What “silent” actually means
Job success is not contract success. Loaders coerce strings into numbers, drop unknown fields, or write nulls into formerly required keys. The warehouse table still grows. The dashboard that joined on the old key goes sparse. Ops sees green; finance sees a wrong number on Monday.
The three drift classes that matter
Additive: new optional columns. Usually safe if consumers ignore unknowns, but still need an explicit accept so models do not invent meaning. Type change: int to string, timestamp without zone, decimal precision. These pass many loaders and poison aggregates. Breaking drop or rename: the field consumers still read disappears or changes name. Block these by default.
Detection that operators will use
Compare expected schema (last accepted) to observed schema on every run. Surface a visual diff: add / type / drop. Page or ticket on break; never auto-merge a drop into production tables. Keep the last-good schema pinned until a human accepts.
How ETL Pack approaches it
ETL Pack detects add, remove, and type-change; blocks breaking drops; requires explicit accept for new columns; alerts on drift. Details and waitlist: https://91crew.com/etl-pack.html
Related: ETL Pack · Schema evolution · Failure recovery