For those following this issue:

Mercurial 1.9 includes my ground-up rewrite of httplib, which you can enable
with ui.usehttp2. When usehttp2 is enabled and you're talking to a 1.9-based
http(s) server, 100-continue is properly used and should avoid the
double-send of the bundle data in (nearly) all cases. If you find defects
(either in the new continue support or regressions caused by usehttp2),
please report them and assign the bug to me.

Thanks!

Degrading to bug.

Augie,

Thanks for all your work on this. I'll take a look at the queues and talk to
my lead about my assisting with testing.

On Sun, Mar 28, 2010 at 8:53 PM, Augie Fackler
<bugs@mercurial.selenic.com>wrote:

>
> Augie Fackler <durin42@gmail.com> added the comment:
>
> I just posted a status update to the hg list. Linking here for sanity.
> http://article.gmane.org/gmane.comp.version-control.mercurial.devel/29897
>
> ____________________________________________________
> Mercurial issue tracker <bugs@mercurial.selenic.com>
> <http://mercurial.selenic.com/bts/issue1876>
> ____________________________________________________
>

I just posted a status update to the hg list. Linking here for sanity. 
http://article.gmane.org/gmane.comp.version-control.mercurial.devel/29897

Interesting. Bear in mind that we'll probably be wanting to revamp our SSL
support to not suck.

The cleanest best path appears to be to fix (that is, rewrite most of) 
httplib. I'm planning on taking 20% time for Q1 to work on that.

The other option is to monkeypatch httplib heavily, which is fragile at best. 
I spent 3 days or so on it, and came to the conclusion that we'd be better off 
making a clean async-safe httplib replacement instead.

What's the status with this one?

I've got the beginnings of a client-side fix for this as well, it seems to be 
working at a basic level.

hgweb actually has code to handle this (look for req.drain() in the code). 
It's inelegant, but server implementations can do that as a stopgap until hg 
is fixed.

I'm going to try and spend 1 day a week on this until it gets fixed.

I don't know that it got discussed in irc. I submitted a preliminary version 
of the patch to the maillist, but I want to improve it some (catch other 
errors, like "ssl required") before resubmitting it again. Now that I've 
gotten single-dir clones done for hgsubversion, I'll be revisiting the patch.

Thanks tonfa, sorry I missed this one on irc. Is that conversation stored
anywhere?

Don't think Augie has a full fix for this yet.

The small request to trigger auth is clever. It would alleviate failures due
to a server closing the pipe on an unauthenticated body. But my read on the
code says the next request would also have to re-authenticate. So a smaller
bundles would stand a better chance of not failing, but still take twice the
traffic on many servers.

Putting Augie in the nosy list, since he worked on a solution to this 
(sending an almost empty bundle to trigger the auth).

Is it the same effort that the one discussed on irc some time ago?

The behavior of mercurial.url, mercurial.keepalive, and urllib2 have every
request sent twice to http servers requiring authentication.

This behavior can cause failures or huge (>8mb documented) re-transmissions
when the 'hg push' sends the 'unbundle' command.

Details and fix suggestions at bitbucket.org/bradobro/liquidhg/wiki/TheFixes.