Skip to content
Lakehouse Compare

Guide, read 10 Nov 2026

The open stack on AWS: Spark on EMR or EKS with Iceberg, Delta or Hudi

All guidesOpen the Compute Bill

The compute layer: EMR on EC2, EKS or Serverless

AWS describes the EMR runtime for Spark as 100% API-compatible with open-source Spark, and EMR runs on EC2, on EKS or as Serverless. EMR on EKS lets teams run EMR-based applications beside other workloads on one EKS cluster, with Airflow or MWAA for scheduling. It supports the Kubernetes Operator for Apache Spark as a job submission model on your own clusters.

EMR Serverless needs no cluster configuration: it determines the resources a job needs and releases them afterward. Pricing for all three modes is on the EMR pricing page and summarized in EMR cost optimization.

The table layer: what EMR bundles

Amazon EMR 7.14.0 packages Spark 3.5.8-amzn-0 with Delta 3.3.2-amzn-3, Iceberg 1.10.1-amzn-1 and Hudi 1.0.2-amzn-3. The EMR Spark feature page says EMR supports Apache Iceberg v3 and Spark 4.0. All three formats are available; you are not forced into one by the platform.

For Hudi, EMR installs the Hudi components by default when Spark, Hive, Presto or Flink is installed, from release 5.28.0, and Hudi is integrated with Trino from 6.1.0. The Delta Lake page lists Delta on EMR from release 6.9.0.

The catalog layer: Glue Data Catalog and S3 Tables

The AWS Glue Data Catalog stores table metadata and integrates with Athena, Lake Formation, EMR and SageMaker. Its Iceberg REST endpoint implements the Apache Iceberg REST specification, supports table specifications v1 and v2 with v2 the default, and for us-east-1 is the endpoint glue.us-east-1.amazonaws.com/iceberg (an API address, not a web page).

S3 Tables stores tables in the Iceberg format in table buckets. Glue ETL jobs can reach S3 Tables through the analytics-services integration, through the S3 Tables Iceberg REST endpoint, or through a Java catalog library. S3 Tables needs Glue version 5.0 or higher, Iceberg v3 needs Glue 5.1 or higher, and the v3 data types need Glue 6.0 or higher. The catalogs guide covers the alternatives.

What you operate

In an open stack you own the maintenance. Iceberg's documentation lists expiring snapshots, removing old metadata files and deleting orphan files as recommended, and compaction and manifest rewriting as optional, all as jobs you schedule. AWS offers managed versions: Glue table optimizers for Iceberg compaction, snapshot retention and orphan file deletion, and S3 Tables maintenance on by default. Athena offers manual OPTIMIZE and VACUUM statements.

You also own Spark tuning, upgrades and the choice of purchase options. EMR enables adaptive query execution, dynamic partition pruning and bloom filter joins, and the docs say adaptive join conversion is on by default.

Published performance and cost for this stack

The one published run of this stack is AWS's own: a TPC-DS 3 TB test on 9 r5d.4xlarge nodes with a Hadoop catalog, in which EMR 7.12 averaged 1,349.62 seconds and open-source Spark 3.5.6 with Iceberg 1.10.0 averaged 6,113.92. AWS's stated costs were $4.81 and $17.65 on On-Demand prices. The setup is in the aws-samples/emr-on-eks-benchmark and aws-samples/emr-spark-benchmark repositories, and AWS notes the results are not comparable to official TPC-DS.

No third-party run of the open stack against managed lakehouses was found on 2026-11-10, so cost comparisons on this site stay in the compute bill, which uses published prices and states every assumption.

Kubernetes and the Spark operator

EMR on EKS applies its uplift to any pod that uses an EMR runtime image, including pods outside job submission. The Kubeflow Spark Operator that EMR on EKS supports is Apache-2.0 licensed and installs with Helm. Self-run Spark on EKS with the open-source operator needs no job rewrite, but you then run the Spark build yourself and carry none of the EMR uplift.

EMR Serverless bills only the vCPU and memory of workers while they run, so a job that scales in and out is billed for less than its peak size held for the whole time. The compute bill line for EMR Serverless does not model that.

Glue as another place to run Spark

AWS Glue ETL also runs Spark. A DPU is 4 vCPU and 16 GB of memory, Spark jobs need at least 2 DPUs and default to 10, and billing is per second. The us-east-1 price list shows $0.308 per DPU-hour for Glue 6.0 and later and $0.44 for older versions. Glue sits outside EMR's uplift model, so a job moved between them changes the pricing unit, not only the price.

Region and cloud limits

All of these services run only on AWS. EMR has regional endpoints across the US regions (us-east-1, us-east-2, us-west-1, us-west-2) and many others. Glue Iceberg compaction pricing, S3 Tables pricing and EMR uplifts differ by region; the figures on this site are us-east-1 list prices read 2026-11-10.

Sources

  1. AWS Big Data Blog, EMR 7.12 and Iceberg benchmark, read 10 Nov 2026
  2. Amazon EMR pricing, read 10 Nov 2026
  3. Amazon EMR on EKS, read 10 Nov 2026
  4. EMR on EKS Spark operator, read 10 Nov 2026
  5. Amazon EMR Serverless, read 10 Nov 2026
  6. Amazon EMR 7.x release versions, read 10 Nov 2026
  7. Delta Lake on Amazon EMR, read 10 Nov 2026
  8. Apache Hudi on Amazon EMR, read 10 Nov 2026
  9. AWS Glue Data Catalog, read 10 Nov 2026
  10. AWS Glue Iceberg REST endpoint, read 10 Nov 2026
  11. Amazon S3 Tables overview, read 10 Nov 2026
  12. S3 Tables with AWS Glue, read 10 Nov 2026
  13. Apache Iceberg maintenance, read 10 Nov 2026
  14. AWS Glue optimizing Iceberg tables, read 10 Nov 2026
  15. Athena Iceberg optimization, read 10 Nov 2026
  16. Amazon S3 Tables maintenance, read 10 Nov 2026
  17. AWS Glue pricing, read 10 Nov 2026
  18. AWS price list, AWS Glue, us-east-1, read 10 Nov 2026

Last reviewed 10 Nov 2026.