Ezra's Research

So yesterday I went down to Amazon's development office in Edinburgh, and gave a talk on Links, at the invitation of Andrew Birkett. The talk went well and I was quite pleased with the response and the sharp questions.

One of the biggest concerns that the Amazon people had, which we haven't tried to address in Links, is failure: how the language allows them to handle failure. In Amazon's systems, there are loads of interdependent services that depend on each other to display a single page, and any of them might go wrong; the developers described to me a wide variety of alternate responses that they might want to give, depending on the situations. For example, in some instances, a page feature might collapse down to nothing (disappearing from the page) when it fails. Other times, if a service doesn't respond, the front-end web software might use cached data to display a feature.

This came up in regard to Links because of the way we're making client-server (Ajax) calls simple RPC calls. The question there is, what happens if the remote call doesn't finish successfully, either because the server code couldn't get the right data, or because the network mechanisms themselves failed; how in Links can we handle that anomaly? Part of the answer might be an ordinary exception mechanism, which we can support fairly easily, but we should think more about how Links programmers deal with exceptional situations.

The crowd was very friendly and engaged with the talk quite deeply, I think. They fed me plenty of questions on particular points. Many of these had to do with various kinds of possible failure, as I mentioned; another theme was metaprogramming, prompted because they noticed a certain amount of repetition in my slides (e.g. (foo=foo, bar=bar) when constructing records; I do hope we can improve on that).

I gather they [present-day Amazon developers] do most of their work in Java, but they weren't thrown off when I started talking about continuations or any of the functional-programming idioms that we use in Links. There were a few self-confessed language afficionados in the crowd, but they weren't the only ones who didn't miss a beat at the functional idioms, the tuple types, or the unusual kinds of variable binding we use for our "formlets" (oh, reader—you don't know about our formlets—I'll have to post about that soon).

Between the talk itself and chatting with developers outside of it, I had a quite nice time. Thanks to Andrew and all of the others for listening so well & treating me nicely, even though I'm an academic!

UPDATE: Slides from the talk are available.