$
We run container orchestration benchmarks at scale — cold-start latencies in microseconds, resource contention across thousand-node clusters, multi-cloud cost models — and publish the data that platform teams actually cite in architecture decisions.
Cold-Start Latency Distribution
p50 / p95 / p99 · microseconds · n=50,000 runs
Pod Schedule Time · Node Count
Runtime Memory Overhead
▸ Research Findings
The data speaks.
We just publish it.
Three findings from our 2025 benchmark cycle. Each represents hundreds of hours of test runs across isolated, reproducible environments.
Median cold-start · k3s · ARM64 · n=50,000
k3s starts containers 3.1× faster than ECS Fargate on equivalent hardware.
At sub-100μs p50 latency, k3s on ARM64 nodes redefines what's possible at the edge. This gap isn't a configuration artifact — it's architectural. Fargate's virtualization layer adds 150–200μs of irreducible overhead that no amount of tuning removes.
Methodology
crictl run · containerd 1.7 · Graviton3 · 50K cold starts per orchestrator · isolated VPC
p99 pod schedule time · 1,024-node K8s cluster · high contention
At 1,000+ nodes, Kubernetes scheduler latency compounds faster than most teams model.
The p99 scheduling time on a 1,024-node cluster under 80% resource contention reaches 487ms — 4.2× the p50. Teams that capacity-plan on median latency are shipping SLAs they can't actually meet during traffic spikes. The tail is where production incidents live.
Methodology
kube-scheduler profiling · 1,024 c5.2xlarge nodes · 48K pods · 80% CPU contention · 10K scheduling events
Cost reduction · ECS → K8s migration · batch workloads · 500-node scale
Migrating batch workloads from ECS to Kubernetes cuts infrastructure spend by 34% at 500-node scale.
Bin-packing efficiency, spot instance integration, and Karpenter's node consolidation close the gap that once made ECS the cost-rational choice. At 500 nodes, the operational overhead of K8s is absorbed. Above 200 nodes, the TCO math consistently favors Kubernetes — and our data shows exactly where the crossover happens.
Methodology
TCO model · 12-month AWS billing data · ECS Fargate vs EKS Karpenter · batch + streaming workloads · same application fingerprint
These are three findings from 94 pages of data.
The methodology and full dataset are in the report.
▸ Research Library
Three reports.
One rigorous dataset.
Annual Benchmark Report · 2025
2025 State of Container Orchestration
Full p99 latency distributions across 6 orchestrators, 3 cloud providers, and 4 runtime environments. Methodology appendix included.
Architecture Decision Report · 2025
Multi-Cloud Cost Modeling for K8s Workloads
TCO analysis across AWS EKS, GKE, and Azure AKS for batch, streaming, and stateful workloads at 100–10,000 node scale.
Comparative Analysis · 2025
Edge Orchestration: k3s vs Nomad at the Boundary
Cold-start benchmarks under constrained resources: ARM64 nodes, intermittent connectivity, and sub-1GB memory profiles.
▸ Methodology
Rigorous by design.
Reproducible by default.
Architecture decisions made without benchmark data are guesses. We built the infrastructure to eliminate the guesswork — for good.
Isolated Test Environments
Every benchmark runs in a freshly provisioned cluster — no shared tenancy, no warm caches, no prior state contaminating results.
Statistical Significance
Minimum 10,000 iterations per data point. We report p50, p95, and p99 with confidence intervals. Outliers are documented, not discarded.
Vendor-Independent
No cloud provider funds this research. Infrastructure costs are self-funded. Results are published regardless of which orchestrator wins.
Reproducible
Full benchmark harness published on GitHub. Any team can reproduce our numbers on their own infrastructure within 4 hours.
▸ Decisions Without Data Are Guesses
The 2025 State of Container Orchestration report is free.
94 pages. Six orchestrators. Three cloud providers. Full p99 distributions, cost models, and a methodology appendix your team can reproduce. No vendor bias. No paywall — just a work email.
14,200+ downloads · cited by engineering teams at 340+ companies