Benchmark details

Cached page RPC · campaign started 2026-10-17 UTC

Results

Successful requests/s

7,496 (range 7,033–7,683)

API memory

32.8 MiB PSS

API CPU time/request

120 µs

Median trial p95 / p99 latency

5.10 / 6.38 ms

Workload and environment

Apple M1 Pro running macOS, with an eight-vCPU ARM64 Colima Linux VM and a Debian API container. One vCPU for the API, four for PostgreSQL, and two for wrk. The database pool had eight connections; wrk used 32 HTTP connections.

The SQL function returned 100 rows from a fixture of 10,000 rows, 100 tenants, and 128-byte payloads. Each request supplied a 1 KiB JSON argument. Frontend HTTP and plaintext PostgreSQL connections were used.

RS256 verification was enabled, with JWT caching and ten reused tokens. The cache allowed 1,000 entries; pgxio also budgeted 1 MiB for accounted claims and digests, excluding metadata overhead. The experimental-libpq image used copied buffers, not zero-copy forwarding.

Method

Three randomized trials, each with 10 seconds of warmup and 30 seconds of measurement. Throughput is the median successful requests per second. Memory is the median of trial mean API-process proportional set size (PSS), excluding PostgreSQL. Latency values are medians of trial percentiles, not pooled percentiles.

These are shared-host local measurements. They do not establish minimum RAM requirements or whole-server capacity. Repeated-token caching results do not represent workloads dominated by unique tokens.

Reproducibility

Download measurements and provenance (JSON)

API image:

sha256:2daebd31eb22c698f232deb1d78d14939959c77dd126ff43eeeeb211c1f2bfcd

PostgreSQL image:

sha256:3645570cccdfa447589da9f57dd740faa29b30938e861289a5574b6ca6b03826