This is the multi-page printable view of this section. Click here to print.

Return to the regular view of this page.

Kubernetes Blog

Scheduling the removal of the kube-proxy IPVS backend

The ipvs backend for kube-proxy was an experiment in providing a service proxy implementation with better performance than the iptables backend. While it succeeded in that goal, the kernel IPVS API has turned out to be a bad match for the Kubernetes Services API, and the ipvs backend was never able to implement all of the edge cases of Kubernetes Service functionality correctly.

In Kubernetes 1.29, the project introduced the nftables backend for kube-proxy, which became GA in 1.33. Part of the plan behind adding the nftables backend was the idea that it would eventually replace both the iptables and ipvs backends. Because the ipvs backend in particular is complex, not 100% feature complete, and architecturally quite different from the other two backends (and thus more work to maintain), Kubernetes SIG Network decided to officially deprecate it as of Kubernetes 1.35. Although it currently remains available in kube-proxy (with a warning about its deprecation), we (Kubernetes SIG Network) now plan to remove it in Kubernetes 1.40. All ipvs users should therefore be planning to migrate to either the nftables or iptables backends.

Why was the ipvs backend written?

The ipvs backend was originally designed to fix two different kinds of performance problems in the iptables backend:

  • First, iptables had problems with data plane latency (the speed at which clients could establish new connections to services), because kube-proxy would write a separate iptables rule to match each service IP, so every new connection would need to be matched against a long series of rules that was O(n) in the number of services in the cluster.

  • Second, it had an even larger problem with control plane latency (the speed at which kube-proxy could process changes to services and endpoints and update the data plane rules), because kube-proxy would rewrite the entire iptables ruleset every time any service or endpoint changed. While the list of rules itself was only O(n) in the total number of endpoints, the amount of time, RAM, and CPU that it took to process the updated ruleset grew faster than that. In very large clusters, kube-proxy would use 100% of one CPU all the time just to process iptables updates. (In fact, the control plane scaling problem was bad enough that for most users it was not even possible to grow the cluster to the size where the data plane scaling problem became noticeable.)

The ipvs backend solved both of these scaling problems. The data plane problem was solved mostly automatically, because the kernel IPVS subsystem uses hash tables rather than linear lookup. The control-plane problem was solved by not rewriting the entire ruleset on every update. (The kernel IPVS API doesn’t have a way of doing batch updates anyway, so this was essentially forced.)

Why is the ipvs backend now being deprecated?

Almost from the start, the ipvs backend also had problems.

The first version didn’t have 100% compatibility with the iptables backend, but it was assumed that this would be fixed over time. Unfortunately, the original authors moved on to other things outside the Kubernetes ecosystem (as inevitably sometimes happens in open source), and it took a while to find new people interested in maintaining the code.

Meanwhile, as people began using the ipvs backend in larger and larger clusters, its own scaling problems started to emerge. In particular, the lack of batch update operations in the IPVS API meant that while the ipvs backend was faster than iptables when making a few small changes, it was potentially slower when making a large number of changes at once. And because the IPVS API is fairly limited in what it can do, kube-proxy's ipvs backend needed to fall back to using iptables to implement various features of Kubernetes Services beyond basic "load balancing", making it run into some of the same scaling problems the iptables backend has. While the ipvs backend tried to mitigate this by using ipset for some of its rules, this again ran into the problem of not having a batch update API. Additionally, the fact that the ipvs backend uses iptables means that it has the same long-term support problem as the iptables backend: Red Hat's plan to remove the iptables API entirely from Red Hat Enterprise Linux means the ipvs kube-proxy backend will also no longer be available to Kubernetes users on RHEL.

What about IPVS schedulers?

Beyond performance, the other reason some users were excited about the ipvs backend was the ability to pick from among various IPVS "schedulers" for distributing traffic to endpoints, rather than relying on the simple random selection used by iptables (and nftables).

Unfortunately, the way IPVS is used by kube-proxy prevents it from being able to use most of the IPVS schedulers in a useful way. The problem is that kube-proxy is a distributed service that runs on all nodes in the cluster, but the kube-proxy processes on different nodes do not share their IPVS scheduler state, so each node picks endpoints without knowing what is happening with clients on any other node. Even the schedulers that are designed to work in a distributed way generally don't work better than random selection in the context of kube-proxy. For example, the mh scheduler uses Maglev Hashing to select endpoints, which theoretically allows for extremely efficient distributed load balancing. But that efficiency is only in the context of a specific sort of two-layer load-balancer design, and kube-proxy does not have the right architecture to be able to act as either of the two layers in that design.

Switching away from ipvs

Switching to nftables

If you are running Kubernetes 1.36 or later, and your nodes are running a Linux distribution with kernel 5.13 or later, and you use the ipvs proxy mode, then Kubernetes recommends switching to the nftables mode of kube-proxy. (Although the nftables backend is GA —and fully usable—as of Kubernetes 1.33, there have been some performance- and scaling-related fixes since then that that weren't backported to releases earlier than 1.36.)

The nftables backend is fully supported, feature-complete, and even has very slightly better performance than ipvs:

kube-proxy ipvs-vs-nftables first packet latency, at various percentiles, in clusters of various sizes

(Taller bars represent longer latency. The "p50" group is the average connection establishment latency. Note that even the largest numbers in this graph are quite tiny; the point is not to claim that nftables is (noticeably) faster than ipvs, just to show that, unlike iptables, it's not slower.)

Switching is as easy as changing the kube-proxy configuration and restarting kube-proxy. For example, in clusters deployed via kubeadm, change mode: ipvs to mode: nftables in the kube-proxy ConfigMap and then restart the kube-proxy pods, as described in the documentation on reconfiguring a kubeadm cluster. (If you are overriding other ipvs-specific kube-proxy configuration, you will also need to change those options to the corresponding nftables ones.)

Switching to iptables: "It's Not As Slow As You Think!"

If you can't switch to nftables, either because you are on an older version of Kubernetes, or because your cluster runs on an older Linux distribution that isn't new enough to support the nftables backend, you may be able to just switch to the iptables backend, which has improved greatly since the original creation of the ipvs backend in 2017:

  • The generated ruleset has been optimized to remove redundant and irrelevant rules (#57461, #60306, #96959, #108251), reducing data plane latency.

  • The code that generates the ruleset has been optimized to do fewer allocations, and to avoid unnecessarily recomputing the same information multiple times (#65755, #65902, #67948, #85617, #90103, #97238), reducing CPU and memory usage when resyncing.

  • Rather than doing “check the current state of the rules, then update them to the new state”, iptables kube-proxy now does “assume the current state is correct, update to the new state based on that assumption, and only explicitly check the current state if that update fails”. This greatly improves the resyncing speed, because it turns out that even reading the iptables ruleset is extremely slow in large clusters (#81517, #110334, #114181).

  • To the extent possible within the constraints of the iptables APIs, kube-proxy’s iptables backend now only updates iptables rules for services/endpoints that have changed since the last sync, rather than rewriting the entire ruleset every time (#110268 / KEP-3453).

The net result of this is that the iptables backend of 2026, while still not as good as the ipvs backend, is at least much better than the iptables backend of 2017 was, and many people who found iptables unusable in the past would not have problems with it now.

As with switching to nftables, switching to iptables is mostly just a matter of updating your kube-proxy configuration, although for technical reasons, in this case we recommend that you reboot each node after changing the kube-proxy configuration, rather than merely restarting kube-proxy.

Learn more

  • KEP-5495 gives the full timeline and plan for kube-proxy IPVS deprecation.

  • The blog post "NFTables mode for kube-proxy" has more information about nftables mode and its performance relative to ipvs and iptables.

From Working Group to SIG Architecture: spotlight on AI Conformance

In this SIG Architecture spotlight we talked with Janet Kuo (Google) and Yuan Tang (TensorScale AI) about the journey of AI Conformance in Kubernetes, from the initial Working Group to the integration SIG Architecture as a subproject.

Introductions

Frederico Muñoz and Kirti Goyal (FK): Hi Janet, Yuan, and thank you for the opportunity to learn a bit more about AI Conformance. Can you tell us a bit about yourselves and how you got involved in Kubernetes?

Janet Kuo: I'm a Senior Staff Software Engineer at Google and have been a Kubernetes maintainer since 2015, joining the community just as we were racing toward the 1.0 launch. I started my journey writing the core workload controllers that standardized how applications are orchestrated on Kubernetes. Today, I lead SIG Apps as a Co-chair & TL, and also lead the AI Conformance program. My current focus is to ensure Kubernetes remains the best platform for the next generation of workloads, such as AI.

