It's one of those difficult things. You have to restart and run all to make sure your code is working and your state hasn't gone fudgy, but you also don't want to re-run that cell that took 3 hours to run unless you have to.

My personal best practice is to 'restart and run all' to run the entire thing before I commit it to github. Jupyter git integration is a whole separate bag of worms, as the combination of presentation/code violates some core git assumptions.