I was under the impression that it is normal to never alter or drop columns in large production tables for performance reasons. Instead, I'd prefer to to deploy

I was under the impression that it is normal to never alter or drop columns in large production tables for performance reasons. Instead, I'd prefer to to deploy a new version which can handle only the old OR both the old and new columns. Then, the data organically evolves to the new schema; maybe a background job does it over time. This is the approach you need to use with schema-less solutions like BigTable.