February 27, 2002 spout

HTTP is Dead?

Your friend and mine, Don Box, caused quite a stir yesterday with his keynote at European DevWeek in London. Peter Drayton has also written up a summary (which is much more technically meaty), as well as a commentary. There has been some quite spiriting follow up on this talk all over the Internet: the .NET mailing list, the Off Topic mailing list, the REST mailing list and XML Deviant on XML.COM. Also, while I absolutely agree with Don that the way we use HTTP today leads to trouble, I thought that Ian Griffiths, a fellow DevelopMentor instructor, had a wonderful point of view that he allowed me to share:

Guest opinion by Ian Griffiths

Basically Don seemed to be saying that there are two problems with HTTP (and saying HTTP is dead is just an effective way of getting people to listen; I was half tempted to start my Windows Forms speech at the UK MSDN DevCon with The Web Application is Dead”).

One of these is that HTTP is unidirectional. Surely .NET remoting shows that this isn’t strictly true: individual connections are directional but it’s entirely possible to do callbacks by having connections go in both directions. (Well duh.) The real problem is that the firewall architectures of the internet are designed to make sure the connections only go in one direction; it’s not a problem with the protocol per se, it’s a problem with the infrastructure. In order to fix this problem you need to change the infrastructure regardless of what you do to the protocol. And if you fix the infrastructure you don’t actually need to fix the protocol, since we already know that it’s fine on networks that don’t deliberately break bidirectional communication.

Arguably one of the main obstacles here is the use of NAT - NAT makes it hard for a client behind a firewall to publish an endpoint. But NAT is fundamentally important because we’d have run out of IP addresses already if it weren’t in such widespread use. So the only way to get rid of NAT is for everyone to upgrade to IPv6. This will presumably happen fairly soon since we will run out of IP addresses in any case in about 3 or 4 years. (Windows XP ships with IPv6 support by the way. Type ipv6 install” at a command prompt if you haven’t already. So Microsoft are quietly making IPv6 ubiquitous on the desktop.)

So presuming ipv6 takes off, that’s a fundamental technical obstacle to the one-way nature of HTTP removed. But another one remains: just because IPv6 pushes the address exhaustion date out of our lifetimes (we hope), doesn’t mean that firewall admins will let connections work both ways through the firewall. The bidirectional problem can only be solved with the blessing of firewall admins. And once you have that you don’t actually need IPv6, strictly speaking - given a suitable protocol between the client machines and the firewall there’s no reason that NAT can’t be done in both directions. (Indeed my sub-$100 firewall appliance can do this on a statically configured basis. It is possible (if a little inconvenient) for machines behind the firewall to publish endpoints successfully.)

UPnP apparently has a solution for this. (A better one than static configuration.) It supports P2P network applications that require clients behind the firewall to be able to accept incoming connections. It defines precisely what you need: a protocol that lets a client machine negotiate with the firewall to open up a port for incoming connections. So it turns out that a technical solution to the problem already exists. (Without even having to invoke IPv6. Although we still need that, due to a shortage of address bits in IPv4.)

So the first problem already has a technical solution. Presume for a moment (and it’s a big presumption) that these solutions can be deployed, and firewall admin policies set up so that they are broadly useable.

At this point, the second complaint Don makes - the fact that long-running requests don’t sit well with HTTP - becomes much less of an issue. We just need to use separate HTTP requests for the request and the response. We connect to the server, send a request, along with some response endpoint info (a URL). When the server is done it connects to us, sends us a response. And we’re done. I’m guessing it would probably be possible to write a .NET remoting channel that works like this. The only obstacle is the ability for clients to advertise endpoints.

So if a client can advertise an endpoint somehow (i.e. it can (a) convince the firewall to let a connection request come in, and (b) work out what the endpoint should be - it needs to be aware of any NAT translation going on) then HTTP can actually solve both the problems raised here. And as far as I can tell, *any* solution to the problems raised is going to have to allow the client to receive incoming requests. In which case why invent a new protocol? Once you’ve solved the fundamental problems in the network that stop you from doing this, HTTP is good enough.

The only issues are: (1) getting everyone to agree on how clients will expose endpoints (UPnP has had nothing but bad press so far, since the only thing most people know about it is that the first security hole discovered in Windows XP was connected to it somehow; so it might have to be something else…), and (2) convincing firewall admins to allow this functionality to be used.

Of course just because a technical solution exists doesn’t mean it’s a great solution. Given that this is a fair distance from HTTP as originally envisaged, it would doubtless be possible to design a protocol better suited to the job. But would the benefits be worth it? Everyone already has an HTTP stack and an XML parser - will they flock to implement a new purpose-built solution?

The problems raised in the ZDnet article can essentially be summed up thus: clients can’t expose endpoints. To me, this doesn’t look like a problem with HTTP, it’s a problem with the infrastructure. Changing the protocol is certainly not sufficient to fix the infrastructure problems; it’s not clear to me that it is even necessary.

So really this is a social problem, not a technical one. ;-)

If you’re interested in this topic or what other interesting things Don will say next, he’s giving the keynote address at the Web Services DevCon on March 21-22 in Beaverton, OR (10 minutes west of Portland).