ClickBench 100M: four runs

pg_otel_probe reports from October 30, 2026. Each run executed 43 queries three times on PostgreSQL 19beta3, with no errors. Container: 32 CPUs, 40 GB limit, NVMe.

Metrics with the probe on

Run	Hot, mean	Hot, total	Cold, mean	Cold, total
GPU on Parquet	18.1 ms	1.19 s	378 ms	≈26.55 s
GPU on MergeTree	34.9 ms	6.20 s	279 ms	≈28.62 s
CPU on Parquet	157.3 ms	16.14 s	≈275 ms	≈20.65 s
CPU on MergeTree	78.2 ms	9.42 s	≈153 ms	≈13.58 s

Reports

GPU · Parquet · gpu-parquet/

GPU on Parquet

The vexec executor on the GPU reads the hits_pq6 table through vexec_parquet. This is the fastest hot run, but the first try costs far more than the others.

Probe on / off: mean 18.1 / 18.1 ms, total 1.19 / 1.20 s, cold 378 / 372 ms
Recorded: 634 statements, 29.20 s, peak of 2 backends
Database size: 31.6 MB
Findings: 1 medium (A25: 497 psql sessions, 2.1 % of the time), 3 low

Plans Traces Logs Metrics

GPU · MergeTree · gpu-mergetree/

The vexec executor on the GPU, with the hits table in MergeTree. Five queries (Q34, Q33, Q27, Q28, Q20) make up 83 % of the hot total: they are the only ones still sending data to the GPU on the third try.

Probe on / off: mean 34.9 / 34.5 ms, total 6.20 / 6.14 s, cold 279 / 284 ms
Recorded: 615 statements, 47.05 s, peak of 19 backends
Database size: 60.1 GB
Findings: 1 medium (A25: 475 psql sessions, 1.4 % of the time), 2 low

Plans Traces Logs Metrics

CPU · Parquet · vpcpu-parquet/

CPU on Parquet

Runs on the CPU through vp_cpu, with vexec_parquet storage and the hits_pq6 table. Faster than ClickHouse on 35 of 43 queries, with correct answers.

VP7, no probe: mean 158.4 ms, total 16.23 s
ClickHouse: mean 229.1 ms, total 21.23 s (VP7's run)
Recorded: 137 statements, 53.27 s, peak of 34 backends
Database size: 31.6 MB
Findings: 3 low

Plans Traces Logs Metrics

CPU · MergeTree · vpcpu-mergetree/

CPU on MergeTree

Runs on the CPU through vp_cpu, with the hits table in MergeTree. Faster than ClickHouse on 11 of 43 queries, with correct answers.

VP7, no probe: mean 79.1 ms, total 9.14 s
ClickHouse: mean 67.4 ms, total 7.99 s (VP7's run)
Recorded: 140 statements, 32.95 s, peak of 35 backends
Database size: 60.1 GB
Findings: 2 low

Plans Traces Logs Metrics

How to read the numbers

Hot is the best of the three tries and cold is the first. Totals of the best tries, recomputed from the traces, differ from the table by less than 2 %.
Mean is not the total divided by 43. The values match a geometric mean shifted by 10 ms, so long queries affect it less than they affect the total. This formula was derived from the traces and does not appear in the original run summaries.
≈ marks values computed from the report's traces because the run summaries don't include them. Trace times run slightly below what the client measured.
Recorded covers everything the probe saw: all tries and housekeeping statements. It is not a benchmark result, so don't compare runs on it.
Probe cost. For GPU, each variant was compared over a single pass, with the probe collecting metrics every second (the default is 15 s) and full query text. For CPU, the baseline is VP7's run without the probe, and the difference is too small to measure.
Gaps in the data. All four reports lack read timing and checkpoint logs, so rules A18 and A28 are incomplete. The advisor asks for track_io_timing and log_checkpoints to be turned on.

Metrics with the probe on

Reports

GPU on Parquet

GPU on MergeTree

CPU on Parquet

CPU on MergeTree

How to read the numbers