This post was originally published on this site

GKE Dataplane V2 is Google Cloud’s next-generation networking dataplane for Google Kubernetes Engine (GKE), built on extended Berkeley Packet Filter (eBPF) and open-source Cilium to replace traditional iptables routing and kube-proxy. We built GKE Dataplane V2 to meet the performance, scalability, and security demands of modern cloud-native workloads.

GKE customers consistently tell us that while they want to adopt GKE Dataplane V2 for its performance and scale, migrating production clusters carrying stateful workloads and established IP allocations has been challenging, often requiring a disruptive parallel cluster setup. Today, we are pleased to introduce the preview of in-place migration, allowing you to transition eligible VPC-native clusters from GKE Dataplane V1 to GKE Dataplane V2 directly.

Let’s take a look at what GKE Dataplane V2 brings to the table and how to start migrating your environment to it. 

Why GKE Dataplane V2?

GKE Dataplane V2 is built on the extended Berkeley Packet Filter (eBPF) and open-source Cilium, with features specifically tuned for performance, scalability, and deep integration with Google Cloud networking. Consider the following capabilities:

  • Scale and performance: By handling packet processing with eBPF maps in the Linux kernel instead of linear iptables rules, GKE Dataplane V2 supports clusters of up to 15,000 nodes with network policy enforcement (and up to 65,000 nodes when network policy enforcement is disabled, which requires recreating clusters). This architecture also handles the high Pod churn, fast sandbox provisioning, and network identity scaling required by large agentic AI workloads.

  • High-throughput accelerator networking: Serves as the foundational networking layer for GPU and Cloud TPU clusters on GKE, helping keep accelerators computing rather than waiting on network initialization. It also powers capabilities like GKE managed DRANET for Dynamic Resource Allocation (DRA) of network devices.

  • Network security and observability: With GKE Dataplane V2, standard Kubernetes NetworkPolicy enforcement and hierarchical ClusterNetworkPolicy work directly in the kernel with optimized performance. GKE Dataplane V2 also provides the underlying dataplane for sidecarless GKE ambient networking, while GKE Dataplane V2 observability delivers native network policy logging, live traffic topology, and Hubble flow metrics.

The in-place migration workflow

Instead of having to manually move workloads to a new cluster, GKE now automates the transition across three stages, as described in the GKE Dataplane V2 migration guide:

1. Eligibility and preflight checks: Before starting, verify that your Pod IP ranges have enough capacity to recreate the nodes by using GKE IP address utilization insights, and check Active Assist for incompatible configurations like custom eBPF programs on managed interfaces.

2. Initiate the migration: Once you’ve verified the preflight checks, start the migration with a single Google Cloud CLI update command.

To initiate the migration on an eligible cluster, run:

code_block
<ListValue: [StructValue([('code', 'gcloud container clusters update CLUSTER_NAME \rn –location=LOCATION \rn –enable-dataplane-v2'), ('language', ''), ('caption', )])]>

3. Automated rollout and finalization: GKE updates your node pools to GKE Dataplane V2 while respecting your configured maintenance windows, exclusions, and upgrade strategies. To move faster, you can also patch specific node pools manually, or roll back to GKE Dataplane V1 at any time prior to finalization.

In-place migration supports VPC-native clusters running GKE 1.35 or later (see the migration guide for exact version requirements). It does not support routes-based clusters or clusters running Calico. 

Get started

In-place migration to GKE Dataplane V2 is available now in preview. During this time, we recommend testing migrations on non-production or test clusters first as you plan your broader transition.

To learn more about GKE Dataplane V2 and to get started: