FOSSWire
databases-storage

Citus matches Neki performance at 10M QPS using fewer data nodes

Brief written by the FOSSWire AI editor from Planet PostgreSQL

1 min read Impact: 5/10

Key Takeaway

Using a new Query Router architecture in StackGres 1.19, open-source Citus achieved 10M QPS using roughly half the data nodes required by PlanetScale's Neki.

Summary

A new benchmark from StackGres compares the performance of open-source Citus against PlanetScale's Neki sharded PostgreSQL. By introducing scalable Query Routers to remove the single-coordinator bottleneck, Citus matched Neki's 10M QPS throughput while utilizing only 48% of the data nodes and significantly fewer vCPUs. The test was conducted on AWS Graviton4 instances using a read-only workload.

Technical details of the Citus vs Neki benchmark

  • Architecture: StackGres 1.19 introduces scalable Query Routers to replace the traditional single Citus coordinator.
  • Throughput: Citus achieved 10,032,084 QPS using 24 workers and 48 query routers.
  • Efficiency: At the 10M QPS mark, Citus used 48% of the data nodes (24 vs 50) and required 0.65x the total vCPUs compared to Neki.
  • Hardware: Tests ran on AWS Graviton4 (arm64) with local NVMe storage; workers used r8gd.16xlarge instances.
  • Workload: A read-only pgbench workload consisting of single-row primary-key SELECTs, including hot and uniform key distributions.

Why it matters

The results suggest that open-source sharding solutions like Citus can achieve massive scale with higher hardware efficiency than proprietary alternatives when using distributed coordinator architectures.

Original reference: postgr.es →