Gah! TCP doesn't won't work the way I want :(

Well, I've been having more fun messing around with Node.js and allowing myself to be distracted by interesting problems. The latest of which was triggered by my desire to integrate the BrowserStack beta API for cross browser testing. This is a nice service that will fire up any number of different versions of browsers and point them at a URL that you specify. Integrating this with Testacular and Mocha means that I can run all my browser javascript tests in all browser variants and get the results right back in my shell immediately, without having to run a myriad of browser versions locally :) This even includes mobile platforms :D

So what's the catch?

Well, in order for BrowserStack to connect to my Testacular server it needs to hit a public URL. Unfortunately my development machine is not reachable on a public URL (nor do I want it to be, at least not really public). The solution suggested by BrowserStack was to use a simple service called LocalTunnel. This service provides a client with which you can create an SSH tunnel to a local port that you specify. The service then allocates a random subdomain of localtunnel.com from which it will forward HTTP requests to your local port. Very useful and sounds easy, right? Unfortunately when I tried the client it didn't work and the only clues were leading me into a world of SSH keys, etc.

Hence the distraction. As I probably want to fire up my tunnel and browsers programatically I'm not so fond of relying on command line interfaces and really I want a node module to do it. What's more if I'm going to dig around in secure connections why don't I take the opportunity to expand my knowledge in a direction that I want it expanded. So I decided I would implement my own tunnel service and client solution in node and thus the tls-tunnel package was born.

Early on I figured I didn't want to mess about with generating random sub domains and trying to route based on the sub domain on which a connection was made so instead I decided to assign ports on the server to satisfy client connections. This way whenever a new client connects and requests a tunnel the server will allocate a port from a predefined range of available ports and start listening on that.

My plan was to use a free Heroku or Nodejitsu instance to then deploy my tls-tunnel server when I needed it.

This is where I learnt a hard lesson in the problems of bottom up development. Although I am applying TDD principles I did in fact fail to validate one of my initial assumptions - that I could use multiple ports! Both Heroku and Nodejitsu will only expose one port to your application... this could/should have been a red flag. I realised this early on but plowed ahead anyway thinking that at a later date I could apply a small change to my tunnel and instead use the random subdomain solution to differentiate between tunnels.

So I got my tunnel working using TLS (hence the name) with clients and servers authenticating each other with their own self signed SSL certificates. I was pretty proud of myself for implementing something that was in theory protocol agnostic - I had noticed that other similar solutions were limited to HTTP traffic... this should have been a red flag!

I next turned to the problem of making it all work on one port. Having already learnt quite a bit about the TLS/SSL problem domain I now learned a hard lesson about the TCP domain or more specifically the Node.js net domain.

I had made the assumption that when a raw TCP socket was connected to a server I would be able to read out the domain name that it had used... Wrong!!!

What LocalTunnel is doing is using the HTTP protocol to get the domain name that was used for the connection. GAH!! and what do you know this is the same reason the Heroku and Nodejitsu limit access to a single port. Double GAH!!!

So now I'm left with a choice. My solution can still work but I'm going to have to put it on an Amazon EC2 instance or something (I can get one for free for now). Or I can bite the bullet and implement the same HTTP restriction (boo) and do subdomain based tunnelling.

It's not such a simple choice though. On the one hand it's easy to integrate Heroku and Nodejitsu into my development and testing process (and even share that) as opposed to the hoops I will have to jump through to get it up and running on an EC2 instance. But on the other I don't want to limit my solution to HTTP and I haven't actually verified yet that I can use random subdomains on either service (once bitten, etc).

Perhaps there is a third way though - maybe if I only support one tunnel at a time I can use a single port...

That said, I'm leaning towards the EC2 solution for flexibility ("lean"-ing might be a bad choice of word here though - if you'll excuse the pun ;))

Grunt watch and the Node.js require cache revisited

Still inspired by James Shore's series "Let's Code: Test-Driven Javascript",  I've been continuing with my endeavors to get the Grunt watch stuff working flawlessly(?).

In my last post I mentioned some niggles that were remaining from my previous workaround.

• The workaround only addresses the Mongoose issue
• The workaround assumes intimate knowledge of Mongoose
• Grunt watch still explodes silently when unhandled errors are encountered in tests
• undefined references
• nonexistent requires
• etc.

