Pains of using Github Copilot
Today I was writing a data pipeline to much some timeseries data.
Usual stuff, clean the columns, join some data together. Check the counts etc.
I was using pandas for most of it because Copilot is faster and more convenient for it.
It felt like a 10x engineer moment where the amount of good code i was able to write was super fast. After 2 hours I was done. When I ran everything and started QAing my work deeper, I found a problem.
It took a bloody 5 hours to debug the problem and it was a missing col name in a list of cols which was created by copilot tab completion. Downstream of it, all the operations missed that col and it was no obvious so I didnt look for it there. I kept sussing that the whole col is null because biz logic of a join is incorrect somewhere. The usual DE assumption. That was wrong!
Here are something I will improve in my day to day because of this:
- Validate early and often: Add data validation throughout the pipeline
- Focus on clear variable names: Not having ambiguous or non-descriptive variable/column names help the copilot make less mistakes
- TDD: Where possible think about what I would test before writing the code itself.