TCP-based RPC server (Erlang or something similar?) for iOS/Android app communication

Using TCP sockets with your own protocol rolled down is quite better than HTTP especially with the nature of resources on the mobile devices. Erlang will do quite well, however lets start from your protocol. Erlang excels well at this especially with the Bit Syntax expressions. However still, you could use plain text as you wish. JSON (would need a parser: Mochijson2.erl found in Mochiweb library) and XML (will need a parser: Erlsom).

I have personally worked on a project in which we were using raw TCP Sockets with our Erlang Servers and Mobile Devices. However, depending on the Port numbers you choose, Routers along the way would block/Drop packets depending on the security policies of service providers. However, i still think that HTTP can work. People chat on Facebook Mobile, send Twits e.t.c from their devices and am sure these social engines use some kind of Long Polling or Server Push or whatever but using HTTP. The mobile devices have advanced in capability of late.

Rolling your own TCP Based protocol comes with a number of challenges: Port selection, Parsing of data both at the client and server, Security issues e.t.c. Using HTTP will let you think of the actual problem than spending time correcting protocol issues at client or server. The Devices you've mentioned above like Android and IOS (Ipad, Iphone e.t.c) are very capable of handling HTTP COMET (Long polling). Am sure when you follow the standards for Web Applications on Mobile devices as well as these W3C Mobile Web Best Practices, your app will function well using HTTP.

Using HTTP methods will quicken the work and there are a lot of libraries on the SDKs of these Devices which would assist you prototype the solution you want as compared to the situation of rolling your own TCP-based plain text protocol. To back up this reasoning, look through these W3C findings.

Let me finally talk of the HTTP benefits on these Devices. If you are to use Web technologies for Mobile devices, such as Opera Widgets, Phone Gap, Sencha Touch, and JQuery Mobile, their SDKs and Libraries have Optimizations already done for you or have well documented ways in which your app can be made efficient. Further still, these technologies have the APIs to access the native Devices' resources like Battery check, SMS, MMS, GSM broadcast channels, Contacts, Lighting, GPS , and Memory; all as APIs in the JavaScript classes. It would become hard (inflexible) if you use native programming languages like J2ME, Mobile Python or Symbian C++ / Qt as compared to using Web technologies like CSS3, HTML5 and JavaScript tools mentioned above. Using the Web tools mentioned above will make your app easily distributable by say Ovi Store or Apple Store, from experience.

Take note that if you use HTTP, testing will be easy. All you need is a public Domain so the Widgets on the mobile device locates your servers over the Internet. If you role your own TCP/IP protocol, the Network Routers may be disruptive against the Port number you use unless you plan on using port 80 or another well known port, but then still your Server IP would have to be made Public. There is a short cut to this: if you put your TCP Server behind the same ISP as your testing Mobile's Internet connection, the ISP routers will see both source and destination as behind its Network. But all in all, there are challenges with rolling your own protocol.

Edit: Using HTTP, you will benefit from REST. Web Servers implemented in Erlang (especially Yaws and Mochiweb) excel at REST services. Look at this article: RESTFUL services with Yaws. For mochiweb, there is an interesting article about: A million User comet application using Mochiweb which is broken into 3 parts. Further still, you could look at the solution given to this question.

There are ZeroMQ builds for android and iOS. Java and ObjC bindings exist as well.

HTTP was created for infrequent requests with large responses. It is highly inefficient for transferring very big amounts of small data chunks. In typical situation, http headers can be twice in size of actual payload. The only strong side of HTTP is its habitualness, its 'One size fits all' karma.

If you want lightweight and fast solution, I guess ZeroMQ can be a perfect solution.

One reason to go with HTTP instead of a custom service is that it's widely supported on a transport level.

With mobile devices, a user might be on Wi-Fi at a hotel, airport, coffee shop, or corporate LAN. In some cases this means having to connect via proxy. Your application's users will be happiest if the application is able to use the device's proxy settings to connect. This provides the least surprise -- if web browsing works, then the application should work also.

HTTP is simple enough that it isn't difficult to write a server that will accept HTTP requests from a custom client. If you decide to go this route, the best solution is the one that you don't have to support. If you can write something in Erlang that is supportive of application changes, then it sounds like a reasonable solution. If you're not comfortable doing so then PHP or J2EE gets bonus points for the availability of cheap labor.

While HTTP does benefit from being widely supported, some successful projects are based on other protocols. The Sipdroid developers found that persistent TCP connections do greatly improve battery life. Their article on the topic doesn't address the server side but it does give a high-level description of their approach on the client.