Headroom

400× today's traffic, on the hardware we already pay for

Load tests on real ClickHouse Cloud from inside the cluster, with production's data size seeded (206,444 journeys). 400× = 730 events/second, sustained. Answers matched an independent SQL rebuild with 0 bad rows.

730/s
events sustained
8.5 s
95% of events visible within
10 s
to drain after load stops
0.6 CPU
of staging's 2 at 400×
95% of events visible within (log scale)
Old worker · 100× · local
423 s
New · 100× · Cloud
4.0 s
New · 400× · Cloud
8.5 s
The old worker never caught up at 100×. The new one is 50× faster at 4× the load.
Backlog per minute stays flat0s6s12smin 1min 2min 3min 4
400×100×
Parallel queries (P95 at 400×)
1 parallel query
9.6 s
3 parallel queries ✓
5.1 s
8 parallel queries
4.7 s
3 is the sweet spot; 8 adds little. Capped by a ClickHouse profile so it can't crowd anyone out.
Spare headroom: page size (P95 at 400×)
1,000 journeys/page ✓
15.0 s
3,000 / page
8.9 s
5,000 / page
7.8 s
Already passes at 1,000. Larger pages are a free 40–50% if we ever need it.