When I wrote my book about technical debt, I decided to concentrate more on personal mindset, team dynamics, and organizational influence. I did this because I thought that the coding aspects of paying technical debt were well-covered, and that I could just refer to them.
Even with these resources, I had still been on teams that struggled with more foundational questions of whether code was debt at all, whether it was worth paying, how to explain it to our management, and how to prioritize and plan projects to do something about it. That’s what my book is about.
If you want to read some free chapters of Swimming in Tech Debt, sign up here:
Here’s my list of resources that are helpful when you need practical steps for addressing tech debt that you have identified and prioritized.
These resources cover Seams, Refactoring, Substitutions, Accretions, and Tidyings. A full toolkit of techniques to tackle debt.
#1: Working Effectively with Legacy Code [ad] by Michael Feathers. In this book, Feathers defines legacy code as code without tests. So it makes sense that the bulk of the book is the description of techniques for adding tests to code that wasn’t designed to be testable. Feathers introduces the concept of a seam, which is code you add to a function that makes it possible to control what that function does from an external source. The key point is that, in production, its behavior is completely unchanged.
Here’s a trivial example to illustrate a seam. It will seem like overkill, but the technique is useful in real-world code. Given a function like this:
def calculate_sales_tax(amount, zip_code):
rate = db_lookup_sale_tax_for_zip(zip_code)
return amount * rateIn a codebase without tests or database mocking, functions like this are risky to change. The goal is to get it under test before we refactor it, so let’s do it with a seam to learn the idea.
The first step is to change it to this:
def calculate_sales_tax(amount, zip_code):
rate = 0.08
if not TESTING:
rate = db_lookup_sale_tax_for_zip(zip_code)
return amount * rateBy adding the if not TESTING: we’re accomplishing two things:
- We have not changed the production code
- We have made the code testable
So, the next two steps are to add tests, and then to refactor (which is safer because the code is under test). To see the utility of seams, you have to imagine doing this systematically in the giant spaghetti functions you find in legacy codebases. Also, in the book, Feathers shows more sophisticated ways to do this than just an if statement.
The beginning of the book sets up the seam concept, and the bulk of the book is examples of how to do this with more realistic code. This is the best “How To” book for tackling technical debt in legacy projects that I have read, and its existence meant that, in my book, I could concentrate more on the questions of “When to” pay technical debt and “Who should” do it.
The point of a seam is to make refactoring safe. So, the next book on this list is the seminal work on refactoring.
#2: Refactoring: Improving the Design of Existing Code [ad] by Martin Fowler. The main contribution of Refactoring was to make the term (“refactoring”) ubiquitous. Unfortunately, even Fowler had to admit, the term is often misused. Whenever you go to daily update, and someone mentions refactoring, you could be sure that they are not doing what Fowler described.
To Fowler, a refactoring is a tiny act. His book catalogs about a hundred types. Examples include Renaming a Variable and Extracting a Variable. Each of these takes seconds. The more complex ones might take minutes. Hardly something you’d avoid or mention. Refactoring is something you are meant to do all of the time. Assuming, of course, that your code is under test.
You can freely refactor in code that is under test because the point of a refactoring is to improve the code without changing the behavior. It’s quick because the verification is automatic. Without that, safe refactoring would take longer because you need to write tests first.
So, enabling refactoring might take time. Instead of saying that you are refactoring, a more correct thing to note in your updates is that you are automating tests on code that you need to change, so that you don’t introduce regressions. In addition, you are reducing the chance of problems being found in review or QA. This is something that does take time, but also has value. If you label this activity “refactoring”, you risk being asked not to do it.
In my book, I talk about refactoring along with the more expansive idea of Substitution, which comes from Barbara Liskov and was formalized in her paper with Jeanette Wing.
#3: A Behavioral Notion of Subtyping by Barbara Liskov and Jeanette Wing. This paper is about how to design classes in an object-oriented language so that they can be subtyped safely (and how to do that subtyping). It’s related to the ideas in design by contract and opposes the notion of allowing for breaking changes. Liskov substitution is about replacing components in systems where you cannot (or will not) update the rest of the codebase to accommodate them.
