Disclaimer – I work for Google but have nothing to do with the self-driving car team. The views expressed here are my own and do not necessarily represent that of my employer.
I found Eric Jaffe’s article on Google’s self-driving car and its handling of city streets immensely fascinating. He describes being a passenger in the self-driving car as it drives around Mountain View, CA (where I’m currently living) with minimal intervention from the Google engineer sitting in the driver’s seat.
It details some of the history of the self-driving car project, from its roots in DARPA’s Grand Challenge, a race through the desert with autonomous vehicles, through driving on the highway, to the vastly more complicated challenge of driving in an urban environment.
The article points out that Mountain View isn’t as crowded or chaotic as a city like New York or San Francisco, but it still has many interesting technological and algorithmic challenges, such as trying to intuit whether people at a crosswalk are really going to cross, or whether a bicyclist is going to get into the lane in front of the car.
I recommend reading the article to see how autonomous vehicles combine software engineering, robotics, machine learning, and computer vision in one sophisticated package.
This is hilarious (if you’re a programmer). Some folks trained a Markov chain on the class names in Spring and then made a game out of it – three Java class names are presented, only one is real. I didn’t do so hot – 4/11.
Robert Anning Bell [Public domain], via Wikimedia Commons
I’ve been programming professionally for five years. One of the things that I’ve learned is YAGNI, or “You aren’t gonna need it”.
It’s taken me a long time to learn the importance of this principle. When I was a senior in college, I had a course that involved programming the artificial intelligence (AI) of a real-time strategy game. For our final project, our team’s AI would be plugged in to fight against another team’s. I got hung up on implementing a complicated binary protocol for the robots on our team to communicate efficiently and effectively, and our team ended up doing terribly. I was mortified. No other team spent much time or effort on their communication protocol, and only after getting everything else up and running.
In this essay I’ll primarily be talking about producing code that’s not necessary now, but might be in the future. I call this “speculative design” and it’s what the YAGNI philosphy prevents.
First, let’s discuss how and why this speculative design happens. Then we’ll discuss the problems with giving into the temptation.
Why does it happen
I can only speak to my own experience. The times I’ve fallen into this trap can be classified into a few categories:
- It’s fun to build new features
- It feels proactive to anticipate needs
- Bad prioritization
Building features is fun
Programming is a creative outlet. It’s incredibly satisfying to have an idea, build it in code, and then see it in use. It’s more fun than other parts of development, like testing, refactoring, fixing bugs, and cleaning up dead code. These other tasks are incredibly important, but they’re ‘grungy’ and often go unrewarded. Implementing features is not only more fun, it get you more visibility and recognition.
Proactive to anticipate needs
A second reason one might engage in speculative design is to be proactive and anticipate the needs of the customer. If our requirements say that we must support XML export, it’s likely that we’ll end up having to support JSON in the future. We might as well get a head start on that feature so when it’s asked for we can delight the customer by delivering it in less time.
This is the case with the story I started this piece with. I overestimated the importance of inter-robot communications and overengineered it to a point where it hurt every other feature.
In this case, the feature was arguably necessary and should have been worked on, but not to the extent and not in the order that I did. In this case I should have used a strategy of satisficing and implemented the bare minimum after all of the more important things were done.
Why is it problematic
I’ve described a few reasons speculative code exists. You’ve already seen one example of why it’s problematic. I’ll detail some other reasons.
Let’s start simple. Time spent building out functionality that may be necessary in the future is time not spent on making things better today. As I mentioned at the start of this post, I ended up wasting hours and hours on something that ended up being completely irrelevant to the performance of teams in the competition, at the expense of things that mattered a lot more, like pathfinding.
Since there is more being developed, it’s likely that the overall software product is less focused. Your time and attention are being divided among more modules, including the speculatively designed ones.
Software complexity is often measured in lines of code; it’s not uncommon for large software projects to number in the millions. Windows XP, for instance, had about 45 million lines.
My point today is that, if we wish to count lines of code, we should not regard them as “lines produced” but as “lines spent”: the current conventional wisdom is so foolish as to book that count on the wrong side of the ledger.
