Embracing the Unscalable: My 3-Month Experimentation Framework

In the startup world, you may have come across the well-known advice from Paul Graham: “Do things that don’t scale.” However, implementing this principle in the realm of coding is rarely discussed. After eight months of developing my AI podcast platform, I’ve crafted a straightforward framework to test unscalable hacks: each one gets a trial period of three months. After that time, if it proves its worth, we’ll build it out; if not, it gets cut.

As engineers, we are often conditioned to seek scalable solutions right from the beginning. We chase complex design patterns, incorporate microservices, and set up distributed systems—all to accommodate millions of users. While these approaches are vital for large corporations, they can be traps for startups.

In a nascent venture, aiming for scalability often leads to wasted time and resources, as we’re optimizing for future users who may never materialize. My three-month framework encourages me to produce straightforward, albeit “messy,” code that can be deployed quickly, allowing me to learn about genuine user needs.

Here are My Current Infrastructure Hacks and Why They’re Actually Beneficial:

1. Single Virtual Machine Setup

I’ve consolidated my database, web server, background tasks, and Redis onto a single virtual machine, costing just $40 a month. While this setup lacks redundancy and relies on manual backups, it has provided invaluable insights. Within just two months, I’ve gained a clearer understanding of my resource demands. My AI-heavy platform only requires 4GB of RAM at peak usage. The complex Kubernetes architecture I considered would have been an unnecessary management headache for empty containers.

When system failures—yes, there have been a couple—occur, I gather reliable data on what actually fails, which often surprises me.

2. Hardcoded Configuration

Instead of relying on config files or environment variables, I utilize constants spread across my files:

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

This simplistic approach means I can search my codebase for configuration values in seconds. Each change is recorded in git history and subject to review—even if I’m the sole reviewer. While developing a configuration service could take a week, my three alterations in three months resulted in a mere 15 minutes of redeployment.