It does this through subtyping, but I have always interpreted it as applying to almost any change I make to a system, not just class inheritance. In Object-Oriented programming, subtyping is one way to add new functionality to a codebase, but in my experience, it’s not the main way. In my current work, I don’t use OO or subtyping much at all. But, I substitute code all of the time.
In the paper, the authors describe a way to subtype that is safe to deploy because it does not violate any assumption in the system. This way of thinking applies to any change you might make to code. If you can, it’s better to change the code in a way that doesn’t invalidate all of your other code. Or, if you must, you know that you aren’t “substituting” any more, and there’s a lot more work to do.
Like a refactoring, a substitution should not break tests, but unlike a refactoring, substitutions do not aim to preserve all behavior. In fact, the entire point of a substitution is to add functionality. The paper describes how to do it safely.
To sum it up, your new code will be substitutable for your old code if (in addition to any language syntax requirements): (1) it does not add required pre-conditions to the caller, (2) does not break promised post-conditions, (3) does not violate any invariants, and (4) does not mutate anything assumed to be immutable.
The paper is formal and might not the best entry-point. But, the Liskov Substitution wikipedia article describes it well, and if you want to read more about how I apply substitution to any code change, I cover that in Chapter 5 of my book.
Another great entry-point on this subject is the video Spec-ulation by Rich Hickey
#4: Spec-ulation (video) by Rich Hickey. In Spec-ulation, Rich Hickey talks about the concept of accretion and how to use it to grow software. Accretion is related to substitution and refactoring and is another way to approach changing code (to pay off tech debt, for example).
Rich makes two interesting (and related) points about accretion. (1) like Liskov, he defines safe actions such as “doing more” and “requiring less”, which map to pre- and post-conditions (2) additionally, he argues that Clojure’s syntax makes many more accretions (i.e. substitutions) non-breaking.
In her definition of substitution, Liskov states that all compile-time, syntax, or language level errors introduced by new code are (by definition) non-substitutable. The paper is about “behavior” breaking substitution that is syntactically correct, but will fail at runtime.
For example, in C, if you add a new parameter to a function, that function is not substitutable because all callers need to be changed. In languages like Clojure, adding a parameter is not a syntax error because there is a default and callers don’t need to change. Rich’s talk is about the many syntax affordances in Clojure that allow future accretion (and is a warning about using new tools to subvert them).
He argues that language features like strong type checking, which offer compile-time correctness, limit software growth through accretion because changes that would be non-breaking (and safe) cause syntax errors. This can work well if you control all of the code and can update it. But it causes problems in public code libraries. Clients cannot upgrade their dependencies to get bugfixes and new features if that means their build will be broken for code that was working fine.
The takeaway here is to look at your language’s features and see some of these affordances as a way to pay off tech debt safely. For example: one, often overlooked, feature is namespaces. Using them, you can keep old and new (replacement) code running side-by-side until you are sure it’s ok to remove the old one.
I learned so much about growing code from Kent Beck in eXtreme Programming: Explained, but a more recent book from Beck is more specific to technical debt.
#5 Tidy First?: A Personal Exercise in Empirical Software Design [ad] by Kent Beck. This book builds on Beck’s previous books and the ones inspired by them (like Refactoring). In it, he introduces the idea of a tidying, which like a refactoring, aims to improve code without changing behavior. It is also meant as something to be done all of the time and to be quick and safe (because of tests).
Tidyings, if possible, are even smaller than refactorings. They are more about making code more readable now rather than making it easier to change (which refactorings do). Tidyings are things you can do to help you be a more active reader of code. For example, getting the reading order right and using whitespace better to chunk the code. Kent describes it as a way to get back into the original spirit of refactoring (tiny, safe changes that don’t alter behavior).
With seams, refactorings, substitutions, accretions, and tidyings, you can do a lot to improve a codebase and make it easier to change. This reduction of resistance to change is what I mean when I say to pay tech debt, and these resources detail the main ways I do it.
