Re: Git Smart HTTP with HTTP/2.0

From: Shawn Pearce <spearce@spearce.org>
Date: Jul 11, 2015, 17:23 UTC
Message-ID: <CAJo=hJs21m1C6+rdvCid311-TapK=QKLkqrH8aUZmzHH7CpVug@mail.gmail.com>
In-Reply-To: <20150711070055.GA4061@LK-Perkele-VII>

On Sat, Jul 11, 2015 at 12:00 AM, Ilari Liusvaara <ilari.liusvaara@elisanet.fi> wrote:

Show 15 quoted lines

> On Sat, Jul 11, 2015 at 11:10:48AM +0800, ForceCharlie wrote:
>> As we known, HTTP/2.0 has been released. All Git-Smart-HTTP are currently
>> implemented using HTTP/1.1.
>
> Nit: It is HTTP/2.
>
>> Frequently used Git developers often feel Git HTTP protocol is not
>> satisfactory, slow and unstable.This is because the HTTP protocol itself
>> decides
>
> Note that there are already two versions of HTTP transport, the old "dumb"
> one and the newer "smart" one.
>
> The smart one is difficult to speed up (due to nature of the negotiations),
> but usually is pretty reliable (the efficiency isn't horrible).

The negotiation in smart-HTTP actually has some bad corner cases. Each round of negotiation requires a new POST resupplying all previously agreed upon SHA-1s, and a batch of new SHA-1s. We have observed many rounds where this POST is MiBs in size because the peers can't quite agree and have to keep digging through history.

The native protocol on git:// and SSH is not as bad. Negotiation is still many rounds, but these are pipelined and each round does not need to repeat the prior round, as the server has a single stream and is saving state.

Show 7 quoted lines

> Now, the old "dumb" protocol is pretty unreliable and slow. HTTP/2 probably
> can't do anything with the reliability problems, but probably could improve
> the speed a bit.
>
> Websockets over HTTP/2 (a.k.a. "websockets2") has not been defined yet.
> With Websockets(1), it would probably already be possible to tunnel the
> native git smart transport protocol over it. Probably not worth it.

Another option is to tunnel using gRPC (grpc.io). libcurl probably can't do this. And linking grpc.io library into git-core is crazy. So its probably a non-starter. But food for thought.

But, at $DAY_JOB we tunnel the native bidirectional protocol in grpc.io's predecessor, and it works quite well for us.