I feel like arguing about plain mix projects and poncho or umbrella style projects is the wrong level.

In elixir you basically have two kinds of “structure”. Structure of code – which should be handled by modules and folders in your source code – and structure of runtime concerns, like dependencies or being able to run parts of your application independently from others – which is what drives how you set up supervision trees or multiple otp applications.

What’s often criticized with umbrellas is using the latter solution (otp applications) for the first reasoning. Don’t just put code in multiple otp applications if those parts of code are not meant to be run in isolation. A single otp application is just fine for that.

On the other hand if you intend to split up your application at runtime e.g. over multiple nodes then go for properly devided otp applications.

As for poncho vs. umbrella. Poncho style project setups are way more explicit especially when it comes to configuration. With umbrellas people often seem to have problems with the fact that config (just like dependencies) are set for the whole umbrella. With elixir 1.9 this is even enforced by the generators, which no longer generate application level config in umbrellas. For the nerves people where configuration is even more complex because of target machine differences the solution was to just not use umbrellas. Personally I feel like both are usable and you need to find what style of thinking about/working with your application suits you more. Either “one piece at a time” (poncho) or “the whole as the sum of all parts” (umbrella).

IMHO there is almost no difference between these two approaches since 1.9 (and earlier the only difference in reality was handling the configuration). Right now I would go with Umbrellas even more, as it provides what I want almost directly OOTB and with Poncho I need to manage everything by myself or use external tools.

If you have incompatibilities you’ll have them with the poncho style project as well. At runtime things are run exactly the same. In umbrellas you’re just forced to spot those issues as you’ll always maintain just one set of configs / run apps together more often, while for poncho applications those things might be discovered later e.g. when switching to a “higher level” application. Both have their place though and it’s a tradeoff between more flexibility and tighter control.

Personally I feel like poncho style applications make the most sense in loosely coupled projects like nerves. There might be an application for talking to a special sensor and it just needs to be included to be called from elsewhere, but there are no other dependencies between applications. If things are more coupled, but e.g. in separate apps to split load or failures predictable between different nodes I’d rather opt for an umbrella.