The good news is that I think I have addressed all of these. In addition to that I've figured out some stuff about how to extend grunt and how to manipulate the Node.js require cache.

First off I thought I'd take a look at Mocha to see if it handled things better. After all Mocha also has a watch function.

• Mocha watch does not explode on undefined references (which is nice)
• Mocha watch does still explode on nonexistent requires (actually I didn't find this out till much later on when integrating with grunt)
• Mocha watch still failed to handle my Mongoose issue
• Unfortunately Mocha watch doesn't integrate with JSHint and actually I'd quite like to lint my code on file changes too

So despite only having a small advantage in not falling over so much I thought Mocha showed more promise than NodeUnit and as James noted it is much more active on GitHub. In fact it's under the same banner as Express and Jade which are definitely very popular and well maintained frameworks for Node.js.

Next thing was to integrate Mocha with Grunt so that i can use the Grunt watch function to both lint and run tests on file changes.

The nice thing about writing my own task to run Mocha instead of NodeUnit is that it was then quite easy to fix the issue of exploding on nonexistent requires... It just needed a try/catch around the mocha.run call. In retrospect I could probably have added this to the existing NodeUnit task but by the time I got to this point, I'd already ported all my tests to Mocha.

[A short interlude on Mocha and Should...]

James noted in his videos that Mocha is targeted as a BDD test framework and as such he is not so keen on it's verbosity. I can see what he means but, to be honest, I don't find it much of an issue and in fact quite like it, so for a while at least, I think i'll stick with it.

I also tried the should.js assert library that provides an interesting take on asserts by making them a bit more natural language like. Things like: thing.should.have.property(things).with.length(5);

On first take I thought cool and went full steam ahead in making all my asserts like this. Currently though I'm not sure I like it.

For one, I keep thinking that I should be able to write something in a natural way but find that it's not really supported - it kinda feels like I'm being teased. This will lessen I guess as I really learn the idioms.

A more annoying problem though is related to the way Javascript handles types and comparisons. I keep finding comparisons that i think should work and don't and then comparisons that I think shouldn't work and do! I think this is made worse by hiding the comparisons inside assert functions. As a result I'm starting to come to the opinion that not only is the should framework more trouble than it's worth but in fact any assert framework that hides comparison logic is not such a good idea to use in tests in Javascript. This includes very standard things like: assert.equal(object1, object2);

I may revert to just a single check function that will better reflect how comparisons would actually be written in production code. Ie: assert(conditionalCodeThatResolvesToTrueOrFalse);

[...interlude over]

So there I have it, I can now run my tests as files change and rely on the watch task to keep going no matter what happens (so far!). Just the mongoose problems to resolve then, and actually I added another.

• If a unit test beforeEach function falls over then the after functions are not run
• This means that as I open a database connection in before and close it in after, when I get such an error I then continue to get failures when files change due to not being able to open the database anymore (it's already open)
• Not as serious as the silent failures as at least the watch process keeps pinging me and I can restart it. But still a little annoying

This new issue got me thinking again about the require cache. My previous investigations here had proven fruitless but then, perhaps I had been led astray by some dubious comments on StackOverflow. Beware, this code does not work:

for (var key in Object.keys(require.cache)) {delete require.cache[key];}

So now I was thinking about the Mongoose module.

• The problem isn't that the changed module is still in cache
• The problem is that the Mongoose module is still in cache
• In fact the problem is that any modules are still in cache
• I must clear the cache completely before running my tests!
• Actually I had tried this and it didn't seem to work
• However I had tried it in my tests themselves, now I could try it in my new grunt task :)
• I had already needed to add code that dropped all my own files from cache to make things work. It made sense to drop the rest too when I come to think about it.

So i fixed the code above:

for (var key in require.cache) {delete require.cache[key];}

Tidied up my mocha task adding support for options and this is what I have in a new module...

To use this I dropped it in a grunt tasks directory and updated my grunt.js file...

Note that the call to loadTasks takes the directory name. Also note that I overrode the built in NodeUnit test task and that the options to pass into mocha are given in the mocha config property.

