Around 6 PM yesterday, our SMIL file delivery broke. At first I figured it for a database connection problem, but the log entries were atypical. I soon determined that retrieval of the SMIL DTD was regularly failing. Most requests would get an error, and those that did succeed took over a minute.
There’s a basic flaw in XML DTD’s and schemas (collectively called grammars). They’re identified by a URL, and by default any parser that validates documents by their grammar retrieves it from that URL. For popular ones, that means a lot of traffic. We’ve run into that problem with the JHOVE configuration schema, and that’s nowhere near the traffic a really popular schema must generate.
Knowing this, and also knowing that depending on an outside website’s staying up is a bad idea, we’ve made our own local copy of the SMIL DTD to reference. So I was extremely puzzled about why access to it had become so terrible. After much headscratching, I discovered a bug in the code that kept the redirection to the local DTD from working; we had been going to the official URL, which lives on w3.org, all along.
Presumably W3C is constantly hammered by requests for grammars which it originates, and presumably it’s fighting back by greatly lowering the priority of the worst offenders. Its server wasn’t blocking the requests altogether; that would have been easier to diagnose. The priority just got so low that most requests timed out.
Once I figured that out, I put in the fix to access the local DTD URL, and things are looking nicer now. Moving the fix to production will take a couple of days but should be routine.
The problem is inherent in XML: The definition of grammars is tied to a specific Web location. Aside from the problem of heavy traffic to there, this means the longevity of the grammar is tied to the longevity of the URL. It takes extra effort to make a local copy, and anyone starting out isn’t likely to encounter throttling right away, so the law of least effort says most people won’t bother to.
This got me wondering, as I started writing this post, why don’t parsers like Xerces cache grammars? It turns out that Xerces can cache grammars, though by default it doesn’t. As far as I can tell, this isn’t a well-known feature, and again the law of least effort implies that a lot of developers won’t take advantage of it. But it looks like a very useful thing. It should really be enabled by default, though I can understand why its implementers took the more cautious approach.