The first 90 days of open data infrastructure should not look like a platform rebrand. It should look like a controlled experiment with teeth.

Days 1 to 30: map control, not tools

Start by mapping where control lives today. Which systems own the data files? Which systems own table metadata? Where are policies enforced? Which workloads create copies because direct access is too hard? Which AI use cases are blocked by missing context or governance?

This is not an application inventory. It is a dependency map. The goal is to find the places where a vendor boundary, private metadata model, or undocumented workflow controls the business more than people admit.

By day 30, you should have a ranked list of ODI candidates and a short list of constraints that would make the first pattern fail.

Days 31 to 60: build one governed open path

Pick one valuable domain. Publish one or two tables through an open table format when that is the right fit. Register them in a catalog. Attach ownership, access rules, lineage, freshness expectations, and quality signals. Then make at least two workloads read the same governed data without creating a new copy.

This is where the architecture becomes real. The question is not whether the first query works. The question is whether the pattern can survive schema change, access review, data repair, and a second engine.

Days 61 to 90: prove the operating model

The last month is about evidence. Measure how much copying was removed, which manual governance steps became platform behavior, which workloads can now reuse the same data product, and which gaps still require engineering.

Document the pattern in plain language. Show the control plane, the data plane, the write path, the read path, and the policy path. If people cannot explain how the pattern works, they cannot operate it.

Core idea: the first 90 days should produce a reusable control pattern, not a pile of disconnected open-source parts.

Avoid the transformation trap

Do not spend the first 90 days naming committees and arguing over the final platform. That work may be necessary later, but it will not prove anything. A small working pattern beats a perfect target architecture nobody has tested.

The first win should make the next decision easier.

Sources to start with

These are the primary sources I would start from when checking the claims in this piece.