Inspecting changes locally before pushing
If you work on your branch you run into the situation that you would like to push your changes to the remote repository. CI will then pick up your changes and run the linting and code quality checks on it. Afterwards, you will see whether you improved the quality. But perhaps there are some new violations that crept into the code? Happens to all of us!
I usually like to see and check if any new issues might come up on CI. This lets me improve them before I push. The checks I run locally depend on the kind of work that I do. Lately that’s a lot of Ruby on Rails again — which is great. I love that framework and the language. To grade my code, I use Rubocop, Reek and Flay. If you run their respective commands on your repository, they will check the whole code base. This might be ok, if you didn’t have any issues before. Since I join teams, these days, that work on legacy projects it is rare that there are no problems with the code. If I run the commands just so, I will get a long list and couldn’t possibly see the issues that I introduced through my changes. Lucikly, there is Git and some “command line foo” that can help us here:
git fetch && git diff-tree -r --no-commit-id --name-only master@\{u\} head | xargs ls -1 2>/dev/null | grep '\.rb$' | xargs rubocop
This command will fetch the current state from the remote and diff your branch/changes to master branch. It then runs rubocop on these changes.
In my ~/.aliases.local file I added three lines for all three linters.
# Code Quality
alias rubocop-changes="git fetch && git diff-tree -r --no-commit-id --name-only master@\{u\} head | xargs ls -1 2>/dev/null | grep '\.rb$' | xargs rubocop"
alias reek-changes="git fetch && git diff-tree -r --no-commit-id --name-only master@\{u\} head | xargs ls -1 2>/dev/null | grep '\.rb$' | xargs reek"
alias flay-changes="git fetch && git diff-tree -r --no-commit-id --name-only master@\{u\} head | xargs ls -1 2>/dev/null | grep '\.rb$' | xargs flay"
I am still working on a way to just call one command and have all thee commands run. That doesn’t yet work. Probably because of exit-code reasons, when one linter finds issues.
These simple commands offer a convenient way to find local issues and correct them before pushing to CI.
Avdi Grimm's view on deleting tests
A few weeks ago I wrote about deleting your tests. Yesterday I received the weekly email from Avdi Grimm, where he touches on this subject.
Some premises about my relationship with unit testing:
I like test-driven development. I like driving out individual object design small, isolated (e.g. no database) unit tests. I think of these unit tests as a design aid, full stop. Any help they provide with preventing regressions is gravy. I treat unit tests as disposable. Once they have served their purpose as design aids, I will only keep them around so long as they aren’t getting in the way. These four premises are strongly interconnected. Take away #1 (test first), and tests are no longer a design aid and none of the other points are valid. Take away #2 (isolation) and we get into integration test territory where there’s a much higher possibility of tests having regression-prevention value. Take away #3 (design focus) and the other points lose their justification. Take away #4 (disposability) and I spend all my time updating tests broken code changes.
This makes it easy for me to find myself at cross purposes with others in discussions about unit testing, because often they come into the conversation not sharing my premises. For instance, if your focus in unit testing is on preventing regressions, you might well decide isolated unit tests are more trouble than they are worth. I can’t really argue with that.
A recent conversation on this topic inspired the thought from the driving-design perspective, maybe unit tests are really just a crutch for languages that don’t have a sufficiently strong REPL experience. While I don’t think this is strictly true, I think the perspective shines a useful light on what we’re really trying to accomplish with unit tests. I almost think that what I really want out of unit tests is the ability to fiddle with live objects in a REPL until they do what I want, and then save that REPL session directly to a test for convenience in flagging API changes.
That conversation also spawned the idea of immutable unit tests: unit tests that can only be deleted, never updated. A bit like TCR. I wonder if this would place some helpful incentives on the test-writing process.
You should subscribe to his newsletter, if you haven’t yet.
My submission to Euruko was rejected. ☹ That’s a pity, would have loved to visit Rotterdam. I still have 7 other proposals that might come through. Fingers crossed 🤞
External forces
I am occupied with learning these days. Learning on my own about visualizations of data among other topics. But also learning about learning. For that I read what other people think about learning. There are many things I have to learn about this whole topic. One thought I saw repeatedly, was about external forces, or limiting factors.
Let me elaborate what I mean by that: There are people that can motivate themselves more easily than others can. They reach their goals or at least try very hard. Others give up more easily when they face some resistance. As always, there are people in the middle between these extremes. You know best which group you belong to. 💪
What has this do with software quality? I am getting there… 😉
I am wondering how external forces could help improve quality. If you need to reach your goal and you don’t belong to the group of highly self-motivated people there are options like hiring a coach. Athletes do that all the time. I pay for a “virtual” coach that guides my running efforts.
