While fixing a bug yesterday, I plowed through the code that does Gnip’s HTTP response code special case handling. The scenarios we’re handling illustrate the complexities around doing integrations with many web APIs. It was a reminder of how much we all want standards to work, and how often they only partially do so. Here are a few nuances you should consider if you’re doing API integrations by hand.
When doing a polling based integration with a “real-time” API, you’re inclined to poll it a lot. That has caused some service providers to tell you to slow down using the “retry-after” HTTP header. Some providers use other, not so standard, ways to cool you down, but those are beyond the scope of this post. When you get a non-200-level response back from a server, you should consider looking for the retry-after header, regardless of whether or not it was a 503 or 300-level code (per HTTP 1.1 specification). Generally, when a services sends a retry-after, they’re intention behind it is clear, and you should respect the value that comes back. Now, the format of that value can be either “seconds”, or in a more verbose time format that tells you when you should wait “until” before trying the request again. In practice, we’ve never seen the latter; only the “seconds” version. When we see retry-after, we sleep that duration; you should probably do the same.
HTTP Response-code ’999′
You can look for it in the spec, but you won’t find it. Delicious likes to send a ’999′ back when you’re hitting them too hard. Consider backing off for several minutes if you see this from them.
non-200 HTTP Response Bodies
While many services don’t bother sending response bodies back for non-200s (and those that do often don’t provide anything actionable), many do. It’s a good idea to write those bodies to a log file (or at least the first n-hundred bytes) for human inspection. There can be some useful information in there to help you build a more effective and efficient integration.
The matrix of services-to-response codes, and how you should respond to them, is big. The above is just a small slice of the scenarios your integrations will encounter, and that you’ll need to solve for.
While a service’s documentation is always some degree out of date, and you can only truly learn the behavioral characteristics through long nights of debugging, here are some pointers to service specific response codes that you might find useful.