Deno Joins Cloudflare: The Cost of Dependence

A team chooses a runtime because it is faster, safer, or more convenient.

Two years later, dozens of services run on it, while the vendor makes a strategic decision without involving them.

The person who has to pay for it is the manager with a deadline and a budget.

Runtime migration delivers neither new features nor revenue growth, and takes engineers away from product work.

The cost of exit must be assessed when choosing the technology.

Willison examines a similar case using EmbeddingGemma 2, released under the Apache 2.0 license.

His reasoning is based on mechanics.

An application with an embedding model calculates thousands or millions of vectors and stores them for later comparison.

Vectors from one model cannot be compared with vectors from another, so changing models means recalculating the entire dataset.

If a model is proprietary and available only as a hosted service, the vendor will eventually retire it because a better model will emerge.

You will have to pay again to recalculate all stored data. With an open model, you can keep the old version for as long as needed.

This leads to a business rule: if a storage system tied to a specific model grows around your system, the model license becomes part of the architecture alongside the database.

Architectural layers mitigate the risk, and choosing the “right” vendor guarantees nothing here. - A gateway in front of the models.

Applications access the LLM through a single gateway.

Changing the provider or model changes the gateway configuration, while the service code remains the same.

That is how our LLM & Security Gateway works: it also tracks requests and filters sensitive data. - The model version next to the vector. In a RAG system, each vector is stored with the model and version identifiers.

Recalculation can then run in batches in the background, and search does not stop. - Integrations through standard protocols.

Agent tools are published through MCP, while communication between systems runs through Apache Kafka.

The service runtime (Node.js, Python, or C#) can be replaced without revising contracts. - Open weights for large datasets.

For large embedding datasets, an open model with a clear license (Qwen, Llama, Gemma) costs less than relying on hosting that may be shut down.

This architecture may look dull, but its purpose is specific: if the vendor is acquired or raises prices next year, the move should take weeks of one team's work. We do not accept a quarter of work by several teams in this calculation.

Engineers often measure a technology by its startup speed.

We measure time to value, but it has a second half: how long it will take to exit when the technology no longer fits. A runtime that lets you build a service in a week but requires six months to migrate after a change of ownership is more expensive than one that takes two weeks to build with and can be replaced over a weekend. I am not claiming that Deno was a bad choice.

For many teams, it will remain operational for the entire promised year.

But those who built long-lived systems on it need a plan for the second year.