Work out cluster autoscaler sizing instantly with clear inputs, formula shown and shareable results.
The cluster autoscaler adds nodes when pods cannot be scheduled, so node count is total requests divided by allocatable capacity — not by nominal instance size, since the kubelet, system daemons and eviction thresholds reserve a meaningful share. Reserving 15 percent for bin-packing inefficiency and adding one spare node covers the common case of a large pod that cannot fit anywhere.
Node count
CPU nodes = ceil(total CPU requests / (allocatable CPU x 0.85)); memory nodes = same for memory; nodes = max + 1 spare
The kubelet reserves memory and CPU for system daemons and eviction thresholds, typically 5 to 15 percent depending on instance size and configuration.
Pods that do not divide evenly into node capacity. A pod requesting slightly over half a node's CPU leaves the remainder unusable for another copy of itself.