KubeCon 2024: The Boring Stuff Won, As It Should
Two KubeCons this year, Paris in March and Salt Lake City in November. The headline is that Kubernetes finished growing up.
I was at both KubeCons this year. Paris in March. Salt Lake City in November. The headline takeaway is unromantic and important: Kubernetes finished growing up.
The ecosystem has stopped chasing novelty. The big talks are about operations, cost, and integration with the rest of the business. The hallway conversations are about hiring, migrations from older platforms, and platform engineering. The vendor floor is consolidating. This is what a mature technology looks like.
What dominated Paris
Three threads ran through KubeCon EU:
Platform engineering as a discipline. Backstage, Crossplane, Port, Humanitec, all framed as ways to give application developers a paved path. The sub-text was "Kubernetes is too complex for product teams to use directly, here is the abstraction layer". I agree with the diagnosis. The market has not picked a winner yet.
OpenTelemetry as the default observability story. Five years in, OTel has become the spec everyone agrees on. Vendors that fought it lost. Vendors that embraced it are now selling backends, not collection. The collection layer has commoditised, which is the right outcome.
WASM as the next runtime, maybe. The third year in a row this was the "next year" story. I am still not betting on it for production workloads, but the tooling is real and the use cases are getting clearer. Edge compute, plugin sandboxing, polyglot extensions in databases. Not a Kubernetes replacement.
What dominated Salt Lake City
Different feel. Smaller crowd than Chicago 2023. More serious. Fewer "we're hiring" t-shirts and more enterprise architects with actual problems.
AI workloads on Kubernetes. The big shift. GPU scheduling, multi-tenant GPU sharing, model serving frameworks (KServe, vLLM operators), inference autoscaling. A year ago this was a side track. In Salt Lake it was the main story. The cluster admins who spent a decade arguing about pod density are now arguing about MIG slicing on H100s.
Cost and FinOps. Three years ago every talk was "scale to a million pods". Now the talks are "we ran a million pods and the bill was insane, here is how we cut it 60%". OpenCost graduated. Karpenter is everywhere. Spot/preemptible adoption is finally mainstream.
Security pipelines. Sigstore, SLSA, in-toto, policy-as-code with OPA and Kyverno. The supply chain conversation is no longer aspirational. Real teams shipping signed artifacts, attesting build provenance, enforcing policy at admission. The Solarwinds-shaped scar tissue is finally producing tooling.
What was conspicuously absent
Service mesh hype. Istio is fine, Linkerd is fine, the religious war is over and most teams have decided whether they need a mesh or not. The talks were operational, not evangelical.
The vendor pattern
The ecosystem is consolidating. Two patterns:
- Big clouds absorbing capabilities. EKS Auto Mode, GKE Autopilot, AKS-but-managed-better. The "do I need to manage my own control plane" question has a clear answer for most teams, and it is "no".
- Mid-size vendors pivoting to platforms. Old monitoring vendors becoming "AI observability" vendors. Old config management vendors becoming "platform engineering" vendors. Old service mesh vendors becoming "zero trust networking" vendors. Some of these pivots are real and some are press releases.
If you are evaluating tooling in early 2025, the question to ask is "does this help us run things in production" not "is this technology cool". The answer separates real vendors from the rest very quickly.
The hallway track that mattered
I had three conversations that have stuck:
- A platform lead at a mid-size bank: "We stopped trying to build the perfect platform. We built the boring one. Now developers like us."
- An SRE at a hyperscaler: "Half our toil is still YAML. Not Kubernetes' fault, but the abstraction never landed where it needed to."
- A CTO at a Series B startup: "I would not start on Kubernetes today for a five-engineer team. Fly.io, Render, ECS, anything else. The complexity tax is too high."
That last one is the post-honeymoon honesty the ecosystem needed for a long time. Kubernetes is not the right answer for everyone. It is the right answer for a specific set of problems at a specific scale. Maturity means we can say that out loud.
What I am taking back to clients
Three things I am pushing into client engagements off the back of these two conferences:
- If you have not adopted OpenTelemetry by mid-2025 you are paying a vendor lock-in tax for no reason.
- If your platform team is not measuring developer experience, your platform is invisible at best and resented at worst.
- If you are running GPU workloads on the cluster, treat them as a different class of resource with different scheduling, different SLOs, and probably different teams. Mixing them with regular pod scheduling is a footgun.
KubeCon 2025 will be all of this, more, and more boring. Boring is good.