Test Driven Debugging

I hate debugging. Loathe it, despise it, pick your synonym and I’ll use it. Computer programming is building something out of nothing, making the computer do things it didn’t know how to do, until I told it.

I absolutely love code refactoring, moving code around, organizing it better, cleaning it up and watching tests pass. But stepping through code looking for the place where I accidentally typed question instead of question? Shoot me in the head.

In addition to the frustration of debugging, a bug can only be explained by one thing: I screwed up. I hate screwing up even more than I hate debugging. My hatred of debugging is one of the reasons I started practicing TDD. After all TDD means I never debug right?

Of course we all debug some of the time, usually as a deadline hangs over our heads. This tends to be where practices break down. You hack and slash, desperately trying to make the program just work only to watch it fail and fail again.

Eventually you get the bug fixed, probably with an ugly hack right before (or right after) a customer demo. I won’t claim that “Test Driven Debugging” is a solution, but it certainly helps me in times of need.

Let me take an example from real life. As our last iteration ended I was on a roll. I had a to do list of items and everything kept clicking into place. I’d implement a feature, and it work the way I wanted to the first time. There was just one last part of the story, moving the items on the screen.

I wrote a test, it passed. I wrote another, made it pass. I got everything working the way I wanted too—then tried the actual application, and it went boom. Moving just wasn’t working the way I expected. Items would jump around the screen seemingly at random.

I realized what the bug was-items retrieved from the database were not coming back in the right order. So how do I go about fixing it?

Well I don’t always practice what I preach, so I started hacking and slashing. I threw order_by and find statements around, and even wrote a method that wasn’t tested in a desperate effort to make the demo.

Of course I didn’t make the demo, and then sat down and calmly did what I should have done. I wrote some tests. Well first I wept quietly, but that’s none of your business.

In our application there’s an ActiveRecord called Design, and it has Widgets which use acts_as_list to keep track of themselves within the Design with a position. Don’t worry about the details if you aren’t familiar with Ruby on Rails, the important thing is the steps.

First I wrote a test in the Design.

 1 it "should order widgets by position" do
 2   widget1 = Admin::Widget.create(:design => @design)
 3   widget2 = Admin::Widget.create(:design => @design)
 4   widget3 = Admin::Widget.create(:design => @design)
 5 
 6   widget3.insert_at(0)
 7   widget3.save
 8   @design.reload
 9 
10   @design.widgets[0].id.should eql(widget3.id)
11   @design.widgets[1].id.should eql(widget1.id)
12   @design.widgets[2].id.should eql(widget2.id)
13   widget3.position.should eql(0)
14   widget1.position.should eql(1)
15   widget2.position.should eql(2)
16 end 

Hmmm…it passed. That’s interesting. It’s coming back correct from the model. I guess it’s time to write a test in the controller:

 1 it "should have the widgets ordered by position - requires real data" do
 2   question = Admin::Question.create(:name => "question")
 3   design = Admin::Design.create
 4   widget_pos2 = Admin::Widget.create(:question => question, :design => design)
 5   widget_pos1 = Admin::Widget.create(:question => question, :design => design)
 6   widget_pos2.position.should eql(1)
 7   widget_pos1.position.should eql(2)    
 8 
 9   widget_pos2.insert_at 2
10   widget_pos2.save!
11   widget_pos1.reload
12   widget_pos2.reload
13 
14   widget_pos2.position.should eql(2)
15   widget_pos1.position.should eql(1)
16 
17   get :edit, :id => design.id
18 
19   assigns[:design].should be_instance_of(Admin::Design)
20   assigns[:design].widgets[0].position.should eql(1)
21   assigns[:design].widgets[1].position.should eql(2)
22 end

There’s a couple things worth noting in this test. I used real data instead of mocks, although RSpec style would typically use mocks.

The reason I didn’t use mocks is that while they are great at decoupling data just like this, in the case of a bug I could very well pass my mock, only to discover that I’ve been calling the wrong methods all the time.

I want the real data here. Some would argue this doesn’t make it a Unit Test, and I’d be hard pressed to disagree, but it still is a valuable test. The second is that I’m doing should eql before I call get, which is the method under test.

The reason for that is I want to test my assumptions. Often times when you have a bug it’s because you made a mistake in your assumptions, such as thinking the widgets were always being returned in position order, and I want to tease out any issues in mine.

The last thing I need is a test that passes by coincidence.

So I ran the test-and it passed! Then I started the application again, and it worked! Did tests have a magical power that fixed the bug? Of course not. I had fixed the bug during my hacking and slashing, but I was frantic so I hadn’t calmly restarted my server and had a passing test.

Had I done this from the get go I would have recognized the bug, which was using an explicit design.widgets.find(:all) instead of design.widgets, thereby retrieving the data from the database in id order, and likely passed the demo.

Once again lesson learned, sometimes I have to learn more than once. I’ll clean up the test above to remove the now unneeded extra checking, but I’ll leave it in the code, so that this bug never shows itself again. Thou art commanded—never reintroduce a bug. Now go forth and program my son.

Eric Smith, Software Craftsman

Eric Smith has a Master's Degree in Video Game Development from DePaul University.