As I’ve mentioned before, I’ve been spending some time at work with a search engine. It works sort of like NewEgg’s search in that you might type a string in (“desktop ram”) and then narrow your search by clicking on links in a side bar (DDR3 or Corsair). The search engine in question powers a product called SubjectLin.es (it’s in beta; you can sign up for a free account to try it out, if you like) where you can search for words that appear in an email subject and then further limit by, say, who sent it (“sender”), what kind of email it is (like coupon or receipt, we call it a “tag”) or what business the sender was in (“market”).

To date, I’ve rewritten the search engine a few times and refactored it a few times. It and one other area of the code base are, by far, the most complex bits of the application, so it has gotten it’s fair share of attention. And it’s gotten more recently when an exception brought our attention to something: We don’t have one search engine… we have four. It’s just that they’re all in one class. Finding the email subjects that a user is interested in is not really related to finding the senders who sent them except that the user’s parameters are the same. So I thought what we’d do is factor out our current search object into 4 smaller ones (subject, sender, tag and market) that each centers around the object it’s pulling out of the DB. Besides code clarity, there are some other benefits we get from this modularity, but I don’t want to get into that in this post.

So now we’re faced with the fact that we’ll have 4 objects we want to create that all depend on the same parameters to build themselves. Add to that the fact that we do some fancy Google-style keyword parsing (you can type “some key words sender:Company” and we pull out the company name) and the fact that if the user is interested in limiting by sender, we want to display things differently than if they aren’t. Now, I thought, it looked as if we should have an object just about the user’s search parameters and that each of our search engines should take only that as an argument.

This way, our view (or presenter… whatever) can ask the search parameter object if the user limited based on sender or not. We can build the one object and then pass it to each of the four engines. We can put all the keyword parsing logic into the search parameter object, too. When I had it, it seemed like such a good idea that it had to have been thought of before. And, lo, it is called the Parameter Object pattern. Or, anyway, I think this is a case of that pattern.

What do you think? Is this Parameter Object? Is it something else? Does my example make any sense at all (I honestly have no idea, since I’m so deep in the system; I tried to balance high-level with need-to-know)? Note: as of this writing, none of this is implemented yet. This is all just design work with @marcosacosta, so far. If anything interesting crops up while we’re writing it, I’ll try to remember to update this post.

One thing in particular sort of caught at the edge of my thought patterns: The Null Object Pattern. I don’t have a CS degree and so I’m missing a lot of the formal training about design patterns that many programmers have (and, probably, forget), so I’d never heard of it. When Avdi started talking about how ActiveRecord’s try method is a code smell, I was like, “Yes!” I would not say that I hate it, but I have seen several times some line of code that looks like this:

1

@user.try(:posts).try(:recent).try(:first)

I mean… bleh. But I didn’t know a way that looked any better to me, really. Anyway, you can look at the notes and slides to learn about what Avdi says about the Null Object Pattern. I thought it was awesome and so when I went home, I decided to take it for a test drive.

The Test Course

So in a project at work, we have Users and they may or may not have one Subscription. Hopefully, you can picture this complex object graph. Subscriptions may or may not be “current” based on various business rules mostly to do with whether you paid us or not. So, naturally, we have a Subscrption#current? method. But we’re using the User as a sort of presenter for Subscriptions. So you don’t want to call @user.subscription.current?. That’s a code smell. So on User we had this method:

1
2
3

def current?
  subscription.try(:current?)
end

There’s that rascal try. “This,” I thought, “is a perfect spot for that Maybe method from Avdi’s talk.” So I rewrote it thusly:

1
2
3

def current?
  Maybe(subscription).current?
end

W00t, right? Wrong. The accompanying NullObject class looks like this:

1
2
3
4
5
6
7
8
9

class NullObject
  def method_missing(*args, &block)
    self
  end
 
  def nil?
    true
  end
end

That method missing treatment, so handy in avoiding the chain of trys in my first example, is the gotcha. It means that if I have a User without a Subscription for whatever reason, calling User#current? returns an instance of NullObject, which will pass, say, the boolean clause of an if statement.

