Not a vendor. Components are current: Amazon EMR release 7.14.0 ships Spark 3.5.8-amzn-0, Iceberg 1.10.1-amzn-1, Delta 3.3.2-amzn-3 and Hudi 1.0.2-amzn-3; Iceberg docs show latest 1.12.0; the Kubeflow Spark Operator (Apache-2.0) was pushed on 2026-11-10.
Read
10 Nov 2026
What it does well
Per AWS documentation, EMR 7.14.0 packages Spark 3.5.8-amzn-0, Delta 3.3.2-amzn-3, Iceberg 1.10.1-amzn-1 and Hudi 1.0.2-amzn-3, so all three formats are available (read 2026-11-10).
AWS ran TPC-DS 3 TB on 9 x r5d.4xlarge and reports $4.81 for EMR 7.12 against $17.65 for open-source Spark 3.5.6 with Iceberg 1.10.0, with repositories to repeat it.
Where it is limited
Per the Apache Iceberg documentation, users run table maintenance themselves: expire snapshots, remove orphan files and compact with rewriteDataFiles as jobs they schedule (read 2026-11-10).
A cost comparison of open-stack Spark against managed lakehouses by a party other than AWS or a vendor: none found on 2026-11-10.
Published speed results
Fact
Source type
Source, read
AWS ran TPC-DS 3 TB (104 SparkSQL queries, 3 rounds, 9 x r5d.4xlarge, Hadoop catalog) and reports EMR 7.12 averaged 1,349.62 s versus 6,113.92 s for open-source Spark 3.5.6 with Iceberg 1.10.0 and 1,535.62 s for EMR 7.5. AWS notes results are not comparable to official TPC-DS. The run is by the vendor of EMR.
EMR on EC2 charges an EMR price in addition to the EC2 price (and EBS), billed per second with a one-minute minimum. AWS's example in us-east-1: c4.2xlarge EMR charge $0.105/hour next to EC2 $0.398/hour; one master and two core nodes for a month total $1,101.57. Spot instances are up to a 90% discount versus On-Demand.
EMR on EKS adds an uplift per vCPU-hour and per GB-hour to EKS/EC2 or Fargate costs, plus $0.10 per hour for each EKS cluster. AWS's worked example uses $0.007084 per vCPU-hour and $0.000777875 per GB-hour (100 vCPU and 300 GB for 30 minutes = $0.4709 EMR uplift). The uplift applies to any pod using an EMR runtime image, including with the Spark operator.
EMR Serverless is billed per vCPU-hour and GB-hour. AWS's worked example uses $0.052624 per vCPU-hour and $0.0057785 per GB-hour for a job with 100 vCPU and 750 GB for 30 minutes, then 200 more vCPU for 15 minutes, totaling $9.596275.
In the same AWS benchmark the cost was $4.81 for EMR 7.12, $5.47 for EMR 7.5 and $17.65 for open-source Spark 3.5.6 with Iceberg 1.10.0. Cost formula uses r5d.4xlarge at $1.152/hour plus $0.27/hour EMR plus EBS, On-Demand.
The EMR runtime for Spark provides 'full API compatibility with open source Apache Spark and Apache Iceberg', so existing Spark jobs run as is. Optimized runtimes are used by EMR on EC2, EMR Serverless, EMR on EKS, EMR on Outposts and AWS Glue. Self-run Spark on EKS uses the open-source Spark operator and needs no job rewrite.
Iceberg docs say users must run maintenance: expire snapshots, remove old metadata files and delete orphan files are 'recommended'; compacting data files with the Spark rewriteDataFiles action (for example target-file-size-bytes of 500 MB) and rewriting manifests are 'optional' maintenance, run as jobs you schedule.
Hudi merge-on-read tables need compaction to merge log files into base files and bound read latency. Compaction has two steps, scheduling (plan written to the timeline) and execution; the docs describe strategies for choosing which file slices to compact.
Delta Lake OSS docs list file management features the user applies: compaction (bin-packing) with OPTIMIZE, auto compaction, data skipping, Z-ordering, multi-part checkpointing, log compactions and optimized write.
Amazon EMR 7.14.0 packages Spark 3.5.8-amzn-0, Delta 3.3.2-amzn-3, Iceberg 1.10.1-amzn-1 and Hudi 1.0.2-amzn-3. All three table formats are available on EMR.
EMR on EKS supports the Kubernetes Operator for Apache Spark (Spark operator) as a job submission model with the EMR runtime on your own EKS clusters. The Kubeflow Spark Operator is Apache-2.0 licensed with Helm installation.
Amazon EMR has regional endpoints in the US (us-east-1, us-east-2, us-west-1, us-west-2) and many other AWS regions per the AWS general reference endpoints table. Spark runs on EC2 instances the user chooses (the AWS benchmark used r5d.4xlarge).
AWS publishes the setup in GitHub repositories (aws-samples/emr-on-eks-benchmark branch tpcds-v2.13_iceberg, and aws-samples/emr-spark-benchmark) with build instructions, so the TPC-DS Iceberg benchmark can be repeated.
EC2 instance-hours only. There is no managed-service charge; the team runs open-source Spark itself.
10 nodes x 2 hours = 20 node-hours. 20 x $1.008 = $20.16.
$20.16
Includes EC2 on-demand compute for the ten workers. Excludes the driver node, EBS volumes, S3 requests, data transfer, and the engineering time to operate Spark, Iceberg and the cluster manager. If the same open stack ran on a self-managed EKS cluster, AWS would add $0.10 per hour for each EKS cluster (Amazon EKS pricing page, read 2026-11-10). That fee is excluded here, as on the EMR on EKS line, because it depends on how long the cluster stays up, not on the two-hour job.
Searched for and none found on 10 Nov 2026
Independent third-party cost comparison (by a named party that is not AWS, Onehouse or another vendor) of open-stack Spark on EMR or EKS against managed lakehouses (searched: news.google.com/rss/search?q=EMR+7.12+Spark+performance+Iceberg+AWS+big+, awslabs.github.io/data-on-eks/docs/blueprints/data-analytics/spark-opera, (404))
Published benchmark of Iceberg, Delta or Hudi compaction time or cost on EMR/EKS (searched: iceberg.apache.org/docs/latest/maintenance, hudi.apache.org/docs/compaction, docs.delta.io/latest/optimizations-oss.html)
Open stack on EMR and EKS side by side with each other entry