Performance

Docker is everywhere. Every application runs in containers. Every deployment uses Docker. Every team containerizes everything. But here’s the thing: Docker adds a runtime layer between your application and the OS. That layer has overhead. That overhead costs money. Containers aren’t free. They consume CPU. They consume memory. They consume disk space. They add complexity. They add operational burden. Most applications don’t need containers. Most applications can run directly on the OS. Most applications don’t need the isolation. Most applications don’t need the portability. ...

Service meshes are everywhere. Istio. Linkerd. Consul Connect. Every microservices architecture needs one. Or so the marketing says. But here’s the thing: service meshes add sidecar proxies to every pod. Envoy, Istio’s sidecar, uses 50-200 MB RAM per pod. Linkerd-proxy uses 20-100 MB. Multiply by hundreds of pods. That’s gigabytes of memory just for service mesh overhead. All of this before your applications run. All of this just for inter-service communication. All of this overhead. ...

etcd sits at the heart of Kubernetes. Before your applications run, etcd is storing cluster state, coordinating elections, and replicating data. It consumes 2-8 GB RAM per node. It requires 3-5 nodes for high availability. That’s 6-40 GB RAM just for cluster coordination. Most teams don’t need distributed consensus. Most teams don’t need high availability at the cluster level. Most teams are running small clusters that would work fine with a single node and backups. ...

Cloud-first is the default. Every startup uses AWS. Every enterprise migrates to Azure. Every consultant recommends GCP. But here’s the thing: 37signals went from $3.2M per year to $1.3M per year after leaving the cloud. Over $10M saved in five years. GEICO spent a decade migrating to the cloud. Result: 2.5x higher costs. They’re not alone. The cloud isn’t always cheaper. It’s often more expensive. Especially when you factor in hidden costs: egress fees, managed services, vendor lock-in. ...

Amazon Prime Video cut costs by 90% by moving away from microservices back to a monolith.

NGINX sits between your users and your application. Before a single request reaches your code, NGINX is parsing configs, terminating SSL, rewriting URLs, and logging everything. All of this overhead. All of this complexity. The Ingress-NGINX controller is being retired in March 2026. About 50% of cloud-native setups depend on it. No more fixes. No more patches. Migrating means rewriting ingress configs across hundreds of services. Staying means increasing security risk. Pick your poison. ...

Kubernetes was built to orchestrate Google’s global infrastructure. You are not Google. Terribly sorry. 82% of container users run Kubernetes in production. Most of them shouldn’t. The Control Plane Tax Before your application serves a single request, Kubernetes needs etcd chewing through 2-8 GB RAM per node. Then kube-apiserver, kube-scheduler, kube-controller-manager, kubelet (reserving 25% of node memory by default), CoreDNS, kube-proxy, and a CNI plugin. All of this before your code runs. ...