Managed Kubernetes is one of the more portable things you can run in a cloud, because the API you deploy against is standardised. A manifest that works on one provider’s cluster will usually work on another’s. The differences are around the edges, and the edges are where the operational work lives.
The control plane
Every managed offering runs the control plane for you, but the commercial and operational terms vary. Some providers charge an hourly fee per cluster for it, others include it and charge only for worker nodes. Upgrade behaviour differs too: some apply control plane upgrades on a schedule you influence rather than control, others leave the timing entirely to you. If you run many small clusters, a per cluster fee changes the economics of how you divide workloads.
Networking is the biggest practical difference
Kubernetes defines how pods communicate but leaves the implementation to a network plugin, and providers choose differently. The consequences are concrete.
- Address allocation. Some integrate pod addressing directly into the provider’s own network, so every pod holds a routable address. That simplifies connectivity to other resources and makes address range planning important, because large clusters consume a lot of addresses.
- Load balancer provisioning. A service of type load balancer creates a provider resource, and what you get, how long it takes and what it costs vary.
- Network policy support. Whether pod level network policy is enforced depends on the plugin in use, and it is not always on by default.
Storage
Persistent volumes map onto the provider’s block storage through a driver. The portable part is the volume claim; the non-portable parts are the storage class names, the performance tiers available and whether a volume can be attached to more than one node. Workloads that assume shared read-write access need checking against each provider’s offering, because support for it is uneven.
Identity and permissions
Connecting a pod to the provider’s own permission system is where the most provider-specific configuration usually lives. Each platform has its own mechanism for letting a workload assume a cloud identity without a static key. It is worth setting up properly, because the alternative is long lived credentials in cluster secrets, but it is the piece least likely to transfer when you move.
Node management
Offerings differ in how much of the node layer you own. Some present pools of virtual machines you size and scale. Others abstract nodes away so you specify the workload and the platform provides capacity. The second reduces operational work and usually costs more per unit of compute; the first gives more control and more to look after.
How to keep a cluster reasonably portable
- Keep manifests free of provider-specific annotations wherever a standard field exists.
- Reference storage classes through a name you define rather than a provider default.
- Put the provider-specific pieces, meaning identity integration, ingress controller configuration and storage classes, in a clearly separated layer.
- Plan address ranges deliberately, especially on platforms where pods consume network addresses.
- Do not depend on a specific Kubernetes minor version, since upgrade cadence is not yours to set.
None of this makes a cluster trivially movable, but it does mean the parts that need rewriting are identified and contained rather than scattered through every manifest.