So, not sure as to whether I’d misunderstood something, was making some dumb mistake or what, I emailed Avdi. He said, basically, “Awesome question. I will answer it in a blog post.” And, lo, he did. Go read that post to see what he said. The comments also have some good ideas.

Noodles

If you read my comment, I said I was going to noodle on stuff and post again. I had more thoughts than it seemed like would fit in a blog comment. Hence this post. So my initial thought was disappointment. It turned out the Null Object Pattern wasn’t as powerful (in Ruby) as I’d hoped, since if you might have something (calling Maybe) the chances that you’ll have some conditional asking a boolean business-rule question about it is not low.

So I thought about how to get around that. You could, for instance, make a more complex method_missing definition that grepped the message name for /\?$/ and returned false. That’s fail, though. It falls down the moment you have something like this:

1
2
3

if Maybe(@posts).empty?
  # Intuitively, you'd expect NullObject#empty? to have put you in here.
end

But then I realized that Avdi was making a higher-level point: since it is not possible to make your own objects look falsey in Ruby, you have to have another solution. Trying to define various question-mark methods on NullObject is trying to untie the knot, but I should be looking for a way to cut it. So it got me thinking: Why the hell to I have Users without Subscriptions, anyway? Shouldn’t User#current? express that business logic clearly, rather than just express the logic that enforces it? Yes. Yes, it should.

We have some Users who are also admins, who have special rights. It’s also conceivable that we could give away a free account for whatever reason. So, really, we want something like this:

1
2
3

def current?
  self.free_account? || subscription.current?
end

But, this thought it incomplete. It expresses the business logic cleanly: The User is current if they’re flagged as free or if their Subscription is up to date. However, if the weird case of a User who is neither free nor has an associated Subscription crops up, we still have to hunt down the “Undefined method ‘current?’ for nil” error. It sort of has be reaching for Maybe again.

Or maybe (heh, you see what I did there) I want to steal another trick from Avdi’s presentation and have User#subscription return :no_subscription_defined_for_user so that the error message makes some more sense. I don’t like redefining ActiveRecord‘s default accessor methods, though, to transparently return the symbol if the real object is missing.

If you’ve got any thoughts, I’d love to hear them.

The thing that sucked so bad was the API, basically. I didn’t really know how I wanted to use it or how best to fit it into a Rails app. I made a reasonable guess and got to work. It wasn’t a terrible first try, but it was a terrible API. However, I’m not sure what better looks like yet. And I think that I sort of got ahead of myself; I put the cart before the horse.

It’s a little like making the transition from drawing on paper to creating art with the GIMP. When you’re on paper, you draw the stuff that’s in front first and you only draw as much of the stuff that’s behind other stuff as can be seen. So you learn to think a certain way about how you build up your picture. In a program that has layers, you can draw in any old order and draw something entirely even if it’s obscured by something else in the long run.

I was thinking that the obvious path would be to develop my little gem of functionality and then use it in the larger Hey Go Vote application. Now I’m starting to think that’s backwards. I’m going to just work on Hey Go Vote and trust that doing so will give me insight into what parts of the announcement machinery can be made portable. Then I’ll extract those bits into a gem and refactor Hey Go Vote to use the new gem. I’m sure I’ll let you know how that goes.

In the run line (Windows key, then R or click Start menu then Run…) put in runas /u:administrator "explorer.exe /separate". You’ll want to replace administrator with an appropriate user name if that’s not a valid one. A DOS prompt will appear asking for the password and away you go. This tip thanks to Stack Overflow.

I also found something else handy here: You can input run-line commands into the “location” prompt when creating a shortcut from scratch in Windows. So if you don’t want to type out all that /u:administrator stuff all day (or, well, probably not that frequently and you’d forget it), then you can right-click > New > shortcut and paste your command into the location. Call it whatever you want and then you’ve got a shortcut to an admin Windows Explorer right on your desktop.

I find this so much easier to deal with than any other solution when I need to muck with file permissions or any number of things in Windows. Helpful? Didn’t work for you? Thoughts on the site design? Let me know in the comments.