Re: those infrastructure components written in Erlang not attracting Erlang users in any permanent way—it’s not so much that they have more users than contributors, but rather just the bare fact of them being infrastructure components. A black-box infrastructure component can’t be a “killer app” for a language.
Rails is a killer-app for Ruby because it’s a framework; a developer who uses Rails in their project is inherently a Ruby developer. People learned Ruby to use Rails.
But you don’t need to write a single line of Erlang to use CouchDB or Riak or Ejabberd. They’re self-contained. (And, in fact, scriptable in not-Erlang languages; CouchDB for example spawns and manages a JavaScript engine to run its “design documents” on.)
Thus, these infra components becoming popular doesn’t cause there to be any more Erlang programmers than there were before.
In that sense, I would say that Erlang has never had a true “killer app” yet. (Outside of a few specialized domains, like telecom and fintech, where the OTP platform itself is Erlang’s killer app.)
And even in the places where Erlang itself is a “killer app”, it’s also not so much a viral thing that the whole industry gets excited about; rather, Erlang is more of a “secret sauce.” Nobody’s learning Erlang in their basement because their friend told them it’d be the next big thing in their industry.
How does anyone learn Erlang, then? Well, usually, engineers are forced to learn Erlang at one job (where the system architect there decided on it); they get to liking it; and then, when they later become a system architect themselves in another job, and choose a stack, they choose Erlang, thus forcing a new wave of engineers to learn it.