Re: RFC: Resumable clone based on hybrid "smart" and "dumb" HTTP

From: Shawn Pearce <spearce@spearce.org>
Date: Feb 14, 2016, 17:56 UTC
Message-ID: <CAJo=hJuvHpWPQGV1gwgwOATy+8OO0Smt-BG9xb3q7eeJSypdaA@mail.gmail.com>
In-Reply-To: <20160214170525.GB10219@sigill.intra.peff.net>

On Sun, Feb 14, 2016 at 9:05 AM, Jeff King <peff@peff.net> wrote:

Show 21 quoted lines

> On Sat, Feb 13, 2016 at 06:14:31PM -0800, Shawn Pearce wrote:
>
>> > And with "resumable=<url>", the client does not have to hit the server
>> > to do a redirect; it can go straight to the final URL, saving a
>> > round-trip.
>>
>> It occurred to me today that to make the above ("resumable=<url>") as
>> efficient as possible, we should allow HTTP clients to include
>> &resumable=1 in the GET /info/refs URL as a hint to the server that if
>> it serves a resumable=<url> it can shrink the ref advertisement to 1
>> line containing the capabilities.
>
> I'm slightly wary of this. The client advertising "resumable=1" does not
> mean "I will definitely do a resumable clone". It means "I support the
> resumable keyword, and if you share a resumable URL with me, I _might_
> use it, depending on things that are none of your business" (e.g., if it
> does not like the server URL's protocol).
>
> It is recoverable by having the client re-contact the server without the
> resumable flag, so it could still be a net win if the client will select
> the resumable URL a majority of the time.

That was my thinking; if the client starts out with &resumable=1 and then doesn't like the URL the server offers in the advertisement it can drop the query parameter and relist the references for a standard dynamic clone.

Show 5 quoted lines

> I'm also not happy about having an HTTP-only feature in the protocol. I
> liked Stefan's proposal for the "v2" protocol that would let the two
> sides exchange capabilities before the ref advertisement. Then the
> client, having seen the server's resumable URL, knows whether or not
> to proceed with the advertisement.

Ok, that's fair. Lets forget the "&resumable=1" for now so that the various protocols are more identical. Later if "v2" makes enough progress we can rely on its capability exchange before advertisement to get this potential bandwidth savings.

Show 9 quoted lines

>> Clients are going to follow the returned <url> to get a bundle header,
>> which contains the references. And then incremental fetch from the
>> server after downloading the pack. So returning references with the
>> resumable URL during a clone is an unnecessary waste of bandwidth.
>
> If the bundle is up to date, the client can skip the follow-up
> incremental fetch, as it knows that it has everything needed for the
> original ref advertisement it got. Whether that's a net win depends on
> how up-to-date the bundles are.

Well, that comes down to "was the repository repacked since last push". If yes the pack contains everything and its pack-*.info / split bundle header thing contains the same contents as packed-refs.

I think in practice clients aren't going to bother with the implementation detail of saving the initial ref advertisement aside, grabbing the bundle header, and comparing them to see if it needs a subsequent fetch. They are just going to discard the ref advertisement, retrieve the pack, anchor the objects using the bundle header, and then reconnect.

Remember this isn't just about holding the refs in memory. To be resumable the client process may have to be restarted, so that initial ref advertisement has to be written out to disk to persist.

Show 13 quoted lines

> If "C" is the cost to contact the server at all and "A" is the cost of
> the advertisement, then a "hit" with this scheme means the overhead is
> C+A (we contact the server only once). A "miss" means we have do the
> followup fetch anyway, and we pay 2C+2A (paying the advertisement cost
> both times). Whereas with your scheme, we pay 2C+A always; two contacts,
> but only the second has an advertisement.
>
> So it depends on the relative cost of C and A, and how often we expect
> it to kick in.
>
> In practice, I suspect it's mostly dominated by the cost of the actual
> clone objects anyway, but maybe that is different for Gerrit. I hear
> refs/changes/ can get pretty big. :)

I hear refs/pulls/ can also get big. :)

But yes, some Gerrit advertisements are not small.