Harper Releases Version 5.2 and Benchmarks Against Vercel-Based Stack

Harper 5.2 Release

Harper, a database platform advocating for a single-runtime architecture, has released version 5.2. This update includes a new record cache and delivers more throughput per node. The platform's architecture aims to keep application code and data co-located.

Benchmark Against Multi-System Stacks

Harper conducted a benchmark comparing its single-runtime approach to a Vercel-based stack. The Vercel stack utilized Vercel Functions, Neon Postgres, Upstash Redis, and Ably for real-time capabilities. The benchmark focused on live, personalized-data workloads, where Harper reported significantly better performance.

The benchmark measured in-process data access at approximately 0.4ms for Harper, compared to a ~3ms network hop for the separate tiers in the Vercel stack. This performance gap widens as a page requires more personalized data.

Architectural Philosophy

Harper's architecture runs counter to trends like Databricks' Lakebase, which separates compute from storage. Harper argues for collapsing data path components—database, cache, and job runner—into a single runtime to simplify operations and reduce costs. This co-located approach aims to keep code next to its data, reducing network latency.

Workload-Dependent Outcomes

The benchmark involved 474 load tests across eight scenarios and two US regions. While Harper's co-located architecture performed better for live, personalized-data paths, the Vercel serverless stack outperformed Harper's single free node under high, sustained fan-out loads due to its autoscaling capabilities. The results indicate that the optimal architecture is workload-dependent, with live data favoring co-located runtimes and high concurrency fan-out favoring serverless stacks.