1.3.16 is getting closer...

From: Willy Tarreau <w#1wt.eu>
Date: Mon, 8 Dec 2008 01:20:08 +0100

Hi all,

for the last few months, haproxy's code was a terrible mess. I could finally finish to reorder things and get it to work again.

Its internal structure has changed a bit so that HTTP is no longer a hack on top of TCP. Indeed, it was becoming increasingly difficult and risky to add even minor HTTP features. Several times we unexpectedly broke lower layers as side effects.

Now, the internal structure looks very much like the diagrams in buffers.fig. There are several layers which can behave independantly :

    layer 7 => eg: HTTP request
    stream protocols (unix, tcpv4, tcpv6)     file descriptors

The goal was to be able to easily add new layer7 protocols or features without breaking the lower layers anymore. I would say that this goal is almost achieved. I'm saying "almost" because I am not completely satisfied with the lower layer, which will still evolve. But at least higher layers should not change much.

Some previous tricks were very easy to implement in the new model. For instance, the tarpit is just a protocol analyser which does nothing until it reaches its timeout. In previous model, it was a state sitting somewhere in the middle of HTTP and TCP states. Also, a request or a response can now be parsed even with one side shut down, which was not possible before, hence the tricks to pretend the connection was still open. Also, doing a turn-around after a failed connection attempt is completely transparent to HTTP now.

Unix sockets and TCP protocols have converged a lot. Not enough yet, but still a lot. Large parts of the code were copy-pasted, which is a good sign.

The stats interface has been cleaned up a bit, and it was so much easy to add a new feature that I added the session dump command I've been missing during all this revamping.

I've merged all the fixes from the 1.3.15 branch. It's very likely that the code still has bugs, but I would say that having it that much modular has profundly eased the debugging sessions, and I think that it will take less time to rule all of them out now.

I still need to implement the ACLs on the response, which was the initial reason to break the old model, and after that, I think it will be time to release 1.3.16.

I'm strongly thinking about making that one the future stable release and opening a new development branch, because having the development branch being used in production everywhere is slowing the development down a lot.

I've pushed the changes to the git tree. A snapshot will be automatically packaged in 3 hours and uploaded there so the brave among us can give it a spin :


I've installed it in front of haproxy's site to ensure there's no problem with real-world traffic, so if you can't access the site, it probably means I've failed somewhere ;-)

Future features might include DSO loading for layer 7 protocols, better tcp-splicing abilities, ability to chain multiple processings (eg: SSL), the beginning of a control socket to set servers up/down, QoS in the request queue, and perhaps the beginning of a minimalistic keep-alive support.

Willy Received on 2008/12/08 01:20

This archive was generated by hypermail 2.2.0 : 2008/12/08 01:30 CET