I once equated lines of code produced to productivity, but nothing could be further from the truth. I now consider it a very good week if I decrease the lines of code in the system, by deleting chunks of code or rewriting them to be simpler and shorter.
The extra code and complexity associated with speculative coding is very expensive.
- It slows down readers of the code
- It slows down building the software (especially if it pulls in more dependencies)
- It adds additional tests that slow down the test suite
- It is likely to add more bugs (more code generally equals more bugs)
Sets unrealistic expectations
Say that you design a feature because you think that the customer is going to want. Imagine that you actually got it right – what they end up asking for is essentially identical to what you’ve implemented. You deliver it to the customer a full week before you promised it.
You might look like a hero, but this sets a very bad precedent. It sets unrealistic expectations as to how much work the feature took to implement, and might lead to the customer setting impossible deadlines for features of similar scope. If you were able to finish that feature early, they might reason, there’s no reason you shouldn’t produce the next feature just as quickly.
You’re probably a bad judge of what will be needed in the future
It’s hard enough to build software from detailed specifications and requirements. Guessing about what the specifications and requirements of a feature that isn’t needed yet is likely to end up with a product that doesn’t make anyone happy. It will likely match the designers’ mental model but not the users, since there was inadequate input from them.
It’s hard to remove features once they exist
Say that you’re designing the export feature of your software. You imagine there will be a whole lot of formats you want to support, but at the moment the only hard and fast requirement is CSV (comma separated value) format. As you’re writing the CSV export code, you see how it would be trivial to implement JSON encoding. And while you’re at it, you throw in XML. You were required to produce CSV but now you have JSON and XML support too. Great!
Well, maybe. Maybe not. A year down the line you notice that only a small percentage of your users export to XML, but the feature has led to a disproportionate number of support tickets. Now you’re in a tough place – if you kill the feature, you’ll irritate these power users. Furthermore, you will have effectively wasted all of the time in implementing the feature in the first place, and all the subsequent patches.
I have seen little-used features remain in production because they’re too much trouble to delete and alienate the few users of said feature. Which leads to…
Increased risk of dead code
Imagine that you’ve implemented a new feature but it’s not ready for prime time yet. Or maybe you used it once or twice but it’s not worth turning on for your normal service. You don’t want to kill the feature entirely, as it might have some utility down the line. (Warning bells should be going off about now) You decide to hide the feature behind a configuration flag that defaults to off. Great! The feature can easily be reenabled should you ever need it again.
There’s just one problem – it gets turned on accidentally interacts catastrophically with the rest of the system. Your software deals with financial transactions and it ends up costing your company 460 million dollars.
This sounds unlikely – except it’s true. This is essentially what happened to Knight Capital in 2012.
Knight also violated the requirements of Rule 15c3-5(b) because Knight did
not have technology governance controls and supervisory procedures
sufficient to ensure the orderly deployment of new code or to prevent the
activation of code no longer intended for use in Knight’s current operations
but left on its servers that were accessing the market; and Knight did not
have controls and supervisory procedures reasonably designed to guide
employees’ responses to significant technological and compliance
- It’s dangerous to leave dead code around in a system
- Speculative development is likely to lead to features that are not used often and are more likely to be dead code than if they were completely spec’ed out as in normal development
- Therefore speculative development puts you at a greater risk of dead code problems
Don’t allow dead code stay in the codebase. If you should ever need it again, you should be able to retrieve it from the version control system. You almost certainly won’t.
As an engineer, it’s easy to fall into the trap of implementing features before they’re actually needed. You’ll look productive and proactive. In the end, it’s best to avoid this temptation, for all of the problems I’ve mentioned. These include
- the extra code takes time to write, test, debug, and code review
- it contributes to a lack of conceptual focus in the system
- if done to please a customer, it sets unrealistic expectations for the development of other features
- it imparts an extra maintenance cost for the rest of the lifetime of said feature
- it will be difficult to remove the feature if and when its lack of use becomes apparent
- it puts you at increased risk of leaving dead code in the system, code which may later be accessed with bad consequences
I love Dijkstra’s notion of ‘lines spent’. Do you want to spend your time and lines of code on a speculative feature? Just remember – you aren’t gonna need it.