benchmarks

The benchmarks are in tidepool-bench. They use release builds and are not part of just ci.

command	measures
just bench	ingest throughput, for each format
just bench-query	query latency on the reference dataset
just bench-live	freshness: 50 WebSocket sessions while tidepool ingests 100,000 rows/s
just bench-restart	the restart time of the reference dataset

The reference dataset is the retail star schema with up to 100 million orders rows.

numbers

All on an Apple M1 Max (10 cores), from the milestone notes in docs/plan.md.

Ingest (just bench, 8 clients, batches of 10,000 rows, timed until the last row is visible):

format	rows per second
NDJSON	about 2.2 million
Arrow IPC	about 2.4 million

A stored orders row takes about 28 bytes.

Queries (just bench-query, 100 million orders, region × category, three metrics, 200 selections that are each a new combination):

	p50	p95
a new selection	about 68 ms	about 113 ms
a selection toggled back	less than 1 ms

The first query after a start takes 170 to 290 ms. It builds the join columns and the relationship maps.

Live (just bench-live, 200 ms epochs):

	result
an epoch of 100,000 new orders, at 100 million orders	0.6 to 0.9 ms
from the ingest acknowledgement to the frame, 50 sessions, 100,000 rows/s	p50 128 to 180 ms, p95 220 to 228 ms

Restart (just bench-restart, 100 million orders): 1.2 s.

arguments

just bench --rows 1000000 --batch 10000 --clients 8
just bench-query --orders 100000000 --runs 200 --metrics revenue,order_count
just bench-live --orders 10000000 --sessions 50 --rate 100000 --seconds 30

Other work on the machine moves the latency by 15 to 20 ms. When you report a number, also report the load average of the machine.