At ~35TB Postgres we ended up separating pg_wal onto its own volume first — write amplification there was quietly eating 20–30% of our NVMe throughput before we caught it.