Blog
Kubernetes 1.35
‘Timbernetes’ Debuts with AI-Optimized Scheduling and In-Place Pod Resizing
Introduction: The Arrival of ‘Timbernetes’ Contents hide 1 Introduction:
The Arrival of ‘Timbernetes’ 2 The Core Philosophy
Introduction: The Arrival of ‘Timbernetes’
The container orchestration landscape has reached a pivotal milestone with the debut of the Kubernetes 1.35 release, affectionately codenamed "Timbernetes." As enterprises increasingly move toward cloud-native maturity, the demand for efficiency, cost reduction, and intelligent workload management has never been higher. This release specifically addresses the friction points that have historically plagued DevOps teams: the inability to resize pods without restarting them and the static nature of resource scheduling.
Kubernetes 1.35 is not merely an incremental update; it represents a paradigm shift in how clusters manage computational resources. By introducing AI-optimized scheduling and stabilizing In-Place Pod Resizing, the community has delivered a robust platform capable of handling the dynamic needs of modern infrastructure. For a granular breakdown of the changelog, you can review our detailed analysis of kubernetes 1.35 features which highlights every depreciation and addition in this version.
In this comprehensive guide, we will dissect the core features of Timbernetes, explore the technical mechanics of AI-driven placement, and discuss the operational impact of resizing pods on the fly. Whether you are a platform engineer or a CTO, understanding these changes is crucial for maintaining a competitive, resilient infrastructure.
The Core Philosophy of Kubernetes 1.35: Efficiency and Intelligence
The "Timbernetes" moniker reflects a focus on structural integrity and growth. Just as timber provides the framework for construction, Kubernetes 1.35 solidifies the framework for next-generation applications. The primary theme of this release is Resource Intelligence.
Previous versions of Kubernetes relied heavily on static definitions and manual tuning. Operators would define requests and limits, and the kube-scheduler would perform a relatively simple bin-packing exercise. However, real-world applications—especially AI-powered applications—are rarely static. They burst, they idle, and they evolve. Kubernetes 1.35 bridges the gap between static configuration and dynamic reality.
Key Highlights of the Release
- In-Place Pod Resizing (Stable): The ability to change CPU and memory resources without restarting the container.
- AI-Optimized Scheduling (Beta): A pluggable scheduling framework that uses historical usage data to predictively place pods.
- Dynamic Resource Allocation (DRA) Enhancements: improved handling of specialized hardware like GPUs and FPGAs.
- SidecarContainers (GA): Full graduation of sidecar container lifecycle management.
Deep Dive: In-Place Pod Resizing
Perhaps the most anticipated feature in the Kubernetes 1.35 release is the graduation of In-Place Pod Resizing to Stable. For years, the immutability of pod resource specifications was a double-edged sword. While it ensured consistency, it meant that vertical scaling was a disruptive event.
The Problem with Restart-Based Scaling
In previous versions (pre-alpha/beta features notwithstanding), if a Vertical Pod Autoscaler (VPA) determined a pod needed more memory, it had to terminate the pod and schedule a new one with the updated spec. For stateless applications, this was manageable. However, for stateful workloads—such as huge Java heaps, databases, or AI inference engines—a restart meant clearing caches, cold starts, and potential downtime.
This disruption often led organizations to over-provision resources significantly, wasting money to avoid the performance penalty of a restart. This inefficiency is a key reason why many companies seek expert technology consultancy to audit their cloud spend.
How In-Place Resizing Works in 1.35
With Timbernetes, the resources field in the Pod spec becomes mutable for CPU and Memory. When a change is applied:
- API Server Acceptance: The API server accepts the update to the Pod spec without triggering a validation error for immutability.
- Kubelet Actuation: The Kubelet on the node detects the change.
- Cgroup Update: If the resizing is possible (i.e., the node has enough capacity), the Kubelet interacts with the container runtime (CRI) to update the Linux cgroups limits directly.
- Signal Handling: The application receives a signal (optional) indicating resources have changed, allowing it to adjust internal thread pools or cache sizes dynamically.
This capability is a game-changer for custom software development, particularly for Java applications that can now dynamically adjust heap sizes without the JVM crashing or requiring a reboot.
AI-Optimized Scheduling: The Smart Cluster
While In-Place Resizing handles the “how much,” AI-Optimized Scheduling handles the “where.” The standard kube-scheduler has always been efficient, but it lacks context. It looks at the current request and the current node capacity. It does not look at time.
Predictive Placement
Kubernetes 1.35 introduces a scheduling profile that integrates with machine learning models. By analyzing historical metrics (from Prometheus or similar sources), the scheduler can predict the future utilization of a pod.
For example, if a batch job runs every Monday at 9 AM and spikes CPU usage, the AI scheduler knows not to place it on a node that is already hosting a latency-sensitive API gateway that also sees traffic spikes at 9 AM. This "temporal awareness" prevents noisy neighbor problems before they happen.
Integration with High-Performance Workloads
This is particularly critical for ai-powered applications. Training models requires massive throughput. The new scheduler in 1.35 can group pods that communicate frequently on the same rack or availability zone more intelligently than affinity rules alone, optimizing network latency based on predicted traffic flows.
