Hot-Warm-Cold Architecture

Think about your closet for a second. You've got everyday clothes you grab every morning, jeans, t-shirts, your go-to jacket. Then there's stuff you wear occasionally, a suit for weddings, a rain jacket for trips. And somewhere in the back, there's that box of college t-shirts you haven't touched in five years but refuse to throw away.

Your data works exactly the same way.

A typical production database has millions, sometimes billions, of rows. But here's the dirty secret: 90% of queries hit less than 10% of the data. Most of your storage is sitting there, eating disk space and costing money, while only a tiny slice actually gets read with any regularity.

Elasticsearch clusters at companies like eBay and Uber discovered this the hard way. They were burning thousands of dollars per month on SSD-backed nodes storing log data that nobody had queried in six months. The data was technically "live", same cluster, same hardware, same cost, but it was effectively dead weight.

Hot-warm-cold architecture is the answer: put your data on hardware that matches how often it gets accessed.

Hot-warm-cold is a data tiering strategy that classifies data into three (sometimes four) temperature tiers based on access patterns, then stores each tier on appropriately priced hardware.

The key insight: you don't need Ferrari-speed storage for data nobody reads. By matching storage tier to access frequency, you can cut storage costs by 60-80% without meaningfully impacting user experience.

This isn't just a storage trick, it changes how you think about data lifecycle. Data is born hot (a new order, a fresh log entry), ages into warm (last month's orders, last week's logs), and eventually goes cold (archived records from two years ago).