Yuan Tang: I'm a co-founder of TensorScale AI where we build AI infrastructure for world models. Prior to that, I was a Senior Principal Software Engineer at Red Hat, where I led the development of Red Hat AI platform and upstream open source strategy. I became involved in Kubernetes early on by fostering collaborations between the Kubernetes community and projects I was leading, including Kubeflow, Argo, and KServe. I later helped found and lead WG Serving and the AI Conformance program, advancing innovation and strengthening community collaboration around running AI workloads on Kubernetes.

Program, Working Group, Subproject: the many shapes of AI Conformance

FK: AI Conformance started as a CNCF Program, was it then you got involved?

Janet Kuo: I was actually involved from day one since I pitched the original idea. It became obvious then that everyone was starting to run AI on Kubernetes, but vendors were building their infrastructure stacks in different ways. I got worried about ecosystem fragmentation. So I wrote up the initial Kubernetes AI Conformance Proposal, took it to the CNCF leadership to get their buy-in, and started the AI Conformance Program. Once we had the green light, I reached out to my fellow co-chairs Yuan Tang, Rita Zhang, and Mario Fahlandt to get the community rallied around it.

Yuan Tang: Yes. I was a co-chair of WG Serving at that time and some of the initiatives such as the Gateway AI Inference Extension quickly gained traction. There was a strong interest among the community to include inference-related requirements in the Kubernetes AI Conformance program.

FK: How would you describe the motivation behind the creation of the AI Conformance Working Group?

Janet and Yuan: The motivation behind creating the Kubernetes AI Conformance Working Group (AI WG) was to address the rapid convergence of AI workloads onto Kubernetes while preventing ecosystem fragmentation.

As AI adoption accelerated, organizations began running increasingly complex workloads, including distributed training, large-scale inference, GPU-intensive pipelines, and emerging agentic AI systems across many different Kubernetes platforms. However, vendors were implementing AI infrastructure capabilities differently, especially around GPUs, networking, scheduling, and inference orchestration. That created portability challenges, operational inconsistency, and growing vendor lock-in concerns.

The AI WG was formed to bring the community together around:

  • vendor-neutral AI infrastructure standards,
  • interoperable Kubernetes-based AI platforms,
  • consistent operational behavior for AI workloads,
  • and common requirements for scalable AI deployment.

FK: WGs are temporary by nature, based on outcomes: would you say the main goals of the WG were achieved?

Janet and Yuan: Yes, the WG achieved its primary initial goals.

The WG has succeeded in moving the ecosystem from: fragmented, vendor-specific AI infrastructure approaches to a shared, community-defined baseline for AI workloads on Kubernetes.

Some concrete indicators of success include:

  • Launching the Certified Kubernetes AI Conformance Program
  • Defining formal Kubernetes AI Requirements (KARs)
  • Getting broad participation from major ecosystem players
  • Growing certified platforms from 18 to 31 within a few months
  • Aligning around newer Kubernetes primitives needed for AI workloads
  • Establishing a roadmap for automated conformance verification

Most importantly, the WG created consensus that AI workloads are now a first-class Kubernetes use case.

KubeCon stage with 4 persons in it, one of them talking.

FK: AI Conformance was specifically highlighted at the latest KubeCon: is this a reflection of the growing importance of AI throughout the project?

Janet and Yuan: Yes, the prominence of AI Conformance at the latest KubeCon + CloudNativeCon is a strong reflection of how central AI has become to the Kubernetes ecosystem overall.

A few years ago, AI workloads on Kubernetes were still viewed as somewhat specialized, mostly tied to ML platforms or GPU-heavy research environments. What’s changed is that AI is now becoming a mainstream infrastructure concern for nearly every large-scale platform operator.

That shift naturally elevates AI from an application running on Kubernetes to a core workload category Kubernetes must optimize for.

The AI Conformance initiative reflects that evolution. It signals that the community now sees AI infrastructure requirements, things like accelerator scheduling, high-performance networking, workload-aware orchestration, and low-latency inference, as important enough to standardize at the ecosystem level.

Its visibility at KubeCon also shows a broader strategic realization: if Kubernetes is going to remain the universal control plane for modern infrastructure, it needs to provide a consistent foundation for AI workloads across clouds, on-prem, and edge environments.

So the emphasis on AI Conformance is not just about chasing AI hype; it's about Kubernetes adapting to a major shift in workload patterns and ensuring the ecosystem evolves cohesively rather than fragmenting into vendor-specific AI stacks.

(slide) Initial adopters and major contributors: Akamai (logo), Alibaba Cloud (logo), AWS (logo), Broadcom (logo), CoreWeave (logo), DaoCloud (logo), Google Cloud (logo), Kubermatic (logo), Microsoft Azure (logo), NVIDIA (logo), Oracle (logo), SUSE (logo), Red Hat (logo), SAP (logo)

From Working Group to SIG Architecture subproject

FK: How would you describe the change from WG to SIG? Is it a reflection of AI as something that requires ongoing architectural governance?

Janet and Yuan: In Kubernetes, Working Groups are meant to be temporary. Our WG had a specific goal to get the initial Kubernetes AI Conformance requirements published and get the certification process off the ground. Once we hit those milestones, we archived the WG and moved AI Conformance to be a permanent subproject under SIG Architecture.

Moving to SIG Architecture definitely proves that AI needs ongoing governance. AI isn’t just a passing trend or a one-off workload. It has become a core part of what Kubernetes runs every day. Because SIG Architecture already owns overall Kubernetes Conformance and API design principles, housing the AI Conformance subproject there makes perfect sense. It makes sure our AI Conformance requirements stay aligned with the rest of the Kubernetes ecosystem for the long haul.

FK: And with that change, how would you describe the main goals and areas of intervention of the subproject?

Janet and Yuan: Now that we’re a subproject, our main goal is to keep the AI Conformance standards current, vendor-neutral, and standard-setting. We want to make sure Kubernetes captures the latest advancements in AI infrastructure while keeping the ecosystem fair for everyone. When we find gaps where no clear community pattern exists, we help foster and form new standards to solve them.

AI Conformance and Kubernetes

FK: AI Infrastructure is evolving very quickly, while Kubernetes APIs are designed to remain stable for years. Is this something you consider a challenge?

Janet and Yuan: It’s definitely a tricky balance. AI moves very fast, while Kubernetes is designed to be stable and boring. But we actually structured the program around this exact tension.

First, we use a tiered graduation model. When new AI infrastructure concepts emerge, we introduce them as recommended SHOULD requirements. That gives the ecosystem room to experiment and agree on best practices without locking Kubernetes into rigid APIs too early. Once those APIs stabilize and with ecosystem adoption, we graduate them to mandatory MUST requirements.

Second, we intentionally keep our focus on the infrastructure layer. We test for core capabilities like accelerator support through Dynamic Resource Allocation (DRA), rather than getting bogged down in specific AI frameworks or model servers. That keeps the base platform stable while letting the AI tooling evolve as fast as it needs to.

FK: DRA has evolved, also due to the work of WG Device Management, as an important part of Kubernetes (and especially so for AI workloads), with SIG Architecture having an important role in how it’s integrated in Kubernetes: do you find any parallel in this? Specifically, do you think there will be a heavier focus in the integration with the evolving Kubernetes architecture as part of SIG Architecture?

Janet and Yuan: There is a huge parallel there. Whenever we face a massive, complicated challenge that touches multiple parts of the codebase and involves multiple SIGs, a Working Group is the best place to start. It lets experts from different SIGs get together to figure out the problem, and build consensus.

Being part of SIG Architecture means there will be a much heavier focus on how AI capabilities weave into the evolving Kubernetes architecture. Instead of treating AI as an add-on layer, we’re looking at how core Kubernetes primitives, like scheduling and resource allocation, need to adapt at a foundational level to support the next decade of AI workloads.

Final remarks

FK: For interested contributors, what’s the best way to get involved with the AI Conformance subproject?

Janet and Yuan: We're always looking for new contributors. The easiest way to get started is by checking out our main repo at kubernetes-sigs/ai-conformance and join our meetings. If you're interested in defining new AI Conformance requirements, you can jump into PR reviews or propose your own using our KAR template. If you want to write code, we always need help expanding our automated Go test suites in the test/ directory.

For vendors looking to certify their platforms, all the details are at cncf/k8s-ai-conformance.

FK: Any final comments you would like to share?

Janet and Yuan: We’ve been blown away by how fast the community has rallied around this. When we first announced the AI Conformance program during the keynote at KubeCon NA 2025, we weren’t sure what the response would be. But by the time we gave our update keynote at KubeCon EU 2026, we had grown from 18 certified platforms to 31 in just a few months.

That kind of growth shows just how much the industry wanted a unified, vendor-neutral standard. It gives organizations the confidence that they can run their AI workloads on any certification Kubernetes platform without getting locked into a proprietary stack.

If you’re a platform vendor, we’d love to see you get certified. If you’re a developer interested in the intersection of AI and cloud-native infrastructure, come join us in SIG Architecture!

FK: Excellent, thank you both for guiding us in this journey on AI and AI Conformance in Kubernetes.

Announcing the Workload Conformance Working Group

We are excited to announce that the Kubernetes community now has a Workload Conformance Working Group. The working group is writing a shared standard for how workloads behave at runtime on Kubernetes. To get it right, we need input from the people who build those workloads and from the people who run the clusters they're deployed on.

Why workloads need a shared standard

Certified Kubernetes tests the cluster, and Kubernetes AI Conformance tests whether a cluster can run AI and ML workloads. Nothing comparable exists for the workloads themselves, so cluster operators still check each application by hand before routine maintenance or a Kubernetes upgrade.

The problems they find are familiar. A routine node drain is often how you discover that an application can't tolerate disruption. An upgrade can break an application that still depends on a removed API version. Others show up in day-to-day operation: workloads that ask for privileges they don't need, or reserve resources they never use.

Kubernetes already documents much of this, including security and handling disruptions. Even so, every cluster operator ends up repeating the same assessment, and application maintainers get the same questions from each environment they support. A common specification with repeatable tests would give both groups the same baseline to start from.

What the working group will deliver

The charter sets out three deliverables:

  • A specification of workload behavior for each Kubernetes minor version.
  • A method for testing each requirement against a running workload.
  • Upstream conformance tooling that implements those tests.

The first step is deciding which workloads and behaviors the specification covers. The scope includes security, resilience during disruptions, resource use, networking, storage, API usage, and observability, but the group hasn't yet agreed on specific requirements or on how to test them. However, one rule is already set: every requirement needs an objective test that a test binary can run, so nothing subjective goes into the specification.

The specification covers how a workload runs, and importantly not what it does. Consequently, application logic is out of scope of workload conformance. The binary won't retest anything Certified Kubernetes already covers, and it won't require a particular workload packaging format, or vendor tool like policy engine, security scanner, etc.

How the work is organized

Two Special Interest Groups (SIGs) are sponsors for the working group.

  • SIG Architecture will own the specification and tooling through a subproject, which will be created once the first artifacts are proposed by the Working Group.
  • SIG Apps will bring the application perspective.

Whenever a requirement touches another Kubernetes SIG's area, such as node behavior or networking, the working group will initiate conversations with those SIGs as well.

The working group is a forum where SIGs can get involved and shape the direction of Workload Conformance while the specification is being written. We're aiming for a stable first specification and working tests within two to three Kubernetes release cycles.

Get involved

The specification will and needs to be built from real examples, so we want to hear your experiences. If you run clusters, tell us how you assess workloads today and what broke during your last upgrade or maintenance window. If you maintain an application, tell us what it needs from the cluster and where the existing guidance left you guessing.

  • Beyond examples, the work needs reviewers for proposed requirements and contributors to write the tests and tooling.
  • When early versions of the tests are ready, we also want early testers to run the binary against your own workloads and tell us where the results differ between runs or environments.
  • If a requirement touches your SIG's area, help us check its scope and the test behind it.
  • And because the subproject will outlive the working group, we're also looking for people who want to maintain it for the long term.

You can find us in #wg-workload-conformance on Kubernetes Slack (visit slack.k8s.io for an invitation) and on the mailing list. The working group meets every other Wednesday at 15:00 UTC. The Zoom link and the agenda are in the meeting notes. If you want to raise a topic, add it to the agenda. Joining the mailing list usually adds the meeting invite to your calendar.

Thanks to everyone who reviewed and supported the working group proposal.

Why Is My Kubernetes Job Still Pending? Tracing Kueue Admission with Headlamp

You submit a Kubernetes Job and wait for its Pods to appear. A minute passes. Then another. The Job is still suspended, there are no application logs, and there may not even be a Pod to inspect.

In a cluster managed by Kueue, this can be completely normal. Kueue may still be deciding whether the Job can use the requested resources. The useful question is not only “Why is the Pod pending?” It is first “Has Kueue admitted the Workload?”

The Kueue plugin for Headlamp brings that investigation into the Headlamp UI. Instead of reconstructing the admission path from several commands and large YAML objects, you can begin with the Workload, read its condition message, and follow links through the LocalQueue, ClusterQueue, ResourceFlavor, and Cohort.

This article follows that Headlamp-first path. The command line appears only in a short optional section for reproducing the example.

Before you begin

You need a Kubernetes cluster with Kueue installed, a Headlamp installation, and the 0.1.0-alpha Kueue plugin. Your Kubernetes identity must be allowed to read the Kueue resources used in this article.

The alpha release provides read-only views for Kueue v1beta2 Cohorts, ClusterQueues, LocalQueues, ResourceFlavors, and Workloads. It does not add its own create, edit, stop, resume, or delete workflows. Headlamp still provides its standard YAML and resource actions where your permissions allow them.

First, separate admission from Pod scheduling

Kueue and the Kubernetes scheduler answer different questions.

  • Kueue admission: Should this Workload start now, considering queue policy, quota, ResourceFlavors, admission checks, and other waiting Workloads?

  • Pod scheduling: After the Workload is admitted and the Job resumes, which nodes can run its Pods?

That boundary changes the investigation. If the Workload is not admitted, stay in the Kueue section of Headlamp. If it is admitted but its Pods are still Pending, move to Headlamp's normal Job and Pod views and inspect scheduler events.

A Kubernetes Job becomes a Kueue Workload. Kueue evaluates its LocalQueue, ClusterQueue, ResourceFlavor, and optional Cohort before resuming the Job. Headlamp reads these resources through the Kubernetes API.

Kueue admission architecture

The Headlamp-first troubleshooting path

The quickest investigation starts with the Workload, not with a shell.

A Headlamp troubleshooting flow that starts with Kueue Workloads, checks admission conditions, follows LocalQueue and ClusterQueue links when admission is blocked, and switches to Job and Pod views after admission.

Tracing a pending Job through Headlamp

Step 1: Open the Workloads view

In Headlamp, open Kueue → Workloads. Select the namespace that contains the Job, then find the Workload created for it.

Kueue creates a Workload to represent the resource request and admission state of a Job. In the alpha plugin, owner references are shown as Kind/name text but are not yet clickable. The Workload name normally contains the Job name, so the namespace filter and table search are the simplest ways to locate it.

Before opening the detail page, the list already gives you a useful first scan:

  • LocalQueue

  • priority and priority class

  • active state

  • admitted state

  • completion state

  • a readable current status such as Admitted, Finished, or Inadmissible

  • age

If there is no matching Workload, check three things before going further: the selected namespace, whether the Job was submitted to a LocalQueue, and whether your Kubernetes account can list Workloads. The plugin performs permission checks and displays an access message when the current identity cannot read a resource.

Step 2: Read status, conditions, and events

Open the Workload. Start with its current status and conditions, not with the full YAML.

For admission troubleshooting, the most important condition types are QuotaReserved and Admitted.

  • Admitted=True means Kueue has admitted the Workload. The Job can resume, so any remaining Pending Pods belong to the scheduling branch of the investigation.

  • QuotaReserved=False with a reason such as Pending means Kueue has not reserved quota. The condition message explains what prevented the reservation.

  • An Inadmissible status means the Workload currently cannot fit the available queue configuration. A missing covered resource, unsuitable ResourceFlavor, stopped queue, or admission check can all lead you here.

For example, a condition message may report that Kueue could not assign a ResourceFlavor because the ClusterQueue has insufficient CPU quota. That message is the first real answer to “Why is my Job still pending?” The rest of the Headlamp journey explains the configuration behind it.

Also check the Events section. Conditions describe the current state; events help show what recently changed or what Kueue attempted.

