Iceberg 3's new type lets AI send messy JSON without breaking your warehouse

Apache Iceberg 3.0 just landed a new data type called V3 Variant. It's a big deal for teams that feed AI models into their data warehouses. The old way meant you had to guess what your JSON payloads would look like and build a rigid schema. If the AI changed its output, everything broke.

With V3 Variant, you just dump the JSON as-is. Iceberg stores it. You can still query it. The data stays structured enough to search but flexible enough to handle whatever the model spits out. No more schema migrations every time the model updates.

This is the kind of tooling the side-hustle crew and small shops need. You don't have an army of data engineers on payroll. You have a cousin who knows Python and a laptop. This type lets you build the thing, test it, and ship it without a week-long schema change.

The post at semyonsinchenko.github.io walks through the technical details of how it works under the hood. TLDR Data's Dan also covered it. The link is at iceberglakehouse.com.

Why this matters for us: When AI tools start talking to our data, we need a way to keep the structure without losing the flexibility — because the models change faster than our schedules.