Tractatus on CICD

As we now know, building products in a problem domain is difficult. We also know we will be working in that problem domain for a long time. It makes sense then to build a tool that allows us to build and maintain products easily. This is what collectively has been called in software as having a “CICD” stack. Surprisingly, lots of companies are not obsessed enough with its importance. They may think of it as a nice-to-have, second-hand infrastructure. Obsessions in software are rarely a mark of good engineering. However, if I had to make a bet on one thing worth obsessing about it’s the CICD stack. This is because a software company is essentially a huge factory of product updates guided by immersion in the customer's problem domain. Products are difficult to make, so they will need lots of fixes and enhancements that we will need to be delivered to users. Making products is also an inherently iterative process, so we can’t plan and deliver all features in one go. We will need to ship them in a never ending cycle of updates that respond to changing customer needs and technology demands of the day. Employees that were hired to solve problems in the problem domain will spend their days interacting with the CICD stack. They will try to deliver their value into customer’s hands and each obstacle in the process will contribute to their burnout. As Elon said, it’s not about building a car, it’s about building a machine that builds the car.

Popular posts from this blog

Terminology of CICD

To be clear, there are two types of software: libraries and applications. Libraries are used by applications to provide some key piece of functionality. Applications are always deployed to some Environment and run to perform computation. It should always be clear which one of the types is being implemented in a given piece of code. Building and publishing a new version of source code which is ready for use is called Releasing the library/application. The release is always associated with a new version number. Act of change to use a new version of an application in a given environment is called a Deployment. Deployments may also be other system or data changes that affect state produced by the application or general behavior of the application. All such changes should be associated with some code changes making them easy to track.
Read more

Perils of classes and pointers

When trying to understand or debug a piece of code the amount of state that one needs to keep in mind is inversely proportional with success. This leads us to an obvious advice to try to limit scope of each function to the essential components. However, there are also certain programming concepts that contribute to the amount of state one needs to track. The first one is the use of classes because class attributes are accessible in each method and may change between calls. The second one is a pointer (or reference). Pointers are problematic because, unless explicitly managed like in rust , they make it difficult to determine the owner of data at a given point of computation. This leads to uncertainty about what state a given variable may hold. Be very careful when adding classes and references to your code. Sometimes a pure function is just better.
Read more

On Testing

Test a piece runs A basic building block of a codebase is a function. So for each building block let's just at the very least test it can be called with some reasonable input and produce the right result. If you feel that’s not needed perhaps you don’t need that function at all. That is a base case for testing that prevents breaking what works. It should be enabled to run on all PRs. Every test case should have a human understandable description of the test (LLMs will be able to generate those tests soon, so better be ready). Test it runs end to end Once a bunch of functions are composed things get tricky to check if it all works. So let's test if a thing can be run end to end and produce something sensible at the end. It can be enabled on all PRs and definitely done before any release. Test pieces produce similar results Once a thing runs end to end it will produce a result. But it's hard to say that the result is right without anything to compare. So instead, we will test...
Read more