Step 3: Compare the Workload's request with its assignment

Stay on the Workload detail page and inspect the PodSets and resource requests. The plugin shows the requested Pod count, CPU, memory, and other resources that Kueue evaluates together.

This matters because a Job with parallelism: 3 and a one-CPU request per Pod asks Kueue for three CPUs in total. Looking at only one Pod's template would underestimate the Workload.

If quota has already been reserved, the same page also shows the assigned ClusterQueue, ResourceFlavors, and admitted resource usage. If those admission assignments are absent, the Workload has not reached that point yet.

In our example, the Workload requests:

  • three Pods

  • three CPUs in total

  • 600 MiB of memory in total

The next question is where that request entered Kueue.

Step 4: Follow the linked LocalQueue

Use the LocalQueue link on the Workload detail page. A LocalQueue is the namespace-level entry point used by the Job. It connects namespaced Workloads to a cluster-scoped ClusterQueue.

The LocalQueue page shows:

  • its namespace and connected ClusterQueue

  • stop policy and current status

  • pending, reserving, and admitted Workload counts

  • Kueue conditions and Kubernetes events

  • related Workloads in the same namespace

This page answers two practical questions: Did the Job enter the queue you expected, and is it the only Workload waiting there?

If the LocalQueue is inactive or stopped, that may already explain the wait. If it is active, follow the linked ClusterQueue to inspect the capacity behind it.

Step 5: Inspect ClusterQueue capacity and usage

A ClusterQueue represents the cluster-level quota and admission policy available to one or more LocalQueues.

On its Headlamp detail page, compare the Workload request with:

  • covered resources

  • nominal quota for each ResourceFlavor

  • current reservations and usage

  • borrowing and lending limits

  • pending, reserving, and admitted Workload counts

  • queueing strategy and stop policy

  • preemption, fair-sharing, flavor-fungibility, and admission-check settings

  • Kueue conditions and Kubernetes events

For the example in this article, the ClusterQueue provides a nominal quota of two CPUs and two Pod slots. The Workload asks for three of each. The condition message and the ClusterQueue page now tell the same story: the Workload does not fit, so Kueue leaves the Job suspended.

This is an admission-capacity problem. It is not evidence that kube-scheduler failed to find a node.

Step 6: Check ResourceFlavor and Cohort when needed

Quota is not always a single pool. A ResourceFlavor can represent a class of compute through node labels, taints, tolerations, or topology settings. Use the ResourceFlavor link from the ClusterQueue when the condition mentions flavor assignment or when the Workload requires a particular hardware class, such as GPU nodes.

If the ClusterQueue belongs to a Cohort, follow that link too. Cohorts let ClusterQueues share unused quota. The Headlamp page shows parent and child Cohorts, member ClusterQueues, referenced ResourceFlavors, resource groups, borrowing and lending limits, and fair-sharing information.

These pages help answer a more subtle question: Is the Workload blocked because the queue has no quota of its own, because it cannot borrow enough, or because no suitable flavor can be assigned?

What we learned without leaving Headlamp

For the sample pending Job, the UI gives us a complete explanation:

  1. The Workload is not admitted.

  2. Its QuotaReserved condition is false and the message points to insufficient quota.

  3. The Workload requests three CPUs and three Pod slots.

  4. Its LocalQueue is active and points to the expected ClusterQueue.

  5. The ClusterQueue provides only two CPUs and two Pod slots.

  6. Kueue therefore keeps the Job suspended until quota becomes available or the queue configuration changes.

The important part is not that Headlamp invents a new status. It reads the same Kubernetes API objects as other clients, then places the related resources and human-readable values in one connected investigation.

If the Workload is admitted, switch views

Suppose the Workload detail page shows Admitted=True, but the application is still not running. Kueue has completed its part of the decision. Continue in Headlamp's standard Kubernetes views:

  1. Open Workloads → Jobs and inspect the Job state and events.

  2. Open Workloads → Pods and filter by the Job name or namespace.

  3. Select a Pending Pod and read its Events section.

At this stage, look for scheduler concerns such as insufficient node capacity, node affinity, taints and tolerations, topology constraints, persistent-volume binding, or image-pull failures.

This handoff is the most useful troubleshooting rule in the article:

  • Not admitted: investigate Kueue resources in Headlamp.

  • Admitted, but Pods are Pending: investigate Job, Pod, node, and scheduler information in Headlamp.

A repeatable Headlamp checklist

When the next Job does not start, use this short path:

  1. Open Kueue → Workloads and select the correct namespace.

  2. Find the Workload associated with the Job.

  3. Check the readable status, Admitted, and QuotaReserved conditions.

  4. Read the condition message and recent events.

  5. Compare PodSet resource requests with any admission assignments.

  6. Follow the LocalQueue link and confirm its status and ClusterQueue.

  7. Follow the ClusterQueue link and compare quota, reservations, usage, and policy.

  8. Inspect ResourceFlavor or Cohort details when the message points to flavor selection or shared quota.

  9. If admitted, move to Headlamp's Job and Pod views.

What Headlamp replaces during this investigation

The plugin does not replace kubectl or kueuectl for scripts, exact object output, or changes. It replaces the repetitive UI-less navigation needed during an interactive investigation.

QuestionHeadlamp pathCommand-line equivalent
Is the Workload admitted?Kueue → Workloads → status and conditionskubectl describe workload <workload-name> -n <namespace>
Which LocalQueue did it enter?Workload → linked LocalQueuekubectl get localqueue <localqueue-name> -n <namespace> -o yaml
Which ClusterQueue provides quota?LocalQueue → linked ClusterQueuekubectl get clusterqueue <clusterqueue-name> -o yaml
Which flavor or shared quota matters?ClusterQueue → ResourceFlavor or Cohortkubectl get resourceflavor <flavor-name> -o yaml or kubectl get cohort <cohort-name> -o yaml
What happened after admission?Workloads → Jobs and Pods → eventskubectl describe job <job-name> -n <namespace> or kubectl describe pod <pod-name> -n <namespace>

For scripts and GitOps workflows, the command line remains the right tool. For a person asking “why is this waiting?”, the linked Headlamp path is usually faster to read.

Optional: reproduce the pending scenario

The plugin repository includes sample Cohort, ClusterQueue, LocalQueue, ResourceFlavor, and Job manifests for local testing. Apply those examples to a cluster where Kueue is installed, then open the Kueue section in Headlamp.

To reproduce the specific quota mismatch used in this article, configure a ClusterQueue with two CPUs and two Pod slots, then submit a three-Pod Job that requests one CPU per Pod. The Job enters its LocalQueue, Kueue creates a Workload, and the Workload remains pending because the combined request is larger than the available quota.

Keeping this setup separate from the main walkthrough is intentional: once the resources exist, the investigation itself should happen in Headlamp.

Current alpha limitations

The 0.1.0-alpha release is an initial read-only view of the Kueue admission path.

  • The plugin does not add its own create, edit, pause, resume, or delete flows.

  • Workload owner references are displayed as text and are not currently navigable.

  • The UI depends on the Kubernetes permissions of the identity used by Headlamp.

  • Metrics and a visual relationship map are possible follow-up areas rather than features of this release.

These limits are useful to state clearly: the plugin's current purpose is to make admission state and resource relationships easier to inspect.

Developed during LFX Mentorship

The plugin was developed during CNCF LFX Mentorship 2026 Term 2 using TypeScript, React, Kubernetes APIs, and Headlamp's plugin system.

Utkarsh Raj implemented the plugin with guidance and review from René Dudfield (@illume) and Kevin Hannon (@kannon92). Feedback from the Headlamp and Kueue communities shaped the resource views, status presentation, navigation, documentation, tests, and first alpha release.

Try it and share feedback

Explore the plugin source and README, read the 0.1.0-alpha release documentation, or share feedback in the release issue.

If a pending Workload is still difficult to explain after following the Headlamp path, that is exactly the kind of feedback that can improve the plugin.

The Shift to cgroup v2 in Kubernetes: What You Need to Know

In Linux, cgroups (control groups) are a kernel feature used for managing system resources. Kubernetes uses cgroups to allocate resources like CPU and memory to containers, ensuring that applications run smoothly without interfering with each other. With the release of Kubernetes v1.31, support for v1 cgroup management moved into maintenance mode. Support for v2 cgroup management has been stable since Kubernetes v1.25.

Compared with cgroup v1, cgroup v2 provides a single unified hierarchy, a more consistent interface, and a stronger foundation for resource isolation and modern resource-management features.

