A missed heartbeat marks a node Unknown first, and evicts only after a grace period.

The two-stage timeout stops a brief network blip from rescheduling every Pod on a perfectly healthy node.

Node conditions are separate signals

Ready	kubelet is healthy and can accept Pods
MemoryPressure	low memory: kubelet may evict Pods
DiskPressure	low disk or inodes: eviction possible
PIDPressure	too many processes
NetworkUnavailable	CNI not yet configured, often right after join

inspect a node

$ kubectl describe node gke-np1-x2 | \
    sed -n '/Conditions/,/Addresses/p'
Ready            True   KubeletReady
MemoryPressure   False
DiskPressure     False
PIDPressure      False
$ kubectl get lease -n kube-node-lease

Nodes report in; silence is tolerated, then acted on

Each kubelet periodically reports its node's status and, in modern clusters, renews a small Lease object in the kube-node-lease namespace as a lightweight heartbeat. The node lifecycle controller watches these. If a node goes quiet for roughly forty seconds it marks the node's Ready condition Unknown and adds not-ready and unreachable taints.

Eviction is deliberately slower. Pods carry a default toleration for those taints with a five-minute timeout (tolerationSeconds: 300), so they are evicted and recreated elsewhere only if the node stays silent for that long. This two-stage design avoids reacting to transient blips with disruptive churn. You can shorten the toleration for workloads that must fail over faster, but the trade-off is more needless evictions during network flaps.

Conditions such as MemoryPressure, DiskPressure and PIDPressure are separate signals reported by the kubelet. When one is true the kubelet starts evicting Pods, lowest QoS class first, to protect the node. NetworkUnavailable is commonly true for a few seconds after a node joins, until the CNI is configured.

The animation reveals the four stages along the timeline. Defaults vary by Kubernetes version and distribution, so verify values in your own cluster before relying on specific timings.