Kubernetes v1.36 marks a pivotal evolution in how the platform manages container memory resources. The release introduces substantial updates to the Memory QoS feature, which has been in alpha since v1.22. This new iteration leverages the cgroup v2 memory controller to provide the kernel with precise guidance on container memory treatment. For cloud engineers and DevOps professionals, this shift represents a move from basic throttling to sophisticated, tiered protection mechanisms. Understanding these changes is critical for those pursuing Kubernetes certifications, as they redefine how resource guarantees are enforced at the kernel level.
Decoupling Throttling from Reservation
One of the most significant architectural shifts in v1.36 is the separation of memory throttling from memory reservation. Previously, enabling memory QoS often conflated these two distinct concepts. In this version, the feature gate activation triggers memory.high throttling, where the kubelet sets memory.high based on the memoryThrottlingFactor, which defaults to 0.9. However, memory reservation is now governed by a separate configuration field within the kubelet settings.
When the configuration is set to None, which is the default behavior, the system does not write memory.min or memory.low values. Despite this, throttling via memory.high continues to function as expected. This decoupling allows for more flexible cluster configurations where you might want to limit write speeds without reserving physical memory pages. For example, a developer might throttle a batch processing job to prevent it from consuming all available RAM during peak hours, without guaranteeing that specific amount of memory remains untouched for other processes.
Tiered Protection by QoS Class
The v1.36 release introduces TieredReservation, a mechanism that writes tiered memory protection based on the Pod's QoS class. This feature fundamentally changes how the kernel handles memory pressure for different workload types. Guaranteed Pods receive hard protection via memory.min. If a Guaranteed Pod requests 512 MiB of memory, the kernel ensures this amount is never reclaimed. The system will invoke the OOM killer on other processes if it cannot honor this guarantee.
Conversely, Burstable Pods receive soft protection via memory.low. Using the same 512 MiB request example, a Burstable Pod will see the kernel write a memory.low value, but this does not provide the same level of absolute security as the Guaranteed class. This distinction is vital for architects designing multi-tenant clusters. You can now explicitly define which workloads require absolute memory stability versus those that can tolerate swapping. This granularity is essential for high-performance computing environments where memory pressure can lead to cascading failures.
Observability and Kernel Warnings
Alongside the core protection mechanisms, v1.36 adds observability metrics to help operators monitor memory behavior. These metrics provide visibility into how the cgroup v2 controller is interacting with the kernel. Additionally, the release includes a kernel-version warning for memory.high. This warning alerts administrators if the underlying kernel version does not fully support the new memory reservation features, preventing configuration errors that could lead to unexpected behavior.
For AI engineers and data scientists managing large model inference services, these metrics are invaluable. They allow for real-time adjustment of memory policies based on actual cluster performance rather than theoretical limits. The ability to distinguish between hard and soft protection ensures that critical inference services do not get evicted during memory spikes, while less critical training jobs can utilize available resources more aggressively.
What This Means For You
The introduction of tiered memory protection in Kubernetes v1.36 requires a reevaluation of existing resource request strategies. Engineers must update their deployment manifests to leverage the new memoryReservationPolicy. This update is particularly relevant for those studying for advanced Kubernetes certifications, as it tests the ability to configure complex resource constraints. By adopting these features, organizations can achieve higher cluster density without sacrificing the reliability of their most critical services. The separation of throttling and reservation offers a new dimension of control that was previously unavailable in standard Kubernetes configurations.