My understanding is that since React has a virtual DOM, it can just render everything on the server with Node.JS, instead of evaluating it in a browser.
the downside to his reasoning is that he is tying himself to a single back end (node) if he used something like mustache he coulkd use any backend, but mustache implies parsing templates on the client. googles closure compiles to javascript functions and I am sure something like that exists for mustache and this would seem very sensical.
Converting a large string to a dom element is not expensive, it is when that dom element is added to the actual display tree when things go a bit awol, this is to be expected, redrawing any GUI costs a lot.
React is an excellent library and I am glad people use it, none of the ideas are radical however.
fwiw, React isn't tied to Node: I've used it with Django via PyExecJS (which just uses raw JavaScriptCore).
this technique is called pre-rendering, and yes, I am amazed people don't do it.
"PyExecJS (which just uses raw JavaScriptCore)." this seems pretty slow or can you cache the execution function. (aka you don't have to parse the js multiple times)
Our backend is jvm so we decided on closure, with this on startup we compile our templates on the server and then pass the rendering functions around pumping the data into it to render initial pages. At the same time, we compile the same templates into javascript functions which are included into our app via the closure dependency system to render on the client. This has the advantage of not having to parse and "compile" a template at the point of delivery.
This is why I think react is an excellent library, because it allows anybody to do this and brings this notion into the minds of developers who jump easily on bandwagons.
I also think that the react team are well on their way to building a library which would fair excellently under all sorts of static analysis
We are actively working on more static analysis tools for React: it's certainly one of our major priorities.