The Three-Month Rule: A Pragmatic Approach to Rapid Learning in Software Development
In the realm of startups and software development, one piece of advice often echoes through the community: “Do things that don’t scale,” a mantra popularized by Paul Graham. However, the challenge lies not in understanding this philosophy but in applying it effectively in the world of coding.
As I delve into the development of my AI podcast platform, a journey spanning eight months, I’ve crafted a straightforward yet powerful framework: any unscalable solution is given a lifespan of just three months. At the end of this period, we either determine its value and refine it for scalability, or it is discarded.
As engineers, we’re indoctrinated to create scalable solutions from the onset. We focus on design patterns, microservices, and robust architectures that can handle vast numbers of users. While this approach suits large organizations, in a startup environment, chasing scalability can often feel like an exercise in costly procrastination. We find ourselves optimizing for potential users who may never materialize and tackling challenges that may not surface. My three-month rule encourages me to create simple, straightforward, and perhaps “messy” code that can be deployed swiftly and provides invaluable insights into the actual needs of my users.
Current Infrastructure Hacks That Foster Learning
1. Consolidated Operations on a Single VM
All crucial components of my platform – the database, web server, background tasks, and caching—all operate on a $40/month virtual machine, without redundancy and with manual backups. This approach has proven insightful. In just two months, I have gained a clearer understanding of my actual resource requirements than any formal planning document ever could. My platform, which I expected to be resource-intensive, averages about 4GB of RAM usage. The complex Kubernetes architecture I nearly implemented would have resulted in managing idle containers.
Each time the system crashes (which has happened twice), I gather concrete data about the points of failure. Interestingly, the failures are never what I anticipated.
2. Simplified Configuration Management
My code features hardcoded configurations like:
python
PRICE_TIER_1 = 9.99
PRICE_TIER_2 = 19.99
MAX_USERS = 100
AI_MODEL = "gpt-4"
There are no configuration files or environment variables—just simple constants across my codebase. While this may seem archaic, it allows me to swiftly search for any configuration value and track changes through Git