Deprecation of cgroup v1

Kubernetes has deprecated cgroup v1. Starting with Kubernetes v1.35, failCgroupV1 defaults to true, so the kubelet does not start on a cgroup v1 node by default. Administrators can temporarily set failCgroupV1: false in the kubelet configuration file, but removal will follow the Kubernetes deprecation policy. Further removal work is tracked in KEP-5573: Remove cgroup v1 support.

If you are still on a release older than v1.35, migrate every Linux node to cgroup v2 before upgrading, or plan to set the temporary failCgroupV1: false override. If you are already on v1.35 or later, confirm that every Linux node runs cgroup v2 (or that you intentionally keep the override). Under the default configuration, a remaining cgroup v1 node fails during kubelet startup.

For kubeadm-managed clusters, Kubernetes v1.35 also makes this an earlier, stricter check. The SystemVerification preflight check, provided by k8s.io/system-validators, returns an error during kubeadm init, kubeadm join, and kubeadm upgrade when it detects cgroup v1 with kubelet v1.35 or later; with an older kubelet, the check remains a warning. See kubernetes/system-validators#1.12.1 release notes for details.

The top FAQs cover three main areas: why to migrate, the benefits and drawbacks, and key points to keep in mind when using cgroup v2.

Limitations of cgroup v1 and Improvements with cgroup v2

The Linux kernel documentation describes both interfaces:

Let's enumerate some known issues.

active_file memory is not considered available memory

The kubelet treats active_file memory as not reclaimable. For I/O-intensive workloads, a large page cache can therefore make the kubelet report memory pressure and evict Pods. This is a known kubelet issue (kubernetes/kubernetes#43916); migrating to cgroup v2 does not by itself change that calculation. The documented workaround is to set equal memory requests and limits for containers that perform intensive I/O, after measuring an appropriate value.

Memory QoS updates in Kubernetes v1.36

Memory QoS was introduced as an alpha feature in Kubernetes v1.22 and updated in v1.27. It remains alpha in v1.36, but now separates memory throttling from memory reservation and adds tiered memory protection:

Memory QoS is available only on Linux nodes that use cgroup v2. It relies on the cgroup v2 memory controller: memory.high provides throttling, while memory.min and memory.low provide hard and soft protection when tiered reservation is enabled. cgroup v1 cannot provide this protection model.

  • Enabling the MemoryQoS feature gate applies memory.high throttling to Burstable containers. The threshold is derived from the request, limit, and memoryThrottlingFactor (default 0.9).

  • memoryReservationPolicy: None is the default. It does not write memory.min or memory.low.

  • memoryReservationPolicy: TieredReservation maps Guaranteed Pod memory requests to memory.min (hard protection) and Burstable Pod requests to memory.low (soft protection). BestEffort Pods receive neither protection.

  • The kubelet exposes Alpha metrics for the total memory.min and memory.low reservations on a node.

  • Kernel 5.9 or later is recommended. On older kernels, memory.high reclaim can trigger a known livelock; from v1.36 the kubelet logs a warning when Memory QoS is enabled on an affected kernel.

For example, to opt in to tiered protection:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
featureGates:
  MemoryQoS: true
memoryReservationPolicy: TieredReservation
memoryThrottlingFactor: 0.9

The overall Kubernetes recommendation is not to enable Alpha features in production; however, if you judge that memory QoS with tiered reservations is useful for your platform in production, make sure to test the configuration and account for hard-reserved memory before you enable the TieredReservation feature gate. See the Pod QoS documentation for the current mapping of Kubernetes QoS classes to cgroup v2 controls.

Container-aware OOM handling

On cgroup v2 nodes, the kubelet defaults singleProcessOOMKill to false. It therefore sets memory.oom.group for each container cgroup so that an OOM event kills all processes in that container together, rather than leaving a partially functioning multi-process container. Set singleProcessOOMKill: true only if you need the cgroup v1-compatible behavior where the kernel may kill one process at a time. See the KubeletConfiguration reference for this setting.

This behavior is scoped to a container cgroup, not the whole Pod. Also, cgroup.kill is a separate administrative interface: writing 1 to it sends SIGKILL to every process in that cgroup and its descendants; it does not configure OOM behavior. The cgroup v2 memory controller additionally provides memory.events counters that monitoring systems and userspace OOM managers can observe.

Rootless support

In cgroup v1, delegating controllers to less privileged containers may be dangerous.

Unlike cgroup v1, cgroup v2 officially supports delegation. Most implementations of rootless containers rely on systemd for delegating v2 controllers to non-root users.

This delegation mechanism is separate from Kubernetes Pod user namespaces, which map container users to unprivileged host users. Pod user namespace support graduated to stable in Kubernetes v1.36; check its filesystem, kernel, CRI runtime, and OCI runtime prerequisites before enabling it.

What else?

  1. eBPF stories:
    • In cgroup v1, device access controls are exposed through interface files.
    • The cgroup v2 device controller has no interface files, and is implemented on top of cgroup BPF.
    • Cilium attaches BPF cgroup programs for socket-based load balancing. Its default cgroup root is /run/cilium/cgroupv2.
  2. Pressure Stall Information (PSI) reports CPU, memory, and I/O contention at node, Pod, and container level. On supported clusters, the kubelet exposes PSI by default (KubeletPSI is stable and locked on). PSI requires cgroup v2, Linux 4.20 or later, CONFIG_PSI=y, and a kernel not booted with psi=0. The kubelet surfaces the data through the Summary API and /metrics/cadvisor.
  3. When migrating, update software that reads the cgroup filesystem directly. The migration guidance recommends cAdvisor v0.43.0 or later and lists compatible Java, Node.js, and automaxprocs versions.

CPU weight conversion in newer OCI runtimes

cgroup v1 uses cpu.shares, whereas cgroup v2 uses cpu.weight. Newer OCI runtimes use an improved non-linear conversion that preserves the default priority and gives small CPU requests more usable granularity. The change is implemented in the OCI runtime rather than Kubernetes: it is available in crun v1.23 and runc v1.3.2. After upgrading a runtime, monitoring or policy tools that predict exact cpu.weight values may need updates. Read New Conversion from cgroup v1 CPU Shares to v2 CPU Weight for the formula, examples, and compatibility considerations.

In-place resource updates

In-place Pod vertical scaling graduated to stable in Kubernetes v1.35. Kubernetes v1.36 then [enabled] (/blog/2026/04/30/kubernetes-v1-36-inplace-pod-level-resources-beta/) in-place vertical scaling for Pod-level resources by default, as a Beta feature. The kubelet coordinates changes between the Pod-level and container cgroups so that increases create headroom before container limits grow, while decreases constrain containers before shrinking the Pod-level boundary. Accurate aggregate enforcement for this v1.36 feature requires cgroup v2.

Adopting cgroup version 2

Requirements

Here's what you need to use cgroup v2 with Kubernetes. First up, you need to be using a version of Kubernetes with support for v2 cgroup management; that's been stable since Kubernetes v1.25 and all supported Kubernetes releases include this support.

  • You need at least one Linux node; cgroup is a Linux-only concept
  • Your OS install must run with cgroup v2 enabled
  • The kernel version must be 5.8 or later (5.9 or later is recommended when using memory QoS)
  • The container runtime must support cgroup v2. For example:
    • containerd v1.4 or later supports cgroup v2; use containerd v2.0 or later for automatic cgroup-driver discovery
    • CRI-O v1.20 or later
  • The kubelet and the container runtime must both be configured to use the correct cgroup driver. See Configure the kubelet's cgroup driver to match the container runtime cgroup driver.

For now, you can opt back in to use cgroup v1; the Kubernetes project recommends using cgroup v2, but in Kubernetes 1.36 (the current release) the cgroup v1 option remains supported as a fallback. That fallback is scheduled for removal in Kubernetes v1.38. If you are running an older cluster, plan to migrate; if you are setting up a new cluster with Linux nodes, you should prefer cgroup v2. In either case, review both the Kubernetes runtime documentation and (if relevant) the containerd compatibility matrix.

kernel updates around cgroup v2

When Kubernetes was first announced, in 2014, only v1 cgroup existed. Version 2 cgroup management first appeared in Linux kernel 4.5, released in 2016.

