Comment by sgarland

Every company I’ve been at that relied on application code to handle referential integrity had orphaned rows, and incidents related to data errors or the absurd pipelines they had built to recreate what FK constraints and triggers already do.

RDBMS are extremely well-tested pieces of software that do their job incredibly well. To think that you could do better, or even equally as well, is hubris. If you want to trade those guarantees for “velocity” go right ahead, but you also need to take into account the inevitable incidents and recoveries that will occur.