Skip to content
Lakehouse Compare

Guide, read 10 Nov 2026

EMR cost optimization when the tables are Iceberg

All guidesOpen the Compute Bill

What the EMR bill is made of

AWS states that on EMR on EC2 the EMR price is added to the EC2 price and to the EBS price if volumes are attached, all billed per second with a one-minute minimum. The EMR uplift is a separate line per instance type. On the AWS price list for us-east-1 (publication 2026-11-10), it is $0.192 per hour for m5.4xlarge, $0.1632 for m7g.4xlarge, $0.2142 for r7g.4xlarge and $0.27 for r5d.4xlarge.

EMR on EKS replaces the per-instance uplift with a charge per vCPU-hour ($0.007084) and per GB-hour ($0.000777875) added to EC2 or Fargate, plus $0.10 per hour per EKS cluster. EMR Serverless bills vCPU-hours and GB-hours of workers; AWS's worked example uses $0.052624 per vCPU-hour and $0.0057785 per GB-hour. The compute bill page carries these in a priced workload.

Instance choice

Source: AWS price lists for Amazon EC2 and Amazon EMR, us-east-1, publication 2026-11-10. Prices only; no performance is implied.
Instance (us-east-1)vCPU / memoryEC2 On-Demand per hourEMR uplift per hour
m5.4xlarge16 / 64 GiB$0.768$0.192
m6g.4xlarge16 / 64 GiB$0.616$0.154
m7g.4xlarge16 / 64 GiB$0.6528$0.1632
r5d.4xlarge16 / 128 GiB$1.152$0.27
r6g.4xlarge16 / 128 GiB$0.8064$0.2016
r7g.4xlarge16 / 128 GiB$0.8568$0.2142

Spot and Graviton

AWS says Spot Instances are available at a discount of up to 90% off On-Demand, and the Spot price moves, so the Spot Instance Advisor is the place to read a current figure. The EMR docs advise caution on which nodes go on Spot: you cannot change a running cluster's purchasing option, and by default EMR keeps YARN application masters off task nodes so that losing Spot task nodes does not fail running jobs. Core nodes on Spot risk partial HDFS data loss.

AWS states Graviton-based instances cost up to 20% less than comparable x86 instances. The price list above shows the EC2 gap for the same 16 vCPU and 64 GiB shape: $0.6528 for m7g.4xlarge against $0.768 for m5.4xlarge. Whether your Spark jobs and libraries run the same on arm64 is a test you have to run; no AWS benchmark for Iceberg compaction on Graviton was found on 2026-11-10.

Managed scaling and instance fleets

EMR managed scaling (EMR 5.30.0 and later, except 6.0.0) raises and lowers core and task capacity from workload metrics. You set a minimum, a maximum, an optional On-Demand limit that sends the rest to Spot, and an optional maximum for core nodes. It works with YARN applications such as Spark. AWS advises keeping Spark dynamic resource allocation on, because turning it off can scale a cluster up to its maximum, and keeping EBS use under 90%.

Instance fleets let one fleet name up to five EC2 instance types and an allocation strategy such as capacity-optimized, lowest-price or diversified. That widens the Spot pools a fleet can draw from.

Compaction is a cost line too

Compaction on EMR is a Spark job you schedule, so it consumes the same EC2 plus uplift hours. Apache Iceberg's defaults give a file group a rewrite when it holds 5 or more files, and a rewritten group is capped at 100 GB. The AWS-managed alternatives bill by their own meters: Glue Data Catalog Iceberg compaction costs $0.44 per DPU-hour in us-east-1, and S3 Tables compaction costs $0.005 per GB processed plus $0.002 per 1,000 objects compacted, or $0.01 per GB for sort and z-order.

Which of those is lower for a given table depends on file counts, sort strategy and cluster shape. No published benchmark of Iceberg compaction cost on EMR against these services was found on 2026-11-10. The maintenance pages compare what each platform automates.

Plugins and hardware options on EMR

AWS publishes its own runtime figures: on TPC-DS at 3 TB with 9 r5d.4xlarge nodes, AWS reports EMR 7.12 running in 1,349.62 seconds on average against 6,113.92 for open-source Spark 3.5.6 with Iceberg 1.10.0, and a run cost of $4.81 against $17.65. AWS ran it, notes it is not comparable to official TPC-DS, and publishes the setup in GitHub repositories.

DualBird says its Spark plugin runs on AWS F2 FPGA instances and reports a 100 GB sort-based Iceberg compaction test in a June 11, 2026 blog post, with about $2 per TB against about $6 to over $13 per TB on CPU clusters. DualBird ran that test and has not published its method outside a sign-up form. DualBird is not a lakehouse. It speeds up Spark on EMR and EKS; a team on a fully managed platform cannot add it. It documents no Spark version list and no table-format support list; Iceberg appears only in its own compaction benchmark.

Sources

  1. Amazon EMR pricing, read 10 Nov 2026
  2. AWS price list, Amazon EMR, us-east-1, read 10 Nov 2026
  3. AWS price list, Amazon EC2, us-east-1, read 10 Nov 2026
  4. Amazon EC2 Spot pricing, read 10 Nov 2026