Often you separate the feature exploration environment from the infrastructure that executes a strategy. You may need to rebuild the feature calculation before

Often you separate the feature exploration environment from the infrastructure that executes a strategy. You may need to rebuild the feature calculation before deploying it into prod, but if you try to provide the quants with solid enough tools so they explore the statistics of features and models without having to think much about the built in error handling, that can work well to avoid the rewrite step. You probably end up wrapping numpy and pandas with a very opinionated internal view of how the business handles the kind of edge cases your team sees in your data.