Work out autoscaling threshold instantly with clear inputs, formula shown and shareable results.
An autoscaler must trigger early enough that new capacity arrives before the target is breached. If traffic grows 25 percent a minute and an instance takes two minutes to become ready, load rises about 56 percent during startup, so the threshold has to sit well below the target. Long startup times force low thresholds, which is why image and warm-up optimisation is a capacity lever.
Scale-out threshold
growth during startup = (1 + per-minute growth)^(startup minutes) - 1; threshold = target / (1 + growth during startup)
It is dead time during which load keeps rising. Halving startup roughly halves the safety margin you must give up in the threshold.
Use a cooldown after scaling, asymmetric thresholds so scale-in is more conservative than scale-out, and scale in by one instance at a time.