Code Migration Using LLMs

Last year, I evaluated how effective Large Language Models were in writing unit tests. This year, I decided to evaluate how effective LLMs were in migrating code from one programing language and tech stack to another. This isn't my initial reactions to migrating some 10 line function. My results cover the migration of an entire polyglot persistent microservice which is written in about 1,000 lines of code and integrates with 4 different backend technologies.

Key Takeaways:

• LLMs can dramatically speed up initial code migration. This includes open source LLMs such as Llama 3.3
• Human review and manual refinement is crucial, especially for framework-specific patterns like dependency injection.
• Copilot's statement completion is a powerful productivity booster.
• LLMs are still evolving, with limitations in complex problem-solving.

You can read all the details in my blog entitled Using Large Language Models to Migrate a News Feed Microservice.