Benchmarks¶

make bench runs wrk against a local gateway and a mock backend that answers after 2 ms, then writes bench/RESULTS.md.

The query: { posts(limit: 100) { id title <author field> { name } } }, 100 posts by 100 distinct authors, with three versions of the author field (see Batch vs no batch).

Results¶

wrk for 10 s with 16 connections, OpenResty with 2 workers, 4 CPUs (x86_64), 2026-10-11:

Variant	Subrequests per query	req/s	Latency avg	Latency p99
`naiveAuthor`	101 (then 101)	425.73	47.98 ms	577.41 ms
`author` (`@batch`)	2 (then 2)	1467.53	10.89 ms	15.21 ms
`cachedAuthor` (`@batch` + `@cache`)	2 (then 1)	1851.52	8.64 ms	12.78 ms

xychart-beta
    title "Requests per second (higher is better)"
    x-axis ["naiveAuthor", "author", "cachedAuthor"]
    y-axis "req/s" 0 --> 2000
    bar [426, 1468, 1852]

Cost of the metrics¶

Same query, with Prometheus metrics (log phase, upstream timing, as generated by gqlc render) and without: 5 alternating runs of 8 s per configuration. Medians:

Variant	Without metrics	With metrics	Difference
`author`	1464.54 req/s	1442.61 req/s	−1.5 %
`cachedAuthor`	1925.46 req/s	1951.19 req/s	+1.3 %

The difference is within the measurement noise of a busy desktop machine. A first version, which built a label string for every increment, cost 4.7 % on cachedAuthor: the keys are now memoized and the cache lookups aggregated per round.

Running it¶

make bench                                   # writes bench/RESULTS.md
DURATION=30s CONNS=64 make bench             # longer, more connections
METRICS=1 OUT=/tmp/with-metrics.md make bench