Today, I want to talk about code metrics. Taking the time to critically analyze your code can provide a number of wonderful benefits. It provides a viewpoint through which you can identify problem areas in your code base. It can give you confidence that you are reaching your goals and confirm suspicions when you are floundering. However, in my experience the best benefit of using code metrics is gaining insight into you and your team’s coding style.
Now that we’ve established some of the benefits of code metrics, let’s talk about some of the pitfalls. Unfortunately, like so many other things in the software development world, code metrics are not the only solution. Measuring productivity by any code metric will result in programmers coding to the metric. For example, if that metric is test coverage, they’ll write a single test that runs an assertion against every file. It may not be as blatant as that, but if you structure any incentives around these metrics, they will be gamed.
Another common pitfall is to assume that these metrics actually mean something. Metrics can provide you with insight, but to make that insight actionable, you’ll have to take thoughtful steps to ensure that there is in fact a problem. Once you’ve decided that there is a problem, you’ll need to see if that problem is worth fixing. One last thing, these are never fire and forget tools. They require cogitation, time, and effort in order to be useful.
In this article I’ll walk through three of the most common code metrics. We’ll first talk about cyclomatic complexity, then about the ABC metric, finally we’ll move on to test coverage. For each of these, I’ll explain what they are in as much detail as this format allows and walk through tools you can use immediately to test your own code. Alright, here we go…
Cyclomatic complexity is a graphical measurement of the number of possible paths through the normal flow of a program. For example, a program with no branching statements has a score of one. Calculation of the cyclomatic complexity of a method is done through the use of a control flow graph. These graphs consist of nodes and edges. Each node in a program represents a basic statement and each edge represents the changes in control flow of the program. This all sounds very difficult to grasp, so let’s look at an example.
Given this piece of Ruby:
def cyclomatic_complexity_of_two(arg) if arg # Node A puts "Foo" # Node B else puts "Bar" # Node C end end # Node D
We’ll begin to construct our graph. We first need to figure out how many nodes, edges, and endpoints we have. We only have one endpoint, at the end of the method. Our
if arg line represents a node, as do our
puts "Foo" and
puts "Bar" lines, and our endpoint. Finally our edges are from:
- A to B
- A to C
- B to D
- C to D
- And our endpoint back to our entrypoint (virtual edge)
We end with four nodes, five edges, and one endpoint. Once we’ve done all that, we find the cyclomatic complexity by taking
Edges - Nodes + Endpoints. A plot of the above method looks like this:
If you want to know more about cyclomatic complexity read this paper by Arthur Watson and Thomas McCabe.
The advantages of keeping this number low are clear. A high cyclomatic complexity signals that the code may be difficult to follow. It also tends to mean that the code is harder to test and debug. Keeping this score low, can ensure that you and your team will be better able to understand what’s happening and make smart decisions.
Getting started with CC
Our first stop in Ruby land is Saikuro. It’s a program used to measure the cyclomatic complexity of your program. You can download it here, and then run
ruby setup.rb all. Alternatively, you could install by:
gem install atoulme-Saikuro, which is a clone of the main Saikuro repository. I find the latter a little more convenient (and not as old school).
Once you’ve done that you can run your first Saikuro report with the following command. Since it generates an html (or text) report, I reccommend making a directory for the output and sending it there. Otherwise, the default is to place the reports into your current directory.
~$ mkdir ./saikuro_output ~$ saikuro -c -o saikuro_output -p euclid_equation.rb ~$ open ./saikuro_output/index_cyclo.html
To view the options run
saikuro -h, the above command tells Saikuro to create a report of the:
- Cyclomatic complexity (
- Output directory (
- And parse (
-p) this file.
Inside the html file you’ll see a report giving you the details of the parsed file.
Saikuro measures complexity, but as I’m sure you’re aware, Ruby is a tricky language. Saikuro adds a branch when it encounters (if, unless, while, until, for, elsif, when), but it also adds a branch if the code uses a block. The rationalization here is that using a block inside your code very often changes the control flow.
For example, invoking the “each” method of an array with a block would only traverse the given block if the array is not empty.
Saikuro also ignores the operators
and, though in a perfect world these would add to the cyclomatic complexity.
You can have it report recursively on a dir or target a troublesome class. If you want to take more control over your usage of Saikuro, you can use the gem version and:
So there you have it, a quick reporting tool that can help you understand where you have excess complexity in your projects. In a moment, we’ll see how this is somewhat similar to another project that is great for code evaluation – Flog.
Luckily there are a few papers on the ABC metric that have been searched enough to actually make this google-able. ::grumble:: Anyway, the ABC metric is another way to measure complexity, though its originator vociferously claims that it is a measurement of the size of the program, similar to LoC. Either way, it is certainly a method for better understanding where your code may have issues.
The ABC method is a method by which scores are gathered for each component- assignment, branching, and calls- then calculated into a score that can be easily understood. The formula is
and is rounded to the nearest tenth. The score is kept in vector format and looks like
<a,b,c> until you want the total, then it is calculated into scalar format.
The ABC metric is a neat measurement because it let’s you sum up a lot of information in a way that is very easily understood. Since this information is easily understood, it’s easy to use it for a whole code base. If you understand how these numbers are forumlated, you can use them to cast a really bright spotlight on problem areas in your code base.
Let’s move straight into the Getting Started section because we get to use Flog.
Getting Started with ABC
Alright, so the tool we are going to talk about is Flog. It’s a tool that has been in widespread use for a long time, so in all likelihood you’ve heard of it or have even potentially used it. As such, I won’t spend too much time on installation, just
gem install flog and you’re off to the races.
