I used to build quant investment notebooks that had to be deployed in production. Lots of problems with that. Mine were: Notebook cells run out of order, so you often have something that works in a session, but not in a fresh run. Developing against limited datasets, so you fail against things you didn’t know to test for. Small adaptions that have to be made every time the notebook is translated into a code file. We streamlined it by making a graph-structured Computation a first class object that tracked staleness as code or data was updated. Then that class could be directly published, and when failures happened in production, the graph could be serialized with the inputs and intermediate calculation data that caused failure, for investigation in a notebook.

We open sourced the implementation https://github.com/janushendersonassetallocation/loman