# EMR Serverless vs EC2 vs EKS — Decision Matrix (July 2026)

| Question | Choose |
| --- | --- |
| Need HBase / Presto / Hue / full EMR app set? | **EMR on EC2** |
| Already EKS-first and want bin-packing with other pods? | **EMR on EKS** |
| Sustained 24/7 Spark, Spot + RI economics critical? | **EMR on EC2** (Spot task nodes) or EKS + Karpenter |
| Intermittent batch, team wants zero cluster ops? | **EMR Serverless** |
| Need Spot directly on Serverless? | Not available — use EC2/EKS |

## Cost notes

- **Serverless:** vCPU-second + GB-second (+ storage beyond included); 1-minute minimum; pre-initialized capacity bills while warm.
- **EC2:** EMR uplift + EC2 (+ Spot savings); idle clusters burn money.
- **EKS:** EC2/EKS control plane + EMR on EKS uplift; share nodes with other workloads.

## Default for new intermittent Spark

Start **Serverless**; move to EC2/EKS when utilization data proves RI/Spot wins.

## Lakehouse

All three modes can share Iceberg tables via Glue Data Catalog with Glue 5.x jobs.