You might be interested in marimo

[dmadisetti]

@amakelov I just implemented something very similar to this in marimo (notebook with different kernel to ipython https://marimo.io), without knowledge of this repo:

marimo-team/marimo#1758

Cool stuff! It looks like ComputationFrame tries to address some of the same challenges.

Differences by bullet points in your readme: 1) compute dependency static analysis in the isn't done by marimo's cache implementation since marimo itself prescribes a compute graph (and cache just leverages off that).

2. But the implementation very much does utilize variable relationships and computed values. I'm curious about your hash implementation for function inputs, marimo's implementation is very naive right now, but I was considering implementing opt-in object __hash__. Let me know if you have thoughts about your joblib implementation. I will also check out the papers and readings

3. Exploration is not utilized, but I have thought about this a bit, esp. with respect to marimo's wasm exports. Imagine you have a notebook with a bunch of UI hyper parameters for an expensive model; you could enumerate the hyper parameters combinations and save the results. The resultant notebook then just lets you click between the model results without much user time spent. You could even export to a blog post, where imports like torch or jax are tidied up, and just the resultant model and graphs remain.

---

Any thoughts about a remote cache?