Has anyone deployed Docling reliably on Databricks?

For a more production-friendly setup, consider running Docling as a separate serving process instead of embedding it directly in every Databricks task. That gives you a scalable endpoint, isolates parser failures, and can be more economical than repeatedly starting the full runtime.

We ran into similar problems when processing too much in a single job. Splitting the workload into smaller batches and spreading them across separate workers made the GIL failures less frequent. It still needs some monitoring, but the setup became usable.

We compared Docling with Kreuzberg, also known as xberg, on a cloud GPU platform. Kreuzberg was noticeably faster and more reliable in our tests, with better task performance. Docling is capable, but it may be worth benchmarking alternatives against your actual document mix before committing to it.