TL;DR 
 On a large bare-metal Kubernetes cluster, kube-proxy was consuming over 100 CPU cores spread across nodes. We were already on IPVS — technically “already optimized”. And yet. Migrating to nftables mode (GA since Kubernetes 1.33) returned that budget to application teams, cut rule sync time from minutes to seconds, and smoothed out tail latencies. Below: how we diagnosed it,…
One of those settings people copy-paste without thinking. Until it causes an incident. 
 TL;DR 
 IfNotPresent with immutable tags — fast, predictable, registry stays quiet. Always with mutable tags ( latest , stable ) — if you know what you’re doing and your registry can handle it. Using latest without Always is the classic way to ship a “hotfix” that changed nothing.
Infrastructure · Kubernetes · Docker 
 Most of my clients come through referrals. If someone sent you here, they probably gave you the short version: I find what’s broken, risky, or silently costing you money in your container stack — and I tell you exactly what to do about it. 
 I work on production clusters, not demo environments. If you’ve read any of my posts , you have a…