Results

Measured on the three-node lab cluster with the burst harness, against the HPA at a 70% CPU target.

Load	Result
Ramp (Locust)	the pressure signal warns 80 ± 20 s before p95 latency passes 500 ms
Step from 2 to 4 users	request rate reacts in about 1 s, pressure in 1–31 s, the HPA in 19–34 s
Step to 8 or 12 users	closed-loop load hides the burst; the SLO breaks within 0–5 s
Open-loop burst, 5 → 75 RPS, default probes	croft’s request rate reacts in 0–2 s; the HPA never scaled: the overloaded pod failed its probes and restarted, and the metrics API returned nothing; about 96% of burst requests failed
CPU heterogeneity	the same request costs 1.69× more CPU on the slowest node; the HPA averaged 67% while one replica throttled 34%

The ramp rule (the episode in progress when the SLO breaks) was chosen after seeing the data; say so when citing it.

Calibration

The lab webapp costs 45–58 ms of CPU per request. Node calibration gives 38–64 ms and speed factors 1.000, 0.792 and 0.597.

Open questions

• sendto counts answered requests, so a saturated pod shows its capacity, not the offered load. The missing signal is the queue of connections waiting to be accepted (issue #8).
• runqueue_s fires with sendto and PSI on bursts (0.1–2.3 s) and tracks throttling per pod, but climbs with the pod count, so it does not size a burst.
• The scaler runs next to the HPA (--scaler croft) but is not yet validated against it (issue #9).

Notebooks

Notebook	Question
research/lead_time.ipynb	how far ahead of the HPA and the SLO the pressure signal rises, on steps and ramps
research/burst_signals.ipynb	which syscall tracks requests; reaction time per signal; open-loop bursts