[chipsy]

I think async architectures are worthy but miss something important - the idea of "typed connections" between processes. Actors passing arbitrary messages doesn't make for an efficient "assembly line." Similarly, the Unix-style concurrency is tied to a single low-level protocol, which is not rich enough to adequately describe all data. It described "enough" of all data for 1970's era tasks, but our needs outgrew that.

When you set out to architect a customized, typed, async architecture, you end up with the "flow-based programming" style which has been captivating me recently. It tends to reduce to two events per component: "start" and "stop." The "integrated program" appears in a tiny top-level definition that sets up the components and connections. Components are relatively small and reminiscent of "pure" algorithmic code. Where synchronized behavior becomes essential, flows can be split into stages of processing and kicked off in a sequence.

This particular style has had a lengthy history of reappearing in numerous domains under various guises, and it has demonstrable effectiveness, but it can also feel alien and more "mathematical." The main issue is that has a lengthy design/prototyping time of weeks to months, and the initial complexity of the system looks high because you need a decent number of components to do anything substantial. This really, really goes against the "move fast and break things" mainstream, even within open source - everyone wants their project to just _instantly_ accelerate from 0 to 100 in terms of progress, and we've put most of our efforts into a toolchain that makes it easier and easier to do that.

The best remedy at present seems to be to embrace asynchronity, embrace static types, and maintain faith in both - i.e. to have a lot of discipline.

[michaelochurch]

I regret that I have only one upvote to give this post.