5x perf increase on writes with FPW disabled in Postgres

2 days ago

Lakebase architecture decouples compute and storage, enabling performance optimizations impossible in monolithic Postgres.
Traditional Postgres uses Full Page Writes (FPW) to prevent data corruption from torn pages during crashes, but this inflates WAL volume by up to 15x.
In lakebase, compute is stateless with no local disk, eliminating the torn-page risk FPW addresses.
Disabling FPW naively could cause unbounded delta chains in storage, increasing read latency and resource use.
Image generation pushdown moves FPW-like image creation to the storage layer, generating full page images based on delta thresholds rather than checkpoints.
Benchmarks show throughput gains scaling with compute size: up to 4.5x+ on 32 vCPU instances and a 94% reduction in WAL traffic.
Production benefits include reduced WAL generation (e.g., from 30 MB/s to 1 MB/s), improved read latencies (p99 down by 30-50%), and higher ingestion throughput (e.g., 3x increase for one customer).
The optimization was rolled out globally without downtime using Postgres's XLOG_FPW_CHANGE mechanism, enhancing scalability and stability.
This is part of a broader effort to offload heavy tasks to scalable storage, eliminating Postgres write bottlenecks and improving performance.

Hasty Briefsbeta