DEVOURED
How lakebase architecture delivers 5x faster Postgres writes
Neon's "image generation pushdown" technique, implemented in its lakebase architecture, drastically improves Postgres write throughput by up to 5x and reduces WAL generation by 94% by offloading full-page write operations to the distributed storage layer.
What: Neon re-engineered Postgres's durability mechanism by moving full-page write (FPW) operations, traditionally handled by the compute layer, to its distributed storage layer. This "image generation pushdown" allows the compute to send only compact deltas, while the storage layer reconstructs full pages as needed, achieving 4.5x throughput on 32 vCPU computes and a 94% reduction in WAL traffic in production.
Why it matters: This demonstrates how a decoupled compute-storage architecture can overcome long-standing performance bottlenecks in traditional monolithic database systems, enabling significant efficiency gains and scalability for write-heavy workloads.
Takeaway: If you're using Neon for Postgres, your write-heavy applications should already see significant performance improvements and reduced WAL usage without any action required.
Deep dive
- Traditional Postgres uses Full Page Writes (FPW) to prevent data corruption from "torn pages" during crash recovery.
- FPW involves writing entire 8KB data pages to the Write-Ahead Log (WAL) the first time a page is modified after a checkpoint.
- This ensures recovery even if a disk page is partially written but can inflate WAL volume by up to 15x, becoming a major performance bottleneck for write-heavy applications.
- Neon's lakebase architecture separates compute and storage; compute nodes are stateless and stream WAL to distributed safekeepers.
- Because there's no local disk page to tear, the original need for FPW is eliminated.
- However, simply disabling FPW could lead to unbounded WAL delta chains and slow read performance.
- Neon introduced "image generation pushdown," where the storage layer (pageserver) takes responsibility for generating full page images.
- The pageserver reconstructs pages by finding the most recent materialized image and applying WAL deltas.
- Images are generated when a page accumulates a threshold of delta records, optimizing image generation based on actual changes rather than arbitrary checkpoints.
- This reduces WAL traffic by 94%, improves network efficiency, and scales image generation across distributed storage.
- Benchmarks show throughput gains up to 4.5x for 32-vCPU instances and a 94% reduction in WAL generation.
- Production data from a 56 vCPU project saw WAL generation drop from 30 MB/s to 1 MB/s, with p99 read latencies improving by 30-50%.
- The feature was rolled out seamlessly across Neon's entire fleet since late March, requiring no customer action or restarts.
Decoder
- Write-Ahead Log (WAL): A sequential log used by Postgres to record all database changes before they are applied to data files, ensuring data durability in case of a crash.
- Full Page Write (FPW): A Postgres mechanism where the entire 8KB data page is written to the WAL the first time it's modified after a checkpoint, preventing data corruption from "torn pages" during recovery.
- Torn Page: A corrupted data page on disk that results from a server crash during a partial write operation, leading to inconsistent data if not handled.
- Checkpoint: A milestone in the Postgres WAL that ensures all data changes up to that point have been written to disk, limiting the amount of WAL replay needed for recovery.
- Lakebase architecture: A database architecture, like Neon's, that separates the compute and storage layers, allowing independent scaling and specialized optimizations.
- Pageserver: A component in Neon's distributed storage system responsible for reconstructing data pages for read requests and generating full page images.
Original article
Neon eliminated a decade-old Postgres performance bottleneck by pushing full-page write operations from compute to its distributed storage layer, achieving up to 5x throughput improvements and reducing WAL generation by 94% in some cases. The "image generation pushdown" technique, now rolled out across Neon's entire fleet, leverages the company's separated compute-storage architecture to solve a durability problem that's structurally impossible to fix in traditional monolithic Postgres deployments.
Explore all 68 articles in today's digest
Fresh dev news, AI-digested daily. Never miss a trend at devoured.fyi