A coworker and I were discussing this yesterday. Our Rails processes take about 50MB of memory each, and can handle only one request. Since our site hits a backend DB that is sometimes slow to return results, requests can last minutes. We could build an out-of-band process to query the backend service and return results asynchronously, but that seems like overkill. However, it's easy for a handful of users to eat up all the connections quickly, if they are all making long-running requests.

It seems silly to have an entire Rails process tied up blocked on results from this backend service. And it seems like a rather severe limitation of Rails' architecture to require 50MB of memory per connection-handling process. How do large sites scale Rails?