It's very unclear to me why the writer ran everything through GCP and Colab. Having everything on a single environment with near-100% uptime is certainly still

It's very unclear to me why the writer ran everything through GCP and Colab. Having everything on a single environment with near-100% uptime is certainly still possible by renting a VPS, which is usually a much cheaper alternative anyways. There's also many other alternatives that could be used if memory efficiency is the goal, e.g. polars/data.table instead of pandas, an OLAP database instead of bigquery etc. While I agree that pasting together configs is tedious and annoying, a lot of this type of work is due to developers themselves trying to replicate new trends at $bigco$ instead of optimizing for their own needs.