“We were essentially able to reduce the cost of that cluster by about 75%. On AWS, DevZero demonstrated they could achieve significantly higher savings than we initially thought possible.”
Mihir Nair
Head of Architecture, Databahn
Cut GCP GKE costs without
restarting your workloads
Rightsize CPU and memory, provision optimal Compute Engine machine types, all without restarting a single pod.
Companies who slashed their Kubernetes
spend using DevZero
Your GKE cluster is running, your bill is too high
Most GKE clusters run at 15–30% utilization. Teams overprovision out of caution, node pools stay fixed at peak size, and egress charges pile up with no per-workload view.
Requests are set once and never revisited
CPU and memory requests are set at deploy time and left untouched. Most GKE pods run at a fraction of what they request. You pay for the full Compute Engine allocation regardless of actual usage.
Node pools stay fixed at peak size
GKE node pools are sized for peak load with a fixed replica count. Underutilized nodes keep running across every zone with no automatic way to consolidate or right-size them.
Cross-zone and egress costs are invisible
GCP charges ~$0.01/GB for cross-zone traffic and ~$0.12/GB for internet egress. Cloud Billing rolls both into a single Compute Engine line with no workload attribution and no way to know where to cut.
One platform that actually reduces your GKE bill
Lightweight operators observe usage, generate recommendations, and apply changes when you are ready.
dakr-op · WorkloadRecommendation applied
Before
api-server-deploy
cpu: 2000m
mem: 4Gi
CPU used17%
Mem used22%
CRIU
checkpoint
restore
in-place
After
api-server-deploy
cpu: 340m
mem: 900Mi
✓ 0 pod restarts
✓ connections live
✓ GKE SLO intact
WorkloadRecommendation · devzero-system
Method
CRIU live
Scope
21 workloads
Savings
−$7,800/mo
Rightsize requests without restarting pods
dakr-op receives recommendations via gRPC, validates them against your policy, then applies CPU and memory changes in place using CRIU checkpoint-restore. Zero restarts. Zero dropped connections.
dzKarp · GCENodeClass provisioning
Compute-heavy
ml-worker · batch-job
Memory-heavy
postgres · pubsub-fwd
General-purpose
api-server · frontend
dzKarp · GCENodeClass
spotToSpotConsolidation · implicitPDB 20%
c2-standard-8
compute-optimized
74% CPU
m2-standard-4
memory-optimized
68% CPU
n2-standard-4
general-purpose
61% CPU
Static pools consolidated · Spot VM interruptions handled automatically
Replace static node pools with right-sized GCE instances
dzKarp provisions the optimal Compute Engine machine type per workload class, consolidates idle nodes, and handles Spot VM interruptions automatically. Installed via Helm with Workload Identity. No static credentials.
zxporter-netmon · eBPF flow attribution
| Workload | Type | Vol | Rate | $/mo |
|---|---|---|---|---|
| frontend | Cross-zone | 921 GB | $0.01/GB | $9.21 |
| ml-worker | Cross-zone | 640 GB | $0.01/GB | $6.40 |
| api-server | Internet | 142 GB | $0.12/GB | $17.04 |
| pubsub-fwd | Peering | 380 GB | $0.01/GB | $3.80 |
| Total attributed | 2,083 GB | $36.45/mo | ||
Attributed per pod via eBPF DNS tracing · no payload inspection
See cross-zone and egress cost per workload
zxporter-netmon runs as a DaemonSet using eBPF to trace pod-level flows. It maps each flow to GCP pricing (cross-zone ~$0.01/GB, internet egress ~$0.12/GB) and attributes cost to the workload that generated it.
Before and after with DevZero on GKE
Same cluster, same workloads. Less spend.
