Portability and lock-in

Lock-in is not a moral failing of vendors. It is the natural result of a component becoming load bearing faster than anyone measured the cost of removing it. The defence is to know that cost per component, before it matters.

For an agentic analytics stack, run the exercise layer by layer.

The model

Cost to replace: low. If prompts and tool definitions are in version control and the harness talks to models through a thin adapter, this is a configuration change. Keep it that way. Any feature that only exists in one provider’s API and sits on your critical path is worth pricing honestly before you depend on it.

The harness

Cost to replace: moderate. This is your code, or a framework you have wrapped. Rewriting an agent loop is a project of weeks, not quarters, as long as the tools it calls are exposed over an open protocol and not embedded in the framework.

The interfaces

Cost to replace: low if open, very high if not. An MCP tool server can be swapped for another MCP tool server. A proprietary agent-to-data integration cannot, and it has typically spread through every team that adopted it.

The engine

Cost to replace: low, if the data is in open formats. This is the whole argument for the lakehouse. When tables are Iceberg over Parquet in your own storage, adding or replacing an engine is a connection change. When the data lives inside a proprietary system, the engine and the storage are the same decision and moving means an export.

The catalog

Cost to replace: high, and rising. It holds the namespace, the grants, and the credential relationship with storage. Everything points at it. If it speaks the Iceberg REST protocol, another implementation can take over. If it has a proprietary API, this is the component that quietly becomes unmovable. Be strictest here.

The semantic layer

Cost to replace: highest, and most often overlooked. These definitions took the longest to agree on and encode the most institutional knowledge. If they exist only inside one BI tool’s proprietary model, moving means reconstructing years of arguments between finance and operations.

This is now the most likely lock-in point in a modern stack, precisely because the layers below it solved their portability problems and this one has not yet. Apache Ossie is the effort to close it. Until it matures, the practical defence is to keep the semantic model in version control, in a format you can export and read without the vendor’s product.

The question to ask each vendor

Not “is it open source”. Ask: if I decided to leave, what would I take with me, and what would I have to rebuild?

A satisfying answer describes files you already have in storage and definitions you already have in a repository. An unsatisfying answer describes an export process. The second answer is the one that costs.

Primary sources

• Open Source and the Data Lakehouse Assessing openness one layer at a time.