DB perf considerations aside, a lot of software pattern around idempotency/safe retries/horiz-scaling/distributed systems are super awkward with a serial pk bec

DB perf considerations aside, a lot of software pattern around idempotency/safe retries/horiz-scaling/distributed systems are super awkward with a serial pk because you don’t have any kind of unambiguous unique record identifier until after the DB write succeeds. DB itself is “distributed” in that it’s running outside the services own memory in 99% of cases, in complex systems the actual DB write may be buried under multiple layers of service indirection across multiple hosts. Trying to design that correctly while also dealing with pre-write/post-write split on record id is a nightmare.