I can say that even now lots of machine learning datasets are kept as rows in gSheets and processed by Python running in Colab. In our team, all our evaluations: read input data from gSheet, run model inference, store outputs in gSheet for manual review if needed, process outputs to compute metrics and plot graphs in Colab. It is great because evaluation logic frequently changes, and without Colab you would sit and wait days each time for code reviews and stupid readability approvals.

And then, of course, if you want reproducibility, you just check in Colab notebooks into the source control.