Skip to content- Sources: primary
- Summary: Across a slice of Neon's production fleet, the compute local cache served about 1.1 billion page accesses per second, pageserver NVMe served about 6 million, and object storage about 39, roughly 1 in 28 million. Pageservers answer GetPage@LSN by replaying WAL deltas on image layers, with p90 under 1 delta and p99 around 6 after images were created more eagerly. Object storage reads mostly come from long-dormant projects waking up and from branches or restores at old LSNs.
- Why it matters: For teams weighing Postgres built on object storage, the data shows cold-read latency mostly falls on dormant projects and point-in-time branches rather than on active queries.