The "throw everything in Postgres" angle ignores how bad most DBs are at large-object lifecycle ops like bulk deletion, partial reads, and massive tree renames. Put a million big blobs in there and backup, restore, and replication get ugly.

People also skip over fast atomic moves and concurrent edits across deep hierarchies, which is exactly where a file system earns its keep instead of pretending SQL is a universal storage layer. If uptime and ops sanity matter, you usually bolt one on anyway.