Dynamic Resource Allocation (DRA) and Hardware
As we look toward the app development trends to watch in 2026, specialized hardware is becoming the norm. Kubernetes 1.35 refines the Dynamic Resource Allocation (DRA) API.
DRA moves the resource management of devices (GPUs, TPUs, network accelerators) out of the core Kubernetes code and into drivers. In 1.35, the structured parameters for DRA allow for more complex requests. A pod can now request "a GPU that shares memory with this other specific GPU," facilitating complex AI training topologies.
Operational Impacts and Cost Savings
The combination of these features in the Kubernetes 1.35 release leads to direct operational benefits.
Reducing the “Safety Margin”
DevOps teams typically set resource requests 20-30% higher than actual usage to account for spikes. With functional In-Place Resizing, this safety margin can be reduced. The VPA can reactively increase limits without disruption, allowing for higher node density. Higher density means fewer nodes, which translates to a lower monthly cloud bill.
Mitigating Performance Issues
By preventing unnecessary restarts and ensuring intelligent placement, 1.35 improves overall application stability. This stability is crucial when comparing containerized architectures to other models; effectively managed K8s clusters can now rival or exceed the reliability of managed functions, helping teams avoid common serverless performance issues such as cold starts and execution timeouts.
Migration Strategy: Moving to Timbernetes
Upgrading to Kubernetes 1.35 requires planning. While the API is backward compatible, the new features often require enabling specific feature gates if they are not yet fully GA in your specific distribution (though Resizing is Stable, specific AI scheduler plugins may be Alpha/Beta).
Steps for a Smooth Upgrade
- Audit API Usage: Check for removed APIs. 1.35 cleans up several beta APIs from the 1.31-1.34 era.
- Upgrade CNI/CSI Drivers: Ensure your Container Network Interface and Storage Interface drivers support the new resizing calls.
- Node Pool Strategy: Roll out the upgrade one node pool at a time.
- Enable VPA in “Auto” Mode: Once on 1.35, switch your Vertical Pod Autoscalers to “Auto” (or “In-Place” mode) to begin testing non-disruptive scaling on staging workloads.
For complex environments, leveraging custom software development expertise to build upgrade automation pipelines is highly recommended to prevent downtime.
The Future of Platform Engineering
Kubernetes 1.35 ‘Timbernetes’ is a clear signal that the platform is shifting from "day 1" setup challenges to "day 2" operational excellence. The focus is no longer just on getting containers running, but on keeping them running efficiently, cheaply, and with high availability.
The introduction of AI into the scheduling logic opens the door for self-healing clusters that are truly autonomous. We are moving toward a future where human operators define high-level SLOs (Service Level Objectives), and the Kubernetes control plane—empowered by AI—makes the micro-decisions on resource allocation to meet those goals.
Frequently Asked Questions
What makes the Kubernetes 1.35 release different from 1.34?
The primary differentiator in 1.35 is the stabilization of In-Place Pod Resizing. While 1.34 focused on security hardening and sidecar improvements, 1.35 focuses on dynamic resource mutability and the integration of AI-driven scheduling logic to optimize hardware usage.
How do I enable AI-Optimized Scheduling in Kubernetes 1.35?
AI-Optimized Scheduling is typically implemented via the Scheduling Framework plugins. You will need to deploy a scheduler plugin that interfaces with a metrics provider (like Prometheus) or an external ML model server. It is not a single “switch” but a configuration of the scheduler profile to use a scoring plugin based on predictive models.
Will In-Place Pod Resizing work with my existing Java apps?
Yes, but with caveats. While Kubernetes will update the cgroup limits, the Java Virtual Machine (JVM) inside the container must be aware of the new limits. Modern JVMs (Java 17+) handle container visibility well, but you may need to configure the application to listen for resizing events to expand its heap usage dynamically without a restart.
Does Kubernetes 1.35 reduce cloud costs?
Yes, significantly. By allowing pods to resize without restarting, you can run with tighter resource requests (higher bin-packing density) and rely on the Vertical Pod Autoscaler to handle spikes. This reduces the need for over-provisioning, allowing you to run the same workload on fewer nodes.
Is ‘Timbernetes’ an official name?
Release names in Kubernetes are chosen by the release team and often follow a theme. ‘Timbernetes’ is the moniker for version 1.35, symbolizing the structural maturity and ‘growth rings’ of the project as it solidifies its position as the foundation of modern cloud computing.
Conclusion
The Kubernetes 1.35 release marks a turning point for container orchestration. By solving the long-standing challenge of in-place resizing and embracing the potential of AI scheduling, Timbernetes offers a platform that is not only robust but adaptable. For organizations striving to innovate, these features provide the agility needed to deploy complex, data-intensive applications without spiraling infrastructure costs.
As you plan your roadmap, consider how these features align with broader app development trends to watch in 2026. The convergence of AI and infrastructure is here, and Kubernetes 1.35 is the vehicle driving it forward. Whether you are refactoring legacy monoliths or deploying cutting-edge microservices, upgrading to 1.35 provides the tools to do so with greater precision and efficiency.
Editor at XS One Consultants, sharing insights and strategies to help businesses grow and succeed.