But we should at
least be able to batch by tree depth.

You're right that we'll never be able to be as efficient as a server to
whom we can say "I care about these paths", though.

-Peff