The latest Kubernetes release, version 1.36, significantly refines its Memory QoS feature by introducing a tiered approach to memory reservation and protection, separating throttling from reservation, and enhancing observability. These updates aim to optimize cloud infrastructure reliability and resource utilization while improving developer workflow around pod memory management.
- Tiered memory reservation offers hard guarantees for Guaranteed Pods and soft protection for Burstable Pods.
- Memory throttling and reservation are configurable independently for flexible resource management.
- New observability metrics aid in capacity planning; kernel version warnings improve stability awareness.
Infrastructure signal
Kubernetes v1.36 updates the Memory QoS feature to separate memory throttling from reservation, allowing more precise control over how container memory is managed at the kernel level. Using cgroup v2 memory controllers, the kubelet now sets memory.min for Guaranteed Pods to provide a hard memory reservation, while Burstable Pods receive a softer memory.low reservation that can be reclaimed under extreme pressure. BestEffort Pods remain fully reclaimable without memory reservation.
This tiered memory protection prevents excessive hard reservation on large Burstable workloads that could previously lock too much memory, leading to system instability. By maintaining lower hard reservation footprints and allowing the kernel to reclaim memory from soft guarantees under high pressure, cluster nodes can maintain better overall memory utilization and reduce likelihood of node-wide out-of-memory (OOM) events.
Developer impact
Developers benefit from configurable memoryReservationPolicy as it allows opting into memory protection incrementally. The policy 'None' disables memory reservation while still allowing memory.high throttling, facilitating controlled testing of throttling without risking hard memory locks. Selecting 'TieredReservation' enables differentiated hard and soft memory reservations by pod QoS class, improving pod resilience by ensuring Guaranteed Pods retain critical memory while still giving flexibility to Burstable Pods.
Additionally, the kubelet now exposes two alpha observability metrics—tracking memory.min and memory.low usage at the node level—directly supporting workload capacity planning and debugging memory pressure scenarios. This observability makes it easier to monitor memory guarantees and prevent resource overcommitment that can affect deployment reliability.
What teams should watch
Platform engineers and cluster operators should evaluate kernel versions on their nodes because the memory.high throttling feature can cause livelock on Linux kernels older than 5.9. Kubernetes v1.36 includes startup warnings for nodes running these older kernels, enabling teams to prioritize kernel upgrades for improved stability.
Teams responsible for workload deployment should monitor the new kubelet memory QoS metrics closely, particularly kubelet_memory_qos_node_memory_min_bytes, to avoid node memory saturation from large hard reservations. Updating deployment resource requests and QoS classifications in conjunction with the new tiered memory policies can optimize cluster resource allocation and minimize OOM risk.