Infrastructure economics & operations
Kubernetes operations, FinOps practice, platform engineering, and AI infrastructure — argued with specifics. No vendor summaries, no neutral surveys, no predictions about the next decade.
Kubernetes FinOps Platform Engineering AI Infrastructure Cloud Security
You can right-size every workload perfectly and still run at 35 percent node utilisation. The gap between allocated and provisioned is a packing problem, and the scheduler defaults are not trying to solve it.
How .ai domain pricing actually behaves: registration volume, comparable sales data, the two-year renewal structure, and a framework for valuing a premium technical domain.
A complete right-sizing process: percentile selection, workload classification, safe rollout ordering, verification, and the organisational mechanics that make the changes stick.
Planning capacity for spiky traffic: arrival-rate modelling, the provisioning latency budget, pre-warmed pools versus overprovisioning, and how to size headroom with real numbers.
The honest cost of running a service mesh: p99 latency, sidecar CPU and memory, ambient mode's actual savings, and the four situations where a mesh pays for itself.
A grounded look at eBPF for runtime security and observability: what the syscall boundary reveals, real CPU overhead numbers, kernel version constraints, and where the vendor demos mislead.
Wiring Sigstore keyless signing into Kubernetes admission control with cosign and policy-controller — the rollout sequence, the failure modes, and how to enforce without an outage.
Controlling logging, metrics, and tracing spend: cardinality budgets, tail-based sampling, tiered retention, and OpenTelemetry Collector patterns that cut cost without losing incident signal.
The recurring failure modes of internal developer platforms — mandatory abstractions, YAML that generates YAML, ticket queues in self-service costume — and how to recover from each.
Treating an internal platform as a product: finding the real jobs to be done, measuring adoption honestly, running discovery with captive users, and knowing when to deprecate.
An honest accounting of multi-cloud portability: what the abstraction costs in engineering velocity, where the leaks appear, and the specific scenarios that justify paying for it.
The architectural split between training and inference infrastructure: interconnect requirements, scheduling models, checkpointing, autoscaling behaviour, and the cost structures that follow from each.
A technical comparison of GPU sharing on Kubernetes — time-slicing, MPS, MIG, and dedicated allocation — with the fairness, isolation, and utilisation tradeoffs of each.
How to build infrastructure unit economics that survive contact with product teams: choosing the denominator, handling fixed cost, tracking marginal versus average, and the traps in cost-per-request.
What FinOps maturity looks like past the reporting stage — forecast accuracy, commitment discipline, anomaly routing, and embedding cost into the engineering workflow rather than a monthly review.
Allocation models for Kubernetes cost: idle distribution, shared-service treatment, requests versus usage as the billing basis, and how to build a showback report engineers actually believe.
When spot capacity is safe for databases, queues, and caches — interruption modelling, capacity-pool diversification, PDB configuration, and the recovery-time math that decides it.
A practical comparison of Karpenter and cluster-autoscaler: provisioning latency, bin-packing quality, disruption behaviour, consolidation risk, and which one fits which cluster.
Why vertical and horizontal pod autoscaling oscillate when they share a metric, the specific configurations that are safe, and how to use VPA in recommendation-only mode as a right-sizing engine.
CPU and memory requests and limits solve different problems. A concrete methodology for setting each, why CPU limits usually hurt, and how throttling shows up in your latency percentiles.