Quantify a host failure: VMs affected, capacity lost, post-failure utilisation, restart window and the availability cost.
Two independent things happen when a host dies. Capacity redistributes onto the survivors, and because the denominator shrinks the utilisation rise is non-linear — losing one host of eight at 65% pushes survivors to 74%, but losing two pushes them to 87%. Separately, HA has to detect the failure, elect a master and restart the VMs in batches, which is what the users actually experience as downtime. HA guarantees a restart, not a fast one: a hundred VMs restarting 32 at a time with a 45-second application start is over two and a half minutes of outage before detection is even counted. Substitute your own cost-of-downtime figure; the default is a placeholder.
Host Failure Impact
VMs affected = ceil(VMs ÷ hosts) × hosts lost; post-failure utilisation = current utilisation × hosts ÷ surviving hosts; restart window = detection time + ceil(VMs affected ÷ concurrent restarts) × restart time.
VMs affected = ceil(VMs ÷ hosts) × hosts lost; post-failure utilisation = current utilisation × hosts ÷ surviving hosts; restart window = detection time + ceil(VMs affected ÷ concurrent restarts) × restart time. Two independent things happen when a host dies. Capacity redistributes onto the survivors, and because the denominator shrinks the utilisation rise is non-linear — losing one host of eight at 65% pushes survivors to 74%, but losing two pushes them to 87%. Separately, HA has to detect the failure, elect a master and restart the VMs in batches, which is what the users actually experience as downtime.
HA guarantees a restart, not a fast one: a hundred VMs restarting 32 at a time with a 45-second application start is over two and a half minutes of outage before detection is even counted. Substitute your own cost-of-downtime figure; the default is a placeholder.
This calculator takes 8 inputs: Hosts in cluster, VMs in cluster, Hosts lost, Cluster utilisation before the failure, Restart time per VM, VMs restarted concurrently, Failure detection and election time, Cost of downtime per hour. The pre-filled defaults are a realistic starting point — replace them with figures from your own environment for a result you can act on.
Losing one host of eight removes 12.5% of capacity, but the same workload now divides across seven hosts instead of eight — a 14.3% increase in per-host load. The effect compounds with each additional host lost, which is why small clusters need proportionally more headroom.