[mikepurvis]

The "managed state" is important, though. For example I deal with cases where I'm doing some big expensive operations upfront (eg, processing a bunch of data coming in over a network share— ROS bag files, which are also bzipped, adding more upfront processing cost), and I'm interesting in tweaking the plots and analysis coming out the other end without having to re-run the processing each time.

Even as a relatively experienced developer, I've found it hard to reason about what exactly it is that Jupyter is doing under the hood; basically things will break in weird ways, and rerunning everything will mysteriously fix it, or I have to use hacks to force it to re-import certain modules in order to get things the way I expect.

Basically, it ends up being easier to just manage this myself, like doing the processing and pickling the result, then using other, standalone notebooks to load and render the pickle. But this shouldn't be necessary.

[mynewtb]

That sounds really weird. The only time I noticed such weirdness was when a colleague and me were simultaneously editing a notebook on Google Colab. The code was shared but not the state, super weird...