The unofficial SRE track for KubeCon NA '26

Some of the most interesting sessions at this year’s KubeCon explore the gap between a successful operation and a reliable system. It’s a good reason to look beyond talks with “SRE” in the title.

KubeCon + CloudNativeCon North America returns to Salt Lake City November 9–12, 2026. Continuing Rootly’s unofficial SRE track tradition, we’ve picked seven sessions about finding failures earlier, understanding them faster, and keeping a way back when automation goes wrong.

When Infrastructure Breaks in Production: How EarnIn Brought Testing to the Full CNCF Stack

A backup job that completes without actually running is a particularly unpleasant kind of green checkmark.

Priya Namasivayam (EarnIn) will share how the team brought infrastructure testing into its Flux GitOps pipeline after encountering failures including broken certificate chains and misleadingly successful backup jobs. Their platform runs more than 15 CNCF add-ons, so testing each component in isolation only gets them so far.

The session covers an in-cluster framework built with Testkube, tests modeled as Kubernetes resources, and quality gates triggered by GitOps changes. This is the pick for anyone who has a thorough application test suite and a much more hopeful approach to infrastructure upgrades.

When: Tuesday, November 10 | 12:20–12:50 pm MST
Where: Salt Palace | Level 1 | Grand Ballroom GI

View Priya’s session

The SRE Is Not Dead: What AI Taught Us about Missing Context in Kubernetes Incident Response

Spotting a CPU spike is useful. Knowing what deployed twenty minutes before it is often more useful.

Muhammed Yunas Chirayath and Ajmal Ali kannu will describe an AI-assisted incident response effort on a financial services platform where missing context sent engineers toward the wrong signals. The problem included deployment history, service ownership, and degrading dependencies—information a telemetry anomaly alone couldn’t supply.

Their response combines OpenTelemetry, Argo CD, Prometheus, Grafana, and Backstage into a workflow that connects those pieces. The reason to attend is the account of what was missing and how they rebuilt around it. If you’re evaluating incident automation, bring a question: which context does your system reliably have when the page arrives?

When: Tuesday, November 10 | 3:40–4:10 pm MST
Where: Salt Palace | Level 1 | Grand Ballroom ACE

View the incident response session

When the Clock Lies: How an 8-Minute Time Skew Locked Kubernetes cluster for a day

A pod is stuck in CreateContainerError for more than nineteen hours. The containers on the node are running. So why does kubelet keep trying to recreate them?

Akhil Mohan will trace a debugging journey from a hypervisor reboot to a clock skew during bootstrap. Service account tokens were being issued in the future, container name reservations were accumulating, and kubelet was caught in a loop with containerd.

The appeal here is the investigation: which assumptions looked reasonable, what evidence overturned them, and how the team followed the failure across the boundary between Kubernetes and its infrastructure. If you’ve ever stared at two components giving contradictory accounts of the same workload, this is your session.

When: Tuesday, November 10 | 4:40–5:10 pm MST
Where: Hyatt Regency | Level 4 | Regency Ballroom CD

View Akhil’s debugging session

GitOps Meltdown: How a Commit Triggered a Cascading Delete of 39 Clusters in Production

What happens when the automation responsible for restoring your desired state deletes itself?

Deepak Goel (Nutanix) will walk through an incident on a management platform responsible for 39 workload clusters. A missing Kustomization triggered cascading pruning, including deletion of the Flux controller itself. The abstract describes orphaned resources, stuck finalizers, and lost Git repository storage.

The session follows the failure chain through inventory tracking, storage ownership, reclaim policies, and deletion ordering, then covers rebuilding the GitOps control plane and resuming reconciliation safely. It’s a compelling pick because it asks the recovery question all the way down: what do you need to restore the system that normally restores everything else?

When: Tuesday, November 10 | 5:30–6:00 pm MST
Where: Salt Palace | Level 2 | 255 BC

View Deepak’s session

Anatomy of an Outage: When Bursty ML Workloads Break the Kubernetes Scheduler

You don’t need to operate a drug-discovery platform to care about a workload burst taking down the control plane.

Cellan Hall and Jenny Vega (Isomorphic Labs), with Karol Szuster (Google), will unpack an outage involving tens of thousands of concurrent ML workloads per cluster. Mixing bursty inference with distributed training led to webhook timeouts and cascading control plane failure.

Their recovery work includes separating queues, tuning Kueue’s preemption fairness, and limiting scheduling signals to protect the API server. This is an AI infrastructure session with a very familiar SRE problem underneath: how to stop a surge in demand from overwhelming the machinery that allocates capacity.

When: Wednesday, November 11 | 11:00–11:30 am MST
Where: Salt Palace | Level 1 | Grand Ballroom H

View the scheduler outage session

The Most Boring Talk at KubeCon – And the One That Will Save Your Job

When did you last test restoring your developer portal’s database? What happens to your delivery pipeline when its credentials expire?

Lev Andelman, Regina Voloshin, Cortney Nickerson, and Julie Gunderson are running a hands-on workshop about keeping an internal developer platform alive. The abstract promises a failure that cascades through version control, the portal, and the GitOps pipeline, followed by recovering together.

Set aside the full eighty minutes for this one. After several talks about other teams’ incidents, this is a chance to practice recovery yourself—and come back with better questions about the dependencies behind your own golden paths.

When: Thursday, November 12 | 11:00 am–12:20 pm MST
Where: Hyatt Regency | Level 4 | Regency Ballroom A

View the recovery workshop

Fencing Tokens and the Three-Party Contract: What Most Distributed Locks Get Wrong

A process acquires a lock, pauses, then wakes up and carries on writing. Meanwhile, someone else has become the owner. What stops the old process from corrupting shared state?

Benjamin Wang will explore that failure through lessons from building coordination primitives on etcd. The session explains fencing tokens and the cooperation required between the lock service, the client, and the resource receiving writes, with live demonstrations over HTTP and gRPC.

This is the deeper systems pick for the week. If your recovery plan includes electing a new leader, it’s worth asking how the rest of the system rejects work from the old one.

When: Thursday, November 12 | 2:40–3:10 pm MST
Where: Hyatt Regency | Level 4 | Regency Ballroom CD

View Benjamin’s session

You can attend all seven without a session overlap, although Tuesday is the busiest day. Leave time for the conversations between talks, too.

Join our KubeCon happy hours

Keep the conversations going after the sessions. We’re co-hosting two happy hours in Salt Lake City:

• Tuesday, November 10: KubeCon NA 2026: Afterparty Social at Gracie’s, with Tailscale, Mezmo, Plakar, and ClickHouse.
• Wednesday, November 11: KubeCon NA 2026: After-Hours Social at Poplar Street Pub, with Twingate, Zesty, and Chainguard.

RSVP through the event links above. We’d love to see you there!