CPU and memory requests influence scheduling; limits constrain resource use and can cause throttling or termination when configured poorly. Horizontal autoscaling reacts to metrics, but adding application replicas can worsen database overload if the shared dependency is the bottleneck.

Queue consumers may scale on lag age or queue depth, with controls to avoid rapid oscillation. Pod disruption budgets and topology spread reduce the risk that maintenance removes too much capacity, but they cannot create spare nodes or repair a single shared dependency.

Test node loss, rolling upgrades, startup delays, and scale-out under load. Monitor capacity before and after autoscaling, and ensure connection pools and external quotas are adjusted for the maximum replica count.