Hmm, this is interesting. The size of a frame and its id are at fixed positions. Bytes 1,2 are length, 3,4 are type and 5,6,7,8 are the id. I'm just not sure I could get multiple frames received on a single socket to independently route through different backends, and for response frames to go back through the correct incoming socket.
On Sun, Jul 10, 2016 at 8:10 AM, Chad Lavoie <[email protected]> wrote: > Greetings, > > On 7/10/16 6:33 AM, Matt Esch wrote: > > I need to load balance a custom tcp protocol and wonder if HAProxy > > could be configured or extended for my use case. > > > > The protocol is a multiplexed frame-based protocol. An incoming socket > > can send frames in arbitrary order. The first 2 bytes dictate the > > frame length entirely (so max 64k per frame). The frame has a type and > > a header format, followed by a payload. > > > > The multiplexing works by assigning long ids in the frame header and > > pairing the responses based on this id. > > Depending on what the id's actually look like this may or may not work, > but before I started writing a lot of C I'd try something such as > payload() per > https://cbonte.github.io/haproxy-dconv/configuration-1.5.html#7.1.5 to > match an acl for use_backend; or stick on (using a table) via > https://cbonte.github.io/haproxy-dconv/configuration-1.5.html#stick%20on. > > Again, entirely possible that its not usable for your use-case, but if > it is it sounds much easier than trying to code another format. > > - Chad > > > > A load balancer would read frames from the socket, parse the frame, > > select the correct backend from a header and route the request through > > an available backend peer. The response would be matched to the > > incoming socket based on the frame id. > > > > I expect (a lot of) c code will need to be written to support such a > > custom protocol. I'm looking for specific pointers about how to add > > such a protocol to the existing codebase in a way that would fit > > cleanly, and preferably in a modular fashion. > > > > Any hints appreciated > > > > > > ~Matt > > > > > >
