Re: Fetching 24 Linux commits = 1.2 GiB

From: Jeff King <peff@peff.net>
Date: Apr 15, 2020, 15:16 UTC
Message-ID: <20200415151652.GB2464307@coredump.intra.peff.net>
In-Reply-To: <xmqq1roo947y.fsf@gitster.c.googlers.com>

On Wed, Apr 15, 2020 at 08:08:17AM -0700, Junio C Hamano wrote:

> Do these (and I think we saw other reports) make us rethink the
> status of protocol v2 as the default?  Are all of these fallouts 
> we saw so far easy-to-fix bugs, or are there more fundamental issues
> in the v2 protocol design?

I don't think we know yet.

I agree with Konstantin that the v2 switch is the likely culprit for these issues, but without having been able to reproduce, I don't think we know exactly what the problem is yet. It could be a protocol design issue, or it could be a minor implementation bug.

Note that there is one other issue that's turned up, that I discussed here:

  https://lore.kernel.org/git/20200328154936.GA1217052@coredump.intra.peff.net/

That's more fundamental to the v2 design, but:

  - it only happens when one side drops the connection, so it's not
    impacting normal operation (it does turn an error case into a hang,
    though, which can be rather annoying)

  - it's not in the network protocol itself, but rather the protocol
    between Git and the remote helper. So we could solve it purely as a
    client-side fix.

-Peff