Embracing the 3-Month Experimentation Rule for Scalable Coding

Navigating the intricate landscape of software development often leads us to the well-known advice from Paul Graham: “Do things that don’t scale.” However, the challenge lies in how we put this principle into action, particularly in the realm of coding for startups.

After dedicating the last eight months to creating my AI podcast platform, I have devised a straightforward yet effective framework: any unscalable hack is given a lifespan of three months. At the end of this period, each approach must either demonstrate its value and undergo a proper transformation or be discarded entirely.

As engineers, we frequently find ourselves drawn to the allure of building scalable solutions right from the outset—think design patterns, microservices, and distributed systems tailored to accommodate vast numbers of users. While this mindset is common in larger organizations, it can often lead to expensive procrastination in a startup environment, where we may end up optimizing for potential users who might never come. My three-month rule compels me to focus on crafting straightforward, albeit imperfect, code that gets shipped and provides real insights into what users genuinely need.

My Current Infrastructure Hacks: Smart Strategies Behind Unconventional Choices

1. Consolidation on a Single Virtual Machine

All vital components of my platform—database, web server, background jobs, and even Redis—operate on a single $40/month virtual machine. There’s no redundancy, and backups are manually conducted to my local device.

This approach has proven to be surprisingly insightful. In just two months, I gained a clearer understanding of my actual resource requirements than any extensive capacity planning document could have provided. My so-called “AI-heavy” platform predominantly peaks at 4GB of RAM. The complex Kubernetes architecture I nearly implemented would have found itself managing idle containers instead. Each crash—yes, it’s happened twice—has yielded unexpected data regarding failure points, always diverging from my initial assumptions.

2. Directly Hardened Configuration

python
PRICE_TIER_1 = 9.99
PRICE_TIER_2 = 19.99
MAX_USERS = 100
AI_MODEL = "gpt-4"

I’ve opted out of configuration files and environment variables in favor of constants scattered throughout my code. Modifying any variable necessitates a redeployment.

The hidden value of this method? I can swiftly search my entire codebase for any configuration item. Not to mention, each price adjustment