Replication and MVCC

In fairness, using hot_standby_feedback does increase bloat on the master and therefore impact master performance. Especially in situations like theirs with high churn tables.

Ideally we'd instead store blocks still needed on a replica, but no longer needed on the master, out-of-line using a copy-on-write mechanism of some form. Problem is that it'd have to do indexes as well as the heap, and when we do a tid lookup from an index we'd have to look up both the main heap block and some possibly large number of block copies. It'd get complex fast, and complex means buggy.

It'd also slow down queries on the replica, make replica change apply slower, etc.

Everything is trade-offs, and I think Uber's article misses that somewhat.