  • In Linux 4.5, the cgroup v2 io, memory, and pids controllers were supported.
  • Linux 4.15 added support for the cgroup v2 cpu controller.
  • Pressure Stall Information (PSI) support began with Linux 4.20.
  • The Kubernetes project does not recommend using cgroup v2 with a Linux kernel older than 5.2 due to lack of cgroup-level task freezer support.
  • Kubernetes documents 5.8 as the minimum kernel version for cgroup v2; the root cgroup's system-level cpu.stat file was added in Linux 5.8.
  • The memory.high livelock fix used by Memory QoS is present in Linux 5.9 and later.
  • memory.peak was added in Linux 5.19.

cgroup driver configuration

Configure the kubelet's cgroup driver to match the container runtime cgroup driver.

If you use kubeadm to manage your cluster, Kubernetes recommends that you use the systemd cgroup driver, because kubeadm manages the kubelet as a systemd service. For other management tooling, check the documentation for the tool you're using to manage your cluster.

If you can pick either option, I recommend using the systemd driver.

Whatever tooling you've chosen, the kubelet automatically tries to detect the runtime's recommended cgroup driver. This automatic detection relies on using a runtime that implements the RuntimeConfig CRI RPC (for example: containerd v2.0+ or CRI-O v1.28+).

If you're using a container runtime that supports cgroup v2 but doesn't support automatic cgroup driver detection, you can manually configure an override by editing the kubelet configuration file. For example:

apiVersion: kubelet.config.k8s.io/v1beta1
kind: KubeletConfiguration
cgroupDriver: systemd

Automatic discovery of the runtime's cgroup driver through the CRI, tracked by KEP-4033, graduated to stable in Kubernetes v1.34. It requires a runtime that implements the RuntimeConfig CRI RPC (containerd v2.0+ or CRI-O v1.28+). When available, the kubelet uses the value reported by the runtime instead of its configured cgroupDriver value.

Tools and commands for troubleshooting

Tools and commands that you should know about cgroups:

  • stat -fc %T /sys/fs/cgroup/: Check whether cgroup v2 is enabled; it returns cgroup2fs.
  • systemctl list-units 'kube*' --type=slice or --type=scope: List Kubernetes-related units that systemd currently has in memory.
  • bpftool cgroup list /sys/fs/cgroup/*: List all programs attached to the cgroup CGROUP.
  • systemd-cgls /sys/fs/cgroup/*: Recursively show control group contents.
  • systemd-cgtop: Show top control groups by their resource usage.
  • tree -L 2 -d /sys/fs/cgroup/kubepods.slice: Show Pods' related cgroups directories.

How to check if a Pod CPU or memory limit is successfully applied to the cgroup file?

Work from the API object down to the node. Identify the node, then compare desired resources in the spec with the enacted values in status (in-place resize can leave those out of sync):

kubectl get pod <pod-name> -n <namespace> \
  -o jsonpath='{.spec.nodeName}{"\n"}'

kubectl get pod <pod-name> -n <namespace> \
  -o jsonpath='{range .spec.containers[*]}{.name}{" spec\t"}{.resources}{"\n"}{end}'

kubectl get pod <pod-name> -n <namespace> \
  -o jsonpath='{range .status.containerStatuses[*]}{.name}{" status\t"}{.resources}{"\n"}{end}'

On that node, as root, with the systemd cgroup driver and cgroup v2 (these snippets also need jq). Pin the container with kubelet labels; a container name alone is not unique on a node. Use crictl ps -a if the container is not running:

CONTAINER_ID=$(crictl ps \
  --label io.kubernetes.pod.namespace=<namespace> \
  --label io.kubernetes.pod.name=<pod-name> \
  --name <container-name> -q | head -n1)

# OCI view (not the CRI protobuf field names)
crictl inspect "$CONTAINER_ID" | jq '.info.runtimeSpec.linux.resources'

# Kernel cgroup path for this container
PID=$(crictl inspect "$CONTAINER_ID" | jq -r '.info.pid')
CGROUP="/sys/fs/cgroup$(awk -F: '$1=="0"{print $3}' /proc/$PID/cgroup)"

cat "$CGROUP/cpu.weight"   # request.cpu → shares → weight; not a limit
cat "$CGROUP/cpu.max"      # limit.cpu as "quota period"; unlimited is "max <period>"
cat "$CGROUP/memory.max"   # limit.memory; unlimited is "max"

# Pod-level cgroup (parent slice), used by Pod-level resources
cat "$(dirname "$CGROUP")/cpu.max" "$(dirname "$CGROUP")/memory.max"

# Present when MemoryQoS is enabled
cat "$CGROUP/memory.high" "$CGROUP/memory.min" "$CGROUP/memory.low"

SCOPE=$(basename "$CGROUP")
systemctl show "$SCOPE" \
  -p CPUWeight -p CPUQuotaPerSecUSec -p CPUQuotaPeriodUSec -p MemoryMax

Do not treat info.runtimeSpec.linux.cgroupsPath as a filesystem path when the runtime uses the systemd driver; that value is a systemd unit path (slice:runtime:id), not a directory under /sys/fs/cgroup.

Expected mapping at each layer:

  • Kubernetes: spec.containers[*].resources (desired) and status.containerStatuses[*].resources (enacted after in-place resize). Pod-level resources are on spec.resources and the parent pod slice.
  • CRI (kubelet → runtime; these protobuf names stay cgroup v1-style even on v2): cpu_shares, cpu_quota, cpu_period, memory_limit_in_bytes, plus unified for Memory QoS. crictl inspect does not show these names; it shows the OCI spec the runtime derived from them.
  • OCI spec: linux.resources.cpu.{shares,quota,period}, linux.resources.memory.limit, and linux.resources.unified. The runtime converts shares, quota, and period into cgroup v2 files. After upgrading crun or runc, the shares-to-weight formula may change; see CPU weight conversion.
  • systemd scope: CPUWeight, CPUQuotaPerSecUSec, CPUQuotaPeriodUSec, MemoryMax. Unlimited or unset values often appear as [not set] or infinity.
  • cgroupfs: cpu.weight (from CPU request; an unset request still yields the default of 2 shares), cpu.max, and memory.max on the container scope. Memory QoS adds memory.high, memory.min, and memory.low.

Further reading

Scaling Kubernetes Workloads with Node Swap

Memory is often the first hard limit a Kubernetes cluster hits. Nodes run out of RAM long before they run out of CPU, and the new wave of agentic AI workloads makes this worse. These workloads demand large memory footprints to start up and run untrusted code, then sit idle waiting for the next prompt. That idle but resident memory is expensive, and it caps how many pods a node can hold. This is where swap helps. Kubernetes support for running nodes with swap enabled reached General Availability in v1.34, and by backing that swap with fast NVMe solid state drives (SSDs), a node can page out dormant memory and pack in far more pods. This post explains how we benchmarked that approach across three workloads, including CI/CD kernel builds, sandboxed headless browsers, and isolated Python runtimes; we found density gains of up to 3×, often with little or no latency cost.

The node density problem

The Kubernetes ecosystem has reached a fundamental physical resource constraint: the strict limits of hardware memory versus the growing demand for dynamic, bursty workloads in the new agentic era.

Historically, administrators provisioning memory-intensive workloads encountered a persistent dilemma: set memory limits too high and you waste expensive infrastructure on idle RAM; set them too low and you risk Out-Of-Memory (OOM) kills.

This conflict is amplified when deploying autonomous AI agents using secure execution environments like the agent-sandbox framework. These agentic pods require large memory footprints to initialize and execute untrusted code. However, after their burst of activity, they typically enter long-tail idle phases waiting for user prompts. Keeping this idle state in physical RAM caps cluster density and makes AI infrastructure expensive to run.

The solution: Kubernetes node swap

With the introduction of Kubernetes' support for running nodes with swap enabled (which reached General Availability in v1.34), this paradigm shifts. By enabling the Linux kernel to page out anonymous memory to disk, node swap acts as a shock absorber during traffic spikes or periods of heavy memory oversubscription.

Historically, swap was discouraged in Kubernetes for two reasons. The first was memory accounting. Under cgroup v1, the controls treated memory and swap as a single combined limit rather than letting operators set an independent limit for disk swap. Without independent tracking, a process could page large amounts of anonymous memory out to disk, which made a container's real memory usage unpredictable and hard to isolate. Kubernetes' swap support resolves this by relying on cgroup v2, whose separate swap accounting tracks disk swap on its own. The second reason was the latency penalty of paging to slow spinning disks, which fast NVMe Local SSDs largely eliminate. Together, these make it practical to increase pod density and buffer against memory spikes without sacrificing cluster stability.