So that's it I no longer have to use my Mongoose workaround as the Mongoose module is cleaned up along with everything else before I run the tests :)

I hope this will save me from similar gotchas in other modules too, but I guess I'll just have to code and find out :D

NodeUnit, Mongoose and Grunt watch

Edit: Although interesting to me as a history to my Node.js testing issues, this article is now pretty much superseded by this one which better addresses all of the below problems

Now that James Shore's "Test Driven Javascript" series has kicked off I've been integrating unit tests into the 5Live hangout project. This has, for the most part, been simpler than I expected. NodeUnit is pretty easy to use and from one of the comment threads I have been introduced to Grunt which allows me to tie all my lint tasks and unit tests into a single automated script (James has been doing this himself in Jake but I figured I would give Grunt a try as it does some of the 'grunt' work for me :)) .

Like I said, for the most part this has all been going swimmingly. One of the nice features of Grunt that I discovered is the watch task. This allows me to watch a list of files and when they change, automagically kick off my lint tasks and unit tests - very nice :D

There are some problems though. My application uses Mongoose to interact with a MongoDB database. As such I follow the standard mongoose pattern of using a singleton and defining my model schemas like this...



As I'm doing TDD I actually start off with something like this in a separate test file...



That's all hunkydory. I can run the tests and they pass. I can kick off grunt watch and leave it running while I start editing my files. Let's see what happens when I change my test, thusly...



As expected grunt watch pings me to let me know that my test has failed :)

So I go back to my model and update the greeting function...



Gah, grunt watch pings me again to say that my test still fails. Puzzlement abounds!

If I stop grunt watch and run the tests manually they pass! So what's going on?

Well I wasted a lot of time messing around with the require.cache object, as I figured it was something to do with Node.js module caching, but that wasn't it at all. Either NodeUnit or Grunt is smart enough to remove the changed files from the module cache (I think it must be Grunt that does this but I didn't check).

Eventually I realised that it was the mongoose singleton that was causing the problem. After all this only happened with my mongoose model tests. As the mongoose singleton persists between test runs it doesn't matter that I change the methods on my models, the old versions also persist.

Again I tried a number of workarounds but so far the best seems to be the following.

First I created a wrapper for the mongoose singleton which allows me to reset the schemas...



Next I integrated this wrapper into my tests (only the tests, i still use the the mongoose singleton directly in the model implementations)...



So why do I prefer this solution and what else did I try?

Well I also had another workaround which fixed the problems with method updates.

• Instead of using Schema.methods to assign methods I used Model.prototype
• Instead of using Schema.statics to assign static methods I just assigned them to the Model directly

Why didn't I like this?

• This solution meant a little rejigging of the code to what seemed like a non standard pattern in the actual implementations
• This did not fix a similar problem with updating the actual Schema - ie. adding or removing fields

I still don't much like my eventual workaround as...

• it depends on knowledge of the the internals of Mongoose (which might change)

But at least it's contained in my tests and seems to work for all changes in the model.

However, even with this workaround in place I'm still  not fully happy with the way grunt watch works.

• Annoyingly, it exits on a number of test failures particularly when things are not yet defined. This happens a lot when doing TDD, it's how we write failing tests.
• When it does exit it doesn't actually ping me. As such I have to keep looking to see if it stopped (If i have to do this all the time, it occurs to me that I may as well not use it and instead run my tests manually)
• I'm now just waiting for the next gotcha as I only worked around a problem with Mongoose.
• It seems to me quite likely that there will be other libraries that follow similar internal patterns and they are likely to trip me up in the same way

I have a solution to suggest though...

• Spawn a new Node.js process at least for every grunt watch event if not for every NodeUnit test
• Wouldn't this fix the problem once and for all?

Required watching!

http://jamesshore.com/Blog/Lets-Play/

Well it might be if it weren't 197 episodes and counting but this series of videos illustrating test driven development (TDD) in Java and Eclipse are really excellent. I'm currently up to episode 25 (since yesterday... addicted!) and already I have learnt so much about:

• writing tests up front
• emergent design
• AND just how magical eclipse is when it comes to generating code and refactoring in Java

I really recommend watching some of these. It starts with a completely blank application so I would go from there.

I really want some of these IDE features in the OCode development we do too. And I can see that there may be some mileage in that.

Now i can't wait for James' new series that I gladly helped fund on Kickstarter.

http://www.letscodejavascript.com/

2 things that have got me excited recently in one package - TDD and server side Javascript... Yay, Node.js!