Before DevZero
Avg cluster CPU utilization20%
api-server CPU request2000m (set at deploy)
Node type selectionManual, static node pools
Spot VM interruption handlingManual drain scripts
Cross-zone traffic visibilityNot tracked per workload
Monthly infra spend$49,600 / mo
After DevZero
Avg cluster CPU utilization68% (rightsized)
api-server CPU request340m (live migration)
Node type selectionAuto Compute Engine provisioning
Spot VM interruption handlingAutomatic live migration
Cross-zone traffic visibilityPer-pod breakdown via eBPF
Monthly infra spend$34,720 / mo (−30%)
Getting started in three steps
bash · prod-gke
# Install read operator (zxporter) via Helm$ helm upgrade --install zxporter \oci://public.ecr.aws/devzeroinc/zxporter \--namespace devzero-zxporter \--set config.clusterToken=<token>✓ namespace devzero-zxporter created✓ 3/3 pods running✓ metrics streaming to DevZeroread-only · no cluster changes made
devzero.io / optimization / prod-gke
21 workloads scanned−$12,640 potential
WorkloadNowRec.
api-server2000m340m
postgres4Gi1.1Gi
ml-worker4000m1380m
First scan complete · No changes applied yet
Read-only · Open-source at github.com/devzero-inc/zxporter · No cluster changes until you enable automation
Deploy the open-source, read-only zxporter via Helm. See cost recommendations in your DevZero dashboard with no cluster changes needed. First scan surfaces savings within 24 hours.
bash · gcloud + helm
# 1: Create GCP service account (karpenter-gsa)$ gcloud iam service-accounts create \karpenter-gsa --project=$PROJECT_ID✓ roles/compute.admin granted✓ roles/container.admin granted# 2: Create dzkarp node pool (e2-medium)$ gcloud container node-pools create dzkarp \--workload-metadata=GKE_METADATA \--machine-type=e2-medium --num-nodes=2✓ Workload Identity binding set
kubectl get gcenodeclass · devzero-system
Kubernetes objects created
GCENodeClassdevzero-classReady
NodePooldevzero-defaultReady
GCE nodes provisioned by dzKarp
c2-standard-8compute-optimized×3
m2-standard-4memory-optimized×1
Workload Identity binding active · dzKarp managing node provisioning · Spot VM interruptions handled automatically
Create a GCP service account (karpenter-gsa), grant compute and container admin roles, create the dzkarp node pool (e2-medium), set the Workload Identity binding, then install dzKarp via Helm from the public ECR registry.
Workload policy · dakr-op
Scope
namespace = production
Workload types
Deployment · StatefulSet · DaemonSet
Migration type
Live (CRIU)Rolling restart
Aggressiveness
devzero.io / WorkloadRecommendations
WorkloadBeforeAfterType
api-server2000m340mlive
frontend1000m290mlive
ml-worker4000m1380mrestart
Savings this month−$7,800
Automation enabled · dakr-op applies WorkloadRecommendations via policy · Savings begin immediately
Review WorkloadRecommendations in the DevZero dashboard, create workload policies for dakr-op, then scale down static node pools to hand provisioning to dzKarp.
Everything GKE teams need on one platform
Workload rightsizing without downtime
CRIU applies new CPU and memory limits without restarting the pod. Connections stay open. Each recommendation is labelled live or restart before you approve.
View write operator →Intelligent node autoscaling on GKE
The DevZero node operator provisions the right Compute Engine machine type per workload, consolidates idle nodes, and handles Spot VM interruptions automatically. Managed by policy, not YAML.
GKE setup guide →Cross-zone and egress network cost visibility
An eBPF DaemonSet tracks pod-level flows and maps each one to GCP pricing: inter-zone, inter-region, internet egress, and Cloud NAT. No changes to CoreDNS or kube-proxy.
Network operator docs →Security scanning and compliance
Scans container images for CVEs, checks clusters against CIS Kubernetes Benchmark and GKE hardening guidelines, and flags RBAC misconfigurations. All results in one dashboard.
Security operator →Terraform and Pulumi providers
Manage clusters and policies as code using the official Terraform provider or Pulumi SDKs for TypeScript and Python. Fits existing GitOps and Config Connector workflows.
IaC providers →Cost-aware pod scheduling
dz-scheduler places pods on the cheapest viable nodes based on real-time Spot VM pricing, cross-zone transfer cost, and CRIU compatibility.
Scheduler docs →What our customers say
Frequently asked questions
Technical questions from platform engineers who've evaluated DevZero on GKE.
Stop paying for compute you're not using.Start rightsizing today.
Connect your GKE cluster in under 30 minutes. No code changes. No pod restarts. First savings visible within 24 hours.
Get started free