Skip to content

Strict OCPQ Q1-Q7 benchmark

pg_ocpm 1.0.0 plus ocpm-engine 1.0.0 passed the strict Q1-Q7 publication protocol against OCPQ 0.6.7 with exact duplicate-preserving parity for every evaluation-tree node. The same-host geometric-mean speedup is 16.137x and the minimum individual-query speedup is 9.522x.

Query OCPQ mean pg_ocpm + engine mean Speedup Exact nodes
Q1 32.517 ms 2.158 ms 15.070x yes
Q2 46.192 ms 3.589 ms 12.870x yes
Q3 25.962 ms 2.490 ms 10.427x yes
Q4 49.905 ms 3.444 ms 14.491x yes
Q5 53.537 ms 5.623 ms 9.522x yes
Q6 105.727 ms 2.813 ms 37.586x yes
Q7 86.676 ms 3.190 ms 27.167x yes
Geometric mean 51.453 ms 3.189 ms 16.137x yes

Both arms used zero warmups, ten measured runs per query, and a fresh container for each query. The candidate timer includes prepared PostgreSQL execution, fetch, native decoding, and complete owned materialization of every node. External-ID canonicalization, sorting, serialization, hashing, and comparison are excluded from both timers but required for every accepted sample.

The 1/4/8/16-client concurrency sweep reached 2,124.6 req/s at 16 clients with 15.661 ms p95. Maximum client RSS above baseline was 6.25 MiB; serving storage was 109.91 MiB. Every pre- and post-epoch node check passed, the result cache was disabled, and source, image, database, host, and clean-tree provenance were complete.

Same-time 0.9/1.0 concurrency diagnostic

The absolute 1.0 p95 was recorded while unrelated host processes were consuming several CPU cores. To distinguish product regression from host load, the byte-identical 0.9 and 1.0 candidate runners were executed back-to-back on the same active Docker host, with their corresponding extension images and the unchanged exactness/concurrency protocol.

Clients 0.9 QPS 1.0 QPS 1.0/0.9 0.9 p95 1.0 p95
1 272.7 277.3 1.017x 7.689 ms 7.422 ms
4 891.5 883.4 0.991x 8.910 ms 8.788 ms
8 1,439.8 1,620.7 1.126x 11.028 ms 9.659 ms
16 2,023.8 2,301.6 1.137x 16.390 ms 14.481 ms

All pre/post answers remained exact. This diagnostic shows no 1.0 concurrency regression, but it is not substituted for the clean 1.0 publication artifact. The checker retains exact raw-latency reconstruction, at most 15% epoch-QPS variation, at least 5x 16:1 scaling, and a portable p95/p50 tail-amplification limit instead of treating an unreserved host as an absolute-latency SLA.

The raw diagnostic artifacts are ocpq-concurrency-ab-pg-0.9.0.json and ocpq-concurrency-ab-pg-1.0.0.json in the repository's docs/results/ directory.

See the four-way comparison for the two PM4Py arms and the interpretation boundary.