Ask HN: How do you like to version datasets for production ML systems?

At my previous company (industrial ai vision platform) we had an ad-hoc mechanism for continuously versioning datasets for user applications that we built internally several years ago.

I am curious how folks are versioning datasets now in practice. Platforms like Hugging Face and Weights and Biases seem to provide good abstractions for dataset versioning. Any feedback on the suitability of these for production systems, or other systems, patterns, or best practices that you have found to work well?