Has anyone successfully run Docling on Databricks?

You may also want to benchmark alternatives instead of assuming Docling is the best fit. In our testing, Kreuzberg, now called Xberg, was significantly faster and more reliable for our documents, with better task-level results.

The GIL failures seem to be a tradeoff for Docling’s parsing capabilities. If you stay with it, isolate parsing into separate processes or services and keep each batch relatively small rather than running everything in one long-lived Databricks process.

We ran into similar problems. Processing smaller batches and spreading the workload across more workers helped keep the GIL from locking up the runtime. It still needs some monitoring, but the approach is workable for moderate workloads.

We ended up running Docling through a serving layer rather than directly inside the notebook or job process. That made the setup more scalable, reliable, and cost-effective, especially when several parsing requests were running at once.