• Sources: primary
  • Summary: StackGres ran single-row primary-key reads on Citus 14.1 and PostgreSQL 18.4 on EKS, with StackGres 1.19 query routers. Throughput reached 10,032,084 tps with 24 r8gd.16xlarge workers and 48 routers, at p50 1.06 ms and p99 11.07 ms. The authors say Neki needed 1.5 to 1.8 times the vCPU for similar throughput. Their workers used local NVMe while Neki used network storage, a difference they acknowledge favours them.
  • Why it matters: It is a published data point for sharded Postgres capacity planning, but it comes from a vendor and cannot be checked until the reproduction code is published.
  • Follow-up: Follow-up post with full setup and reproduction code.