How could you hire a “virtual” coach for your coding efforts, for reaching your targets on your software quality metrics? You could hire me or other “real” coaches, of course. But that doesn’t scale too well and might be too expensive.
Again, for some people it is easy enough to use static analysis or linting — a kind of coach in it’s own right — and follow their guidelines. Yet, still there are people that ignore the warnings or guidelines imposed upon them by the tools. Reasons may be a hard deadline or too much workload. How could we offer external forces, limiting factors that help them, guide them, towards doing the right thing?
One solution I can think of is to have a robot not accept your code when it is below standard or ignores guidelines. A robot could be anything that measures and grades your code and reports back to your team. Some tools already offer this, for example GitLab. If you want to merge code that decreases the overal quality metrics, you are not allowed to do so. So that would be one.
Another idea: If you try to commit or merge such code, you need to consult with another developer about the code. Once you worked on it together, the other dev has to enter her secret key, to remove the lock on the merge. This forces you to pair on code more often.
When it comes to teaching there is this saying of the “glass has to be empty (enough).” You cannot pour water into it, when it’s already filled. Said ideas 👆probably won’t work for a team that isn’t aiming for learning and improving.
I will continue to think.
You probably know AC/DC. But have you heard of MC/DC?
If you need to test a complex conditional, you should take some time to learn about this one.
Today I am working on a new visualization of fragmentation of ownership for Git repositories. The more fragmented a repo is, the more developers work on it together. This leads to no clear ownership and makes it easier for defects to creep into the code.
Complex conditionals
The other day we dealt with code coverage and gnarly conditionals. I promised to offer a way to be able to test them properly.
THERE IS NONE.
Ha, what a bad joke. But the real answer might not be better, depending on your point of view. What you have to do is create a table.
(A || B) && C
This is our conditional.
| m | A | B | C | RESULT |
----------------------
| 0 | T | T | T | T |
| 1 | T | F | T | T | x
| 2 | F | F | T | F | x
| 3 | F | T | T | T | x
| 4 | T | T | F | F |
| 5 | T | F | F | F | x
| 6 | F | T | F | F |
| 7 | F | F | F | F |
# m is the test case
# A, B, C are the atomic parts of the conditional
# RESULT is the result of the evaluation of the conditional
For three terms in a conditional, you can have 8 different cases (2^3). You don’t need to test every case. You have to find those cases where switching one term (A, B or C) changes the RESULT. You take those cases and write tests for them. You can ignore the rest as they don’t bring you any new information. For our example these could be the test cases 1, 2, 3, 4. I marked them with an x,
The general rule of thumb is that you can solve this with n + 1 test cases where n is the number of terms.
This technique is called Modified Condition/Decision Coverage or short MC/DC. I like this name, it’s easy to remember 🤘.
It gets harder to do, when one term of the conditional is used more than once (coupled). Another thing to take note of is that depending on the decision statement in the code, it may not be possible to vary the value of the coupled term such that it alone causes the decision outcome to change. You can deal with this by only testing uncoupled atomic decisions. Or you analyse every case where coupling occurs one-by-one. Then you know which ones to use.
You’ll have to do this in the aerospace software industry or where you deal with safety-critical systems.
If you read this far: Congratulations. This is one of the more in-depth topics of testing software. You deserve a 👏 for learning about these things!
I am reading the new edition of Refactoring by Martin Fowler these days. I am preparing a new workshop on Software Design and Architecture. I bet there are some ideas in there that will inspire me to create some specific exercises. #refactoring
I am now using Micro.blog for my microposts. I will crosspost my newsletters there. And whatever I think makes sense. I will figure it out by just trying. 😎
Code coverage can be misleading
During the last week, I had two discussions about code coverage. Code coverage is the metric of how many lines of code are covered by your automated test suite. Many test frameworks have built-in ways to measure this. Other times you have to install another tool manually. When you run your tests you then see how many lines are not covered by a test. That means that no test was run where this line of code was evaluated or executed or interpreted.
When you reach 100% code coverage, what then? Are you done? Could you guarantee that there are absolutely no bugs in your code?
If you are tempted to say Yes, or “maybe?” then let me tell you that you are wrong.
Consider this piece of code.
If you write a unit test for this method, the line eval... will be interpreted because of the if emergency at the end. The line is thus covered.
But the code is not covered or tested.
Admittedly, this is a very trivial example that I made up. In reality, there are some more profound things to consider.
If you have complex conditionals you might need a logic table where you compare all possible combinations of the atomic parts of the conditional.
You cannot possibly evaluate this in your head and know whether you checked for every possible, sensible combination. Yet when you cover that line you are at 100% coverage and can go home, right?
So what do you do? Let’s look at this tomorrow.