SPDY Review by Opera Software

During my interview with that team, one of the kids proudly took credit for the span tag. A weekend hack. Committed the code and shipped it. No review. SOP.

Being more standards-minded at the time, I wanted to throttle the kid.

With the benefit of hindsight, I see that all the jitter and experimentation was pretty much ideal. If a feature proves useful, other browsers may adopt it, and it may become a de facto standard.

First, HTTP pipelining exists.

Second, HTTP's inefficiencies establish a bounds on your performance. HTTP itself does not define your application's performance, unless you've already mined every optimization opportunity at the levels above. Do you ensure all your CSS and JavaScript assets are bundled into single files and minified? If the answer is "no," you have no business complaining about HTTP's performance—managing your assets better will improve performance far more than using SPDY alone will. What about your caching story? What about your database, is it indexed properly? These are tangible improvements that you can make in your application today that can and will improve performance for the end user.

At my work, there is a CGI app written in C that does some atmospheric calculations against a database. The old database was Oracle, the new one is PostgreSQL. The CGI now takes twice as long to run as it did before, which is noticeable because it used to take about 5 seconds, so now it's taking 10, for one day of calculations. Looking into the problem, we found a place where an index would be a good thing to add. So we added it, and the performance didn't change at all.

It turns out, the app is sending a separate SQL query for individual data points. The app needs hundreds of thousands of data points to do its work, so it's sending many thousands of queries on each request. One query can fetch all of the data this program needs in about 0.3 seconds. So this app is, in effect, not measuring the performance of the database itself at all; instead it's measuring the overhead of running an instantaneous query on Oracle and Postgres, multiplied by several thousand. It turns out that the overhead of setting up and performing a query is about twice as high with Postgres. This fact never matters in practice though, because when you notice it you're doing something stupid. Nobody ever says "you should switch to Oracle, because that Postgres's queries take 4 milliseconds to setup instead of 2." (The fact that this app is written in C and could easily be outperformed by a shell script is also telling.)

This is exactly why HTTP is in no sense incomplete or narrowly defined. Google is the one with the narrowing requirements: they know exactly how much money HTTP's inefficiencies are costing them and are in a position to throw engineering time and energy at that number to decrease it. They are also in a position to optimize every other corner of their stack, and presumably they have. This is not true anywhere else.

SPDY is somewhat beneficial for consumers. But it undermines the simplicity and clarity of HTTP. That's what makes this late-game technology. SPDY is SOAP and CORBA to HTTP's RPC. Is it a better definition? Probably. But it's also harder and the benefits are insignificant except at scale.

It's a bit like redesigning a car from scratch just because you need snow chains for certain roads, and upholstery covers for certain passengers.

The two descriptions you make below describe much more essentially than superficially or "astronautically", the same kind of thing, optimized for slightly different use cases (not even THAT different).