Neon's lakebase architecture separates compute and storage, eliminating the need for full-page writes (FPW) — a decade-old PostgreSQL durability mechanism that protects against torn pages on local disk. Since Neon's compute is stateless with no local disk, torn pages can't occur, so FPW is unnecessary. However, simply disabling FPW would cause unbounded delta chains in storage, hurting read performance. Neon solved this with 'image generation pushdown': the distributed pageserver layer now generates full-page images when a page accumulates too many deltas, rather than relying on the compute's WAL stream. This reduced average WAL per transaction from 58KB to under 4KB (94% reduction), yielding up to 4.5x write throughput gains on 32-vCPU instances, 30–50% lower p99 read latencies, and one customer's sync table ingestion jumping from 17k to 62k rows/second. The optimization has been rolled out globally to all Neon databases with no restarts required.
Table of contents
The hidden cost of traditional Postgres durabilityThe Neon solution: eliminating the risk of torn pagesInnovation: image generation pushdown to distributed storageQuantifying the impact: from lab to productionSeamless rollout: performance without interruptionWhat is next for managed Postgres performance?Sort: