Writing code is often the best way to figure out what exactly needs to be done to solve a problem. In the Agile world, they call this a Spike.
A spike is when a software developer sits down and starts writing little bits of code with the goal of building a map of all the work they are going to need to eventually do to complete a project.
It’s a way of quickly hacking out code that does stuff like:
- Querying the database to see what’s in there and how it fits together
- Calling 3rd party APIs to see how they work
- Testing the algorithm that Fred said would work in Slack to see if he’s right
- Anything else that might need verified before committing to a design.
Spikes usually result in a bunch of messy code that kinda/sorta does most of the things that the product is going to need to do in the end. To most programmers, it feels good to do a spike. It gives more instant gratification than other methods of writing code. It makes me feel like an unstoppable force of nature, slicing through time and space, totally unaware of my colossal mediocrity. Because of that, there’s a temptation to feel that most of the work is done when the spike is done.
What comes next is often the developer saying something like, “I just need to clean up the code from the spike and open a pull request”. This creates the impression that there’s not much work left to do. The bad news is that that impression is often false. The good news is that I’m going to tell you how to deal with that bad news right now.
One of the main differences between doing a spike and writing production code is what one does when an edge case becomes clear.
In production code, I immediately write a test and/or alter the code to cover that scenario. In a spike, I just keep on rolling. At best, I’ll make a note to myself about it.
These build up over time. Three days worth of hacking out spike code can leave a lot more of these un-handled cases in the code then you might expect.
Code written during a spike often has much less thought about maintainability put into it.
Here’s an example from a spike I recently did that involved pulling all data about cohorts of students from an internal API.
def cohort_ids open = cohorts_api_client.search_cohorts(status: 'open', limit: 5000) active = cohorts_api_client.search_cohorts(status: 'active', limit: 5000) upcoming = cohorts_api_client.search_cohorts(status: 'upcoming', limit: 5000) (open + active + upcoming).map(&:cohort_id).map(&:to_i).sort end
We had significantly less than 5000 cohorts at the time. So for the purposes of my spike, setting the limit to 5000 and making one call was fine. But what would happen if we had more than 5000 cohorts with any one particular status? That code would miss them. The correct thing to do is to page through the results until there are none left.
The code that actually made it into production looked like this
BATCH\_SIZES = 100 def all %w(open active upcoming).map do |status| fetch(status: status) end.inject(&:+).uniq end protected def fetch(status:, limit: BATCH\_SIZES, offset: 0) batch = cohorts_api_client.search_cohorts(status: status, limit: limit, offset: offset) if batch.count == limit batch + fetch(status: status, limit: limit, offset: offset + limit) elsif batch.count < limit batch end end
That’s a pretty important refactor. It’s also exactly the type of thing that can be easily missed when a coder is rushing to get spike code into production. The spike version of the code will work fine for now. But some day in the future (probably at the worst possible moment), there will be a big influx of new cohorts and that original code would fall on its face. That would shatter the coder’s illusion of having any idea what they’re doing, not just in their work, but in their life. So lets avoid that.
Spike code is typically not driven by writing tests, or the type of tests used to drive it are not the best kind for producing good designs.
For the spike code listed above, I drove the development with one high level integration test. That single test ran all of the code in the spike end to end. Driving code with unit tests, which execute small portions of the code in isolation, typically leads to much better designs. This ends up being a major cause of the less-than-production-ready-ness of spike code.
Another great thing to think about is what the Pull Request is going to look like. In most cases, without some deliberate effort, it is going to be way too big.
I like to compare the branch with master (or develop) on Github so that I can see exactly what it would be like. That’s a good starting point for understanding that the spike code is in fact, not production ready.
I’ll spare you a rant about how I love emacs/org-mode for this sort of thing…
2. Look at each one and type up what it does, how it does it, and (most importantly) what’s not perfect about it.
You might think you already know these things and can skip this step, but don’t!
3. Create a new branch off of master (or develop or whatever) that does not include your spike code.
In my most recent spike, I’m aiming to turn it into pull requests that do basically the following
- Setup the credentials for accessing the APIs involved
- Pulled the data from the apis
- Setup the OO design to do the computations on that data
- One PR for each field that was computed
There will probably be 5 or 6 separate PRs.
5. Build that functionality. Reach for the code snippets in your notes from your spike as you need them.
6. Take a special pass with each snippet you copy in to make sure all paths through the code are fully covered with unit tests and any short comings that you wrote up earlier are addressed.
This is the most important step!
It’s not always going to be easy to do this. For example, Sometimes it will be really hard to break your spike into a bunch sane small PRs. Over time one develops a partly intuitive taste for what’s going to work out best. The only way to develop that taste is to try your best every time you find yourself in this situation.