Google Cloud has opened a preview that enables VPC‑native GKE clusters to switch from Dataplane V1 to Dataplane V2 directly, eliminating the need for a separate parallel cluster. The shift replaces the iptables‑based routing and kube‑proxy model with an eBPF‑driven dataplane built on Cilium, and it brings observable performance and security improvements that matter to AI, platform, DevOps, and security teams.
What changed with GKE Dataplane V2 migration
- Kernel‑level packet processing: eBPF maps handle traffic inside the Linux kernel, removing the linear
iptablesrule chain used by Dataplane V1. - Node scalability: The new dataplane supports up to 15,000 nodes when
NetworkPolicyenforcement is active, and up to 65,000 nodes when policy enforcement is disabled (cluster recreation required). - Accelerator‑friendly networking: Designed as the base layer for GPU and Cloud TPU workloads, it reduces network initialization latency and powers GKE managed DRANET for dynamic resource allocation.
- Integrated security and observability:
NetworkPolicyand hierarchicalClusterNetworkPolicyare enforced directly in the kernel, and native observability provides policy logging, live traffic topology, and Hubble‑style visibility.
Why the change matters for AI and platform workloads
- High‑throughput AI jobs: Large agentic AI workloads generate rapid pod churn and require fast sandbox provisioning; the eBPF dataplane scales to those demands without the bottlenecks of iptables.
- Accelerator utilization: By keeping network latency low, GPU and TPU clusters spend more time computing and less time waiting for network setup.
- Policy enforcement at scale: Teams can enforce fine‑grained
NetworkPolicyacross thousands of nodes without a proportional increase in CPU overhead.
Operational and security considerations
- Eligibility: Only VPC‑native clusters qualify for the in‑place migration preview; clusters using routes‑based networking must be re‑configured first.
- Node limits: When retaining
NetworkPolicyenforcement, the 15,000‑node ceiling applies. Exceeding it requires disabling policy enforcement and recreating the cluster, which may affect security posture. - Observability transition: Existing monitoring pipelines that rely on iptables metrics need to be updated to consume the new dataplane’s native logs and topology data.
- Ambient networking compatibility: The dataplane underpins sidecar‑less ambient networking; teams adopting ambient mode should verify that their service mesh expectations align with the new kernel‑level implementation.
- Migration risk: Although the preview avoids a parallel cluster, practitioners should test the migration in a staging environment, validate pod IP continuity, and confirm that any custom iptables rules are no longer required.
Related CloudNinjas coverage: Google Cloud.
What This Means For Practitioners
Plan a pilot migration on a non‑critical VPC‑native cluster to confirm that your workloads retain expected IP allocations and that policy enforcement behaves as intended. Update monitoring and alerting to ingest the dataplane’s native logs and topology feeds. If you run GPU/TPU workloads or anticipate scaling beyond 15,000 nodes, evaluate the trade‑off of disabling NetworkPolicy enforcement versus recreating the cluster. Finally, track the preview’s graduation timeline to ensure you adopt the fully supported migration path before committing production workloads.


