Of all the exciting stuff that’s happening at Google, one of the things I’ve been most excited about is SPDY, Mike Belshe and Roberto Peon’s new protocol that upgrades HTTP to deal with many of the new use-cases that have strained browsers and web servers in the last couple of years.
There are some obvious advantages to SPDY; header compression means that things like cookies get gzipped, not just content, and mutliplexing over a single connection with priority information will allow clients and servers to cooperate to accelerate page layout based on what’s important, not only what got requested first.
But the the really interesting stuff from my perspective is the way SPDY enables server push both for anticipated content and for handling Comet-style workloads. The first bit is likely to have the largest impact for the largest set of apps. Instead of trying to do things like embed images in data: URLs — which punishes content by making it uncacheable — SPDY allows the server to anticipate that the client will need some resource and preemptively begin sending it without changing the HTTP-level semantics of the request/response. The result is that even for non-cached requests, many fewer full round trips are required when servers are savvy about what the client will be asking for. Another way to think about it is that it allows the server to help pre-fill an empty cache. Application servers like RoR and Django can know enough about what resources a page is likely to require to begin sending them preemptively in a SPDY-enabled environment. The results in terms of how we optimize apps are nothing short of stunning. Today we work hard to tell browsers as early as possible that they’ll need some chunk of CSS (per Steve’s findings) and try to structure our JavaScript so that it starts up late in the game because the penalty for waiting on JS is so severe (blocked rendering, etc.). At the very edge of the envelope, this often means inlining CSS and accepting the penalty of not being able to cache for things that should likely be reusable across pages. On most sites, the next page looks a lot like the previous one, after all. When implemented well, SPDY will buy us a way out of this conundrum.
And then there are the implications for Comet workloads. First, SPDY multiplexes. One socket, many requests. Statefully. By default. Awwwww yeah. That means that a client that wants to hang on to an HTTP connection (long polling, “hanging GET”, <term of the week here>) isn’t penalized at the server since SPDY servers are expected to be handling stateful, long-lived connections. At an architectural level, SPDY forces the issue. No one will be fielding a SPDY server that doesn’t handle Comet workloads out of the box because it’ll often be harder to do so than not. SPDY finally brings servers into architectural alignment with how many clients want to use them.
Beyond that, SPDY allows clients to set priority information, meaning that real-time information that’s likely to be small in size can take precedence on the wire over a large image request. Similarly, because it multiplexes, SPDY could be used as an encapsulation format for WebSockets, allowing one TCP socket to service multiple WebSockets. The efficiency gains here are pretty obvious: less TCP overhead and lowered potential for unintentional DoS (think portals with tons of widgets all making socket requests). There’s going to need to be some further discussion about how to make new ideas like WebSockets work over SPDY, but the direction is both clear and promising. SPDY should enable a faster web both now and in the future.
Some Orthodox Heresies
Just back from a long weekend, a couple of things struck me this morning that are evident but not perhaps obvious. Reality always has multiple sides, so in true Jenny Holzer style, please accept (but not entirely) these small web-development truisms: