[iagooar]

Could you elaborate on that? How would one accept the tradeoff of having one more piece of software to manage vs performance or other considerations? I've heard cowboy does a pretty solid job serving requests, even static files (although not as good as Nginx).

[brightball]

That's basically what it boils down too. Cowboy is more of a pain to configure update and manage so it's less of a choice of managing another piece of software vs choosing which one you want to manage. With Phoenix using Cowboy under the hood you'll never actually touch it.

At that point it just becomes a matter of "do I want to configure nginx or cowboy?"

If you have SSL at the load balancer and you're using a CDN for assets then it's less compelling. If you're looking to deploy a full encapsulated binary (server and all) then avoiding nginx makes sense.

That's basically it though.

[klibertp]

I disagree. I worked with both Nginx and Cowboy for quite some time, and I don't think that Cowboy is harder to "configure, update and manage." I had opposite impression.

I like Nginx very much, by the way - I was very happy when I switched from Apache to Nginx and I never looked back. Nginx is still my default choice for serving static files, as a frontend for Python (and similar) applications, as a (reverse) proxy, etc. However, I don't feel the need for it when working with Erlang or Elixir, exactly because of Cowboy.

Cowboy is not as fast, but it still is quite performant, even for static files handling. I don't have the numbers, and I can't seem to find a relevant comparison (there's one against Node.js and Tornadoweb here: http://www.ostinelli.net/a-comparison-between-misultin-mochi...) right now, unfortunately.

However, upgrading and management of Cowboy are either on the same level as Nginx or simpler. The upgrade process has two steps: change version in mixfile and re-fetch deps, done. I know that it's one step more than plain apt-get, but it's not that complicated, either.

As for management, Cowboy has an advantage that you can configure it in a Turing complete language (with clear semantics - recall if or allow/deny semantics in Nginx config), not only by initial configuration but also by writing handlers, where you can do arbitrary processing. It's also runtime configurable and hot-swappable (you can reload Nginx config, you can't (AFAIK) replace the currently running nginx with newer version "in place", you need to restart it).

On the other hand, Cowboy has much less built-in functionalities, which is why I said it's not good for example for being a plain proxy - you'd need to write your own handler module for that, while in Nginx it's a one proxy_pass directive and it will be faster to boot. However, when you have an Elixir app and you want to just serve some static assets in addition to the app itself I think it's completely fine to rely on Cowboy for this.

Of course, the advantage of Turing-complete language disappears if you use OpenResty. It allows you to write your app inside Nginx in Lua. I used it in a few cases and it was quite pleasant experience, however, then the Elixir backend becomes unnecessary...

So, to summarize: I think that you don't need Nginx if you use Erlang/Elixir for your backend and that you gain almost nothing if you do include it.