The first thing about Flog that you need to know is that it’s opinionated. Very opinionated. The scoring is custom built to apply commonly accepted design patterns for Ruby. In this way, we need to think of Flog as using a modified ABC metric. For example, using inject humorously (if a little snarkily) nets you two points.
Branching terms are
:and, :case, :else, :if, :or, :rescue, :until, :when, :while. There are other rules that add to the branch total, but this is a good start. Assignment is pretty simple, add one to the score per assignment. Calls are any method call that takes the flow out of the current scope.
In a talk Jake Scruggs (Author of metric_fu) gave back in 2008, he neatly summarized the rule of thumb that I like to remember when using Flog:
- 0-10 – Awesome. But does it do anything?
- 11-21 – Good.
- 21-40 – Might need to refactor.
- 41-60 – You could possibly justify the code. (i.e. cost of refactor prevents action)
- 61-? – Danger, investigate when possible.
Alright, with that let’s use it. Given the following Ruby:
# euclids_equation.rb def gcd(x, y) while y != 0 r = x % y x = y y = r end x end
We can run Flog against this code with
flog ./euclids_equation.rb. And we get the output:
As you can see this method receives a 4.4. The default summary also shows a method average; if there were more methods, this would reflect the average complexity of each method. Flog has many options including parsing support for both 1.8 and 1.9 syntax. As you can see, the code executes quite speedily which makes it perfect for a githook.
A pre-commit hook that checks code complexity and returns non-zero for any score higher than 61 works quite well. And since you can always bypass with
--no-verify, it won’t prevent you from your normal workflow if you don’t want to bother.
Now that you’ve got your footing with Flog, let’s talk about coverage.
Code coverage is a measure used in software testing. It describes the degree to which the source code of a program has been tested.
This seems a simple concept, but has surprising depth. Also, the terms used to describe the types of coverage are difficult to research and disambiguate. Before we go much further, I’ll describe a few terms that you’ll be expected to know if you want to dive into measuring code coverage in your projects.
- C1 coverage – This is basic statement coverage (also what simplecov uses). One of the most basic measurements. Simply measures that each statement is executed (or at least loaded) through tests.
- C2 coverage – This is branch coverage. Measures that there is a test for each edge of the flow.
There are a number of other types of coverage measurements that are useful. If you are curious, I’d reccommend reading this paper on the various types of coverage. We’ll primarily be focusing on C1 and C2 with
simplecov, so I won’t belabour the point anymore. But know that there are many ways to define test coverage.
The advantages of tracking code test coverage is pretty obvious. You’ll be able to look at which areas in your code base are the most secure, and which are in most need of attention.
A word of warning: Code coverage as mentioned earlier, can be gamed with relative ease. Be sure to use this only to give you an idea of where you stand and don’t build incentives around it. As always with code metrics, use thoughtful discretion when trying to gain actionable insight into your programs.
Getting Started with Coverage
As briefly mentioned before, we’ll be using
simplecov. Simplecov is a gem that can be easily integrated into a Rails app or other Ruby project, and provides a snazzy little html page that shows your test coverage.
Installation is a little bit more intense than the previous tools listed here, but is still pretty straightforward. Add the following line to your test/spec_helper and run your tests.
require 'simplecov' SimpleCov.start
coverage/index.html and you’ll see something that looks like this:
Image from simplecov’s github page.
To see a highlighted version where red shows the lines not executed, you can click the file paths. And that’s basically that.
C1 (statement coverage) is what simplecov reports, and while you can’t use this information to quantify exactly how healthy your program is, you can use this overview to give you a place to start. This is especially useful for legacy applications, where you may not know yet exactly where the problem areas are.
We looked at these:
- Cyclomatic complexity- a graphical measurement of the number of possible paths through the normal flow of a program.
- Saikuro – Analyzes cyclomatic complexity and writes a nice report for evaluation.
- ABC metric- a method by which scores are gathered for each component- assignment, branching, and calls, then calculated into a score that can be easily understood.
- Flog – An opinionated and robust commandline tool that runs the ABC metric.
- Code Coverage- a measure used in software testing used to describe the degree to which the source code of a program has been tested.
- Simplecov – A nice test-coverage report tool that integrates easily into Rails and other Ruby projects.
Where to Begin
All of the tools discussed above are a good way to get started. If I were to start using code metrics today, I’d begin with Flog. There are good resources out there with information on how to get up and running with it, and I believe its opinionated nature will get you to start asking questions that you might not think about otherwise. Start by using the measurement Jake Scruggs (see above) came up with as a rule of thumb and really think about where that complexity is coming from. Once you’ve gotten a feel how Flog identifies overly complex code, move on to the other tools and begin to get a feel for how they relate.
Start by building a workflow that involves code metrics. It doesn’t have to be something as complicated as a githook even. Maybe just make it a habit to run Flog before merging a branch, or run a suite of metrics before your next code review.
Hopefully, this has whet your appetite for information about your code. Understanding how complexity is measured and where that complexity lives in your code is an incredible advantage for development. This article has barely even scratched the surface of code metrics. Really smart people have been working on this for decades and it shows. There are tremendous resources out there to explore all the different aspects of this study. Go and investigate. Oh, and thanks for reading. ^_^
Special thanks to @telemachus and @pat_shaugnessy for their editing help.
- Cyclomatic Complexity Revisited (infoq), Gavin Terrill, 2008
- “Structured Testing: A Testing Methodology Using the Cyclomatic Complexity Metric”, Arthur Watson,Thomas McCabe, 1996
- Saikuro Website,Zev Blut, 2005
- Applying the ABC Metric to C, C++, and Java, Jerry Fitzpatrick, 1997
- Using Metrics to Take a Look at Your Code, Jake Scruggs, 2008
- Wikipedia Code Coverage