Choosing a compute model
Use the tree as a first cut, then validate with a pilot. If your goal is almost no node operations and default behaviour is acceptable, Auto Mode is the simplest. If your fleet is mixed, spiky or Spot-heavy and you want to tune instance choice and consolidation, self-run Karpenter gives the best cost control. If the fleet is steady and predictable, or your organisation wants exact control of AMIs and groups, managed node groups remain solid and familiar.
Fargate is a niche inside the cluster rather than the whole answer: it suits small stateless workloads and strong per-Pod isolation without node management, with the DaemonSet and hostNetwork constraints you saw earlier. Combining models is normal; for example Karpenter or Auto Mode for general workloads plus a small managed node group dedicated to cluster-critical components.
A cost-shape reminder: steady always-on workloads are usually cheapest on EC2 with commitments (Savings Plans) or Spot for tolerant parts; spiky workloads with long idle gaps can favour Fargate or aggressively consolidating Karpenter. Measure utilisation before deciding, and revisit as workloads change.
Shopwave's choice: orders and loyalty run on Karpenter with a Spot-weighted general pool and a small on-demand pool for the stateful pieces, with a managed node group for system components during the transition.