Eric J Ma's Website

As I've been working on automated Git tooling (i.e. automated commit messages and release notes powered by LLMs), I've been thinking about how this applies to data scientists' workflows.

When I was a front-line data scientist, I found that committing work was sometimes a psychological barrier. The reason for this was because of the commit messages. Having been previously taught that we commit code when it feels done, committing in-progress exploratory work seemed to break the rule of not committing work that was in-progress.

However, the exploratory nature of data science work means that we often need to commit work that is in progress. That may look like a Jupyter notebook with commented code (but not fully-fleshed out commentary in the Markdown cells). Committing the notebook in its unfinished state at the end of the day with an uninformative commit message feels like a high cognitive barrier.

However, if we had an automatic commit message writer, we might have a commit message that looks like this:

commit 0f0155cea572dc980ecddb3dfcce9723c7904a3f
Author: Eric Ma <eric**********@*****.com>
Date:   Sun Sep 17 20:19:23 2023 -0400

    feat: Implement binary genotype-phenotype model

    This commit adds the implementation of a binary genotype-phenotype model in Python. The model is based on a mathematical framework proposed in a research paper by Yeonwoo Park. The implementation includes functions to calculate zeroth, first, and second-order effects, as well as to predict the phenotype value for a given genotype. The code has been tested with a simple binary genotype system with 3 positions and will be further tested with a 5-genotype system. This implementation aims to provide a reference-free analysis of genotype-phenotype relationships.

Notice the phrase:

and will be further tested with a 5-genotype system

This is the kind of thing I wouldn't write on my own! Leaving notes for myself in the commit log, based on the notes I leave for myself in my notebooks, actually makes the commit log useful for me. My own commit message would look more like this:

commit 0f0155cea572dc980ecddb3dfcce9723c7904a3f
Author: Eric Ma <eric**********@*****.com>
Date:   Sun Sep 17 20:19:23 2023 -0400

    Saving work.

My lazy brain's commit message is uninformative and unhelpful for resuming work later on.

I think data scientists don't take advantage of commit logs as much as we possibly could because we're too lazy to write commit logs ourselves. But if we have good commit logs, then we'll have the equivalent of a digital lab notebook that summarizes our work automagically as we go about our work. And that lab notebook is going to be incredibly helpful for our work.

If you're interested in trying out a service I've built that will automatically compose commit messages on every commit, please check out GitBot!