The benchmark data

To quantify the performance boundaries and cost-saving potential of Local SSD-backed node swap, this analysis covers three distinct workload categories: a Traditional Build Workload for CI/CD pipelines, High-Density Browser Sandboxes, and Isolated Python Sandboxes.

Workload ProfileBaseline Capacity (No Swap)Local SSD Swap CapacityDensity Improvement
Linux CI/CD Kernel Build600 MB RAM Limit300 MB RAM Limit-50% RAM Footprint
Headless Chrome (Kata)40 Concurrent Pods50 Concurrent Pods+25% Pod Density
Headless Chrome (gVisor)80 Concurrent Pods160 Concurrent Pods+100% Pod Density
Python Sandbox (gVisor)80 Concurrent Pods240 Concurrent Pods+200% Pod Density

1. Traditional workload: Linux kernel build

Before exploring specialized agentic architectures, swap was validated against classic batch workloads by running a complete Linux 6.1.1 kernel build. The kernel compilation process leverages concurrent worker threads, balloons in memory to hold compiled object files, and requires a large memory spike during the brief linking phase.

This workload mirrors the memory behavior of enterprise CI/CD pipelines. Because earlier compiled objects sit inactive in memory while the pipeline progresses, CI/CD jobs frequently hoard unused physical RAM, which makes them well suited to node swap compression.

On a baseline node without swap, the minimum memory limit to prevent an OOM crash during compilation was 600 MB. Routing swap to a Local SSD cut the container memory limit by 50% to 300 MB without incurring any execution slowdown (in fact, it ran cleanly in 374s vs the baseline 433s). However, as an explicit tradeoff, compressing the limit further to 200 MB forced the active working set into swap, causing long I/O wait times and increasing execution time by over 40%. This reinforces that swap serves as an insurance policy for burst memory, not a replacement for active RAM.

2. High-density agent workloads: headless browser runtimes

AI agent workloads frequently require manipulating headless browsers via Chromium. However, trusting external code execution often requires stricter security isolation than standard Linux namespaces. This benchmark cross-evaluated several container runtime environments. The raw logs and testing methodologies for the default runtime are available in the Agent Sandbox GKE Swap directory.

  • Unsandboxed baseline limits (runc): To test the limits of the environment without the overhead of security runtimes, plain runc containers were swept on a c4-standard-32 node (32 vCPU, 120 GB RAM). Without swap, the node exhausted physical memory and failed past 512 pods. Enabling Local SSD swap allowed the node to support 768 concurrent pods.
  • Advanced security runtimes (for example: gVisor, Kata Containers): Enabling strict security sandboxing increases memory overhead and normally reduces pod density. However, memory swap naturally absorbs this overhead penalty. Without swap, a gVisor environment hit a hard limit at 80 pods. Local SSD swap doubled that capacity, which allowed 160 concurrent gVisor pods on a single node. Similarly, Kata Containers microVMs exhausted physical RAM at 40 concurrent pods without swap, but using GCP Local SSD swap expanded this to 50 stable Kata microVMs before CPU saturation.

At these maximum densities, the per-pod latency increase is driven mainly by pods competing for CPU, not by swap I/O. An operator tuning for a specific latency target would run at a lower density than the peak numbers here and see a proportionally smaller latency cost.

For a comprehensive architectural breakdown and density metrics for gVisor and Kata, see the Agent Sandbox GKE Swap Runtimes directory.

3. Beyond browsers: sandboxed Python runtimes

The advantages of node swap also extend to untrusted, isolated code-execution environments. This sweep deployed simultaneous Python sandbox sessions analyzing 5 million rows of data from the MovieLens 20M dataset, requiring a ≃375 MiB resident memory footprint per execution. The in-depth scaling results and deployment code for this sweep can be reviewed in the Agent Sandbox GKE Swap Python Density directory.

Without swap, heavy concurrent bursts exhausted physical memory, causing the node to hit a hard RAM limit and fail at 80 concurrent sessions. Enabling Local SSD swap offloaded dormant anonymous memory, freeing up physical RAM and preserving the node's page cache. This allowed the node to scale to 240 concurrently isolated Python sandboxes—a 3× density improvement. As with the browser workloads, the latency rise at peak density comes mainly from the sandboxes competing for CPU rather than from swap itself.

Density Benchmarks with and without Node Swap

How to use it

If you manage Kubernetes infrastructure for developer environments, browser testing farms, JVM applications, or AI execution runtimes, leveraging Local SSD swap can multiply your density efficiency.

In Kubernetes v1.34+, node swap is Generally Available. You enable it via the kubelet configuration:

kind: KubeletConfiguration
apiVersion: kubelet.config.k8s.io/v1beta1
failSwapOn: false
memorySwap:
  swapBehavior: LimitedSwap

Pairing this upstream configuration with your cloud provider's high-speed local disk gives you dynamic memory balancing. For example, this is natively supported on Google Kubernetes Engine via Node Memory Swap configured on Local SSD profiles.

To get the benefits, configure your workloads with Burstable QoS: set your container's memory limits higher than its requests. The node automatically rations fast swap space based on idle application memory usage while keeping active processes responsive.

Conclusion

As the Kubernetes ecosystem transitions into the agentic era, administrators face a growing conflict between finite hardware memory limits and the bursty behavior of AI workloads. Frameworks like Agent Sandbox provide the security isolation required for running untrusted agents, but that isolation traditionally demands large amounts of idle memory overhead.

By configuring the kubelet with LimitedSwap and routing it to Local SSDs, you can mitigate this conflict. Fast swap offloads the dormant states of idle agents, allowing you to increase pod density and node utilization on the same infrastructure without compromising security boundaries.

Spotlight on SIG Apps

As Kubernetes adoption has grown, the conversation has shifted beyond running containers to managing increasingly complex application lifecycles. Modern platforms support stateless web services, stateful databases, batch processing, AI workloads, and platform services. At the same time, they must remain reliable during upgrades, scaling events, and infrastructure failures.

Every Kubernetes user relies on SIG Apps, whether they realize it or not. Deployments, StatefulSets, DaemonSets, Jobs, and CronJobs form the foundation of how applications are deployed, updated, scaled, and operated across the Kubernetes ecosystem.

SIG Apps is focused on improving workload resilience, refining application lifecycle management, and addressing the operational challenges that emerge when applications encounter node failures, rollout disruptions, and increasingly complex infrastructure environments.

In this spotlight, we sit down with two of the three SIG Apps chairs Janet Kuo and Maciej Szulik to discuss the evolution of Kubernetes workload management, the challenges of balancing application reliability with operational simplicity, and the future of application lifecycle management within one of Kubernetes’ most influential Special Interest Groups.

Introducing SIG Apps

Natalie Fisher: Can you introduce yourself, your role, and how you got involved in SIG Apps?

Janet Kuo: I'm a Senior Staff Software Engineer at Google and have been a Kubernetes maintainer since 2015, joining the community just as we were racing toward the 1.0 launch. In those early days, my focus was on building the core Workloads API, specifically developing controllers like Deployment, ReplicaSet, StatefulSet, and DaemonSet, defining their rollout behaviors, and bringing them from initial designs to GA. That hands-on work was my entry point into SIG Apps.

Since then, I've stayed deeply involved in both the technical and community sides of Kubernetes. I have led SIG Apps as Co-Chair and Tech Lead since 2019. Currently, in addition to maintaining the workloads API, I am driving new subprojects like the Agent Sandbox to ensure Kubernetes is ready for next-generation agentic and AI workloads.

Maciej Szulik: I started contributing to Kubernetes all the way back in 2014. Since then, I've worked across various areas of the project: controllers, kubectl, and apimachinery, which eventually led me to become one of the Chairs and Tech Leads for SIG Apps. My current focus is reliability of the workload controllers under the SIG Apps umbrella and stability and ease of use of kubectl as part of my SIG CLI Tech Lead role. I also care about overall community health and growth as part of my Steering Committee role. Outside of Kubernetes, I work as a Staff Platform Engineer at Defense Unicorns, where I'm helping make Kubernetes more airgap-native with a project called zarf.

The problem and the solution

SIG Apps is responsible for the core workload APIs that power how applications run on Kubernetes. From Deployments and StatefulSets to Jobs and CronJobs, these controllers determine how workloads are created, updated, scaled, and recovered when things go wrong.

As Kubernetes expands to support increasingly diverse workloads – including AI, batch processing, and large-scale distributed applications – SIG Apps continues to evolve these APIs while balancing reliability, backward compatibility, and operational simplicity.

NF: For readers who may not be familiar, what is SIG Apps, and what role does it play within the broader Kubernetes ecosystem?

MS: SIG Apps is the Kubernetes Special Interest Group responsible for the workloads APIs. CronJob and Job help running batch workloads, whereas DaemonSet, Deployment, ReplicaSet, and StatefulSet serve the majority of other applications. More broadly, SIG Apps owns the layer most developers actually touch day-to-day: the controllers that turn a workload specification into running, self-healing pods. It's the group deciding how Deployments roll out, how Jobs retry, how DaemonSets place a pod per node.

JK: Adding to what Maciej described, as the industry shifts, we are seeing a massive demand to run complex, non-traditional workloads like distributed AI training, batch computing, and dynamic agent environments. Our role is expanding: we aren't just maintaining the classic workloads API, but we are actively evolving it and establishing new patterns (like the Agent Sandbox) to make sure Kubernetes remains the best platform for the next generation of workloads, such as AI.

NF: Looking at the workload APIs owned by SIG Apps (Deployments, StatefulSets, DaemonSets, Jobs, and CronJobs), which areas are receiving the most attention from maintainers and contributors?

MS: After a long stretch focused on making batch workloads run smoothly on Kubernetes, we’ve shifted attention to make sure serving workloads (DaemonSets, StatefulSets, etc) aren’t left behind. This means performance and high-scale improvements to rollout and scaling behavior, plus working through our backlog of user-reported issues, prioritizing the ones with the strongest support from the user base.

Current focus areas

As Kubernetes workloads grow in scale and complexity, the challenges facing workload controllers evolve as well. We asked the SIG Apps chairs where contributors are focusing their efforts today and which resilience problems they believe are the highest priorities.

NF: From your perspective, what are the most important workload resilience problems SIG Apps is trying to solve today?

MS: Node lifecycle challenges have come up repeatedly across SIG Apps, SIG Node, and SIG Autoscaling discussions. DaemonSets and Jobs are just where the pain is most visible, since they're the workloads most directly bound to node state. Rather than solve it piecemeal within one SIG, we've settled on spinning up a dedicated Node Lifecycle Working Group to focus on this properly and hopefully land long-term solutions instead of one-off patches.

JK: From an AI perspective, resilience is critical. When you are running a massive distributed LLM training job that spans hundreds of GPUs, a single node failure can halt the entire pipeline. Similarly, if a DaemonSet that runs your logging or GPU monitoring agent gets stuck on a bad node, it impacts the entire cluster's health.

In addition to the work in the Node Lifecycle WG to handle infrastructure-level degradation, SIG Apps is addressing this at the orchestration layer through subprojects like JobSet (for distributed training) and LeaderWorkerSet (LWS) (for sharded LLM inference). These APIs introduce patterns like "all-or-nothing" failure handling, where a single pod or job failure triggers a coordinated group-level restart to resume from the last clean checkpoint, rather than letting stuck workloads hang in an inconsistent state.

Real-world impact

The work happening within SIG Apps extends far beyond controller implementations and API design. We wanted to understand what these improvements mean in practice for platform teams operating Kubernetes clusters in production.

NF: For platform teams operating Kubernetes in production, what practical improvements would they notice if the node lifecycle and workload resilience work currently under discussion is successfully delivered?

MS: I’m mostly looking from the sidelines, the folks actually in the Node Lifecycle Working Group would give you a sharper answer. But from where I sit, I’m hoping their work translates into fewer 3am pages that turn out to be “a DaemonSet rollout got stuck because node X was flaky, and someone had to manually cordon/delete/restart to unstick it.”

JK: +1 to what Maciej said, and beyond reducing manual intervention, platform teams will also see much better resource predictability and cost efficiency. For example, in AI workloads where GPU idle time is extremely expensive, having Kubernetes automatically detect a degraded node and reschedule the training coordinator or agent before the job crashes means less wasted compute and more stable job execution.

Challenges and trade-offs

Evolving APIs that millions of workloads rely on requires careful engineering and even more careful decision-making. We asked the SIG Apps chairs about the technical and operational trade-offs they weigh when introducing changes to Kubernetes’ core workload controllers.

NF: What are some of the hardest technical or operational trade-offs SIG Apps encounters when evolving core workload controllers?

MS: Honestly, a few tensions keep coming up: how aggressively a controller should give up on stuck pods, and what signals it actually needs to make that call correctly. At the same time, we always have to think about backward compatibility. Deployment, DaemonSet, and Job behavior has been depended on for a decade [by Kubernetes users, tooling, automation, and higher-level controllers], so even a change that’s clearly “more correct” can break automation people built around the old behavior without meaning to.

JK: One of our hardest trade-offs is resisting the urge to make "elegant" design changes that break backward compatibility. Instead, we have to design opt-in features that let users adopt new behaviors without forcing them on legacy workloads. When we need to support completely new paradigms, we prefer introducing them as CRDs first rather than bloating the core APIs, like we are doing with Agent Sandbox, JobSet, and LWS.

Looking ahead

While much of SIG Apps’ work focuses on maintaining the stability of existing workload APIs, the group is also shaping the future of Kubernetes through new enhancements and proposals. We concluded by asking about one proposal that recently returned to active development and what it represents for the future of workload management.

NF: The SIG recently discussed reviving KEP-4443 with a target release of Kubernetes 1.38. What opportunities or challenges does this proposal aim to address, and why is now the right time to revisit it?

KEP-4443 addresses a small but real gap in the Job API: a PodFailurePolicy can be configured to add a condition reason to the JobFailed condition, but different pod failure policy rules targeting different container exit codes all produce that same generic reason. The proposal is simple: an optional Name field on each PodFailurePolicyRule, which gets appended to the JobFailed condition reason, so higher-level tools like JobSet can finally react differently depending on which rule triggered the failure.

As for timing, the answer is as simple as it always is in open source: we lost the original contributor who was driving this. Now we’ve got someone new interested in picking it up, that’s why we’re targeting the next release.

Getting Involved

NF: For someone interested in contributing to SIG Apps, where would you recommend they start, especially if they are not yet a Kubernetes maintainer?

MS: The best place to start is the #sig-apps slack channel and our regular SIG Apps meetings. We’ve all started there, and if it feels intimidating, or nobody replies right away, that’s completely normal. Everyone's busy. It's not personal.

JK: In addition to what Maciej answered, I'd suggest looking at our newer subprojects and initiatives. Contributing to stable APIs like Deployment or StatefulSet can be daunting because the barrier for making changes is very high due to backward compatibility, and there is much less low-hanging fruit.

If you are new to the community, projects like the Agent Sandbox are fantastic entry points. They are actively evolving, have a friendly group of maintainers, and offer plenty of greenfield development opportunities where you can make a significant impact quickly.

Summary

SIG Apps has shaped how Kubernetes applications are deployed and operated since the project’s earliest days. While users often interact with Deployments, StatefulSets, Jobs, and DaemonSets without thinking about the controllers behind them, the work within SIG Apps continues to shape the reliability and scalability of workloads across the Kubernetes ecosystem.

From improving workload resilience and node lifecycle behavior to enabling new patterns for AI and distributed computing, the SIG is evolving Kubernetes while remaining committed to one of the project’s core principles: preserving the stability and backward compatibility that users depend on. Whether you’re interested in core workload APIs, emerging projects like Agent Sandbox, or helping improve the operational experience of Kubernetes users everywhere, SIG Apps offers many opportunities to get involved.

Kubernetes v1.37: Tracking When a PersistentVolumeClaim Was Last Used (Beta)

Kubernetes v1.37 promotes the PersistentVolumeClaimUnusedSinceTime feature gate to Beta (enabled by default). With this feature, the PersistentVolumeClaim (PVC) protection controller adds an Unused condition to each PVC, telling you whether any running pod currently references it — no custom tooling or cross-referencing required.

For the API definition of PVC conditions, see the PersistentVolumeClaim API reference. Read on to learn how the Unused condition works and how to use it.

Why track PVC usage?

In large-scale Kubernetes clusters, it is common for users to create PVCs and then delete the associated pods without cleaning up the storage, because Kubernetes does not automatically delete PVCs when their pods are removed (to protect against accidental data loss). Over time, these orphaned PVCs may