C++ – For nodejs what are best design practices for native modules which share dependencies

I think the critical piece you're missing is that the result of require("foo") is always the same object. Consider this REPL example:

> var myHttpModule = require("http")
{ ... }
> myHttpModule.someNewProperty = "someValue"
'someValue'
> require("http").someNewProperty
'someValue'

The second call to require("http") did not re-create the http module -- it summoned up the only http module that exists in Node's current execution environment, which had been altered on the previous line.

So, in your example, there's only one Alpha module. When Beta and Gamma call require("alpha"), they're each getting the same reference to the one-and-only Alpha module singleton object.

(Or, to be perfectly precise, they're getting the same reference to the exports value of the Alpha module. require creates a Module object, but returns only the exports property of that module. If you don't fully understand the distinction right now, it's not a big deal.)

Behind the scenes: what's really going on with require caching?

The first time you use require to include a module, Node actually runs the code in that module. The resulting module is stored in require.cache. Subsequent attempts to load the module with require first check for an already-loaded module object in require.cache. (Note: environment-native built-in Node modules like http are exceptions in that they don't use require.cache -- you'll need to test out my code below with a custom module.)

require.cache is an object whose keys are module file path, with associated values that are module objects. For example, assume some module named "foo" in C:\node_modules\foo.js:

> require.cache                             // empty cache
{ } 
> require("foo")                            // require foo
{ bar: 'baz' }
> require.cache                             // cache now populated
{ 'C:\\node_modules\\foo.js':
   { id: 'C: \\node_modules\\foo.js',
     exports: { bar: 'baz' },
     parent: { ... },
     filename: 'C:\\node_modules\\foo.js',
     ...
     paths:
      [ ... ] } }

The current value of the foo module is in require.cache["C:\node_modules\foo.js"].exports. We can use require.resolve to get the file path of the module from the name, so we can express it also as require.cache[require.resolve("foo")].exports.

If we call require("foo") a second time, Node sees that require.cache[require.resolve("foo")] is defined, and so it returns the value of require.cache[require.resolve("foo")].exports instead of re-running the module creation code. This exports property of the foo module in require.cache is the single instance of the foo module export value, returned with every subsequent call to require("foo").

One interesting implication here is that you can delete require.cache[require.resolve("foo")] to force a reload of the foo module with the next call to require("foo"), because the object describing the module is removed from the require.cache object.

So what does this mean for me?

Without require, you could still share values between modules using global variables. For example, your Alpha/Beta/Gamma case would work just as well with just Beta and Gamma setting and reading the global value global.a. Instead, with Node's require system, you're actually setting and reading global.require.cache[require.resolve("alpha")].exports.a, which Node lets you read neatly as just require("alpha").a.

In fact, if you just want to share a value and don't need to import your shared data as a module, you could also use a namespacing object, i.e., have Beta and Gamma set and read properties of the object global.alpha. The major advantage to using require is that you don't need to set up require("alpha") external to the other scripts, whereas you would need to define global.alpha = {} before requireing Beta and Gamma. (Alternatively, you could conditionally define global.alpha in each module based on a typeof global.alpha == 'undefined' check instead, to see if it's the first module to use it.)

Treat Web services the same as any other data access:

To approach client side MVC architecture, without frameworks, we treated Web services in the same manner as we did localStorage and the local IndexedDB database. All code that involves requests to remote servers or that query a database or that read/write to localStorage happen in the model layer as DAO objects.

These were written using either the prototype pattern, the module pattern, or in some cases a combination of the two, using a modified factory pattern to treat objects as singletons in production but allowing us to "reset" them when unit testing so we keep tests self-contained.

They're generalized so that no application logic exists in this layer, and most of the code at this level can essentially be reused across future applications.

The services layer encapsulates the DAO layer, keeping the specific storage technology separate from the controller logic. Thus, pushing data to a remote server is seen as the same as inserting data into IndexedDB.

Used raw XMLHttpRequest to interface with Web services:

We chose not to use jQuery AJAX. Instead, we wrote a wrapper around XMLHttpRequest. There's no right or wrong answer here, but choosing not to use jQuery here allows us to stay focused on the rule of having no DOM manipulations in this layer. By wrapping the AJAX logic inside a prototype class, we set certain default headers, since most of what Web services deals with in our case is application/json data.

However, we did make an exception for jQuery Deferred. It didn't seem to make sense to write our own Observer implementation simply to be purists, and our logic behind this decision is that future professional programmers who work on this project with us have a better chance of understanding $.Deferred than some hand-rolled implementation. Many developers, on the other hand, we argued would have been exposed to XMLHttpRequest at some point in time, so being a purist here seemed less risky from the perspective of communicating to other developers what a piece of code does.

Analysis:

Earlier, I mentioned rules. We wrote up a series of "rules" designed to keep the code maintainable. For instance, rule #1 is that jQuery DOM manipulations only happen in a "view" layer. So the only jQuery we use in other layers is the $.Deferred object and nothing else.

We use $.Deferred throughout the application to keep certain logic in the layers where that belongs. For instance, we keep a persistent $.Deferred observer in the DOM click handlers in order to "notify" the controller that a click event has occurred so the controller can then delegate to other layers of the application to get/set, fetch/store, or perform other actions.

Using TDD as a development methodology, we've created small functions, where the largest is never more than a page, and most layers have a great deal of test coverage.

After a few months, the jury is still out on whether we should have used a framework, as much of what we've done could be generalized further into a reusable framework. We've learned quite a bit trying this, and I'm hoping this experience helps others who wish to develop client-side applications that involve a lot of asynchronous operations.

To give credit where it's due, I took inspiration from the following people:

• Ben Nadel's article on A Better Understanding of MVC Model View Controller
• My colleague Varun Vairavan, who has written some tools for AngularJS and node-webkit, and who was heavily involved in many of these technical decisions based on his experiences with Backbone.js and AngularJS.
• Joel Spolsky's Law of Leaky Abstractions, which we personally dealt with when using a specific JavaScript framework.
• Ben Nadel's Using jQuery Deferred To Create Compound Objects From Multiple Asynchronous Data Sources

For further reading on going framework-less, see Joe Gregorio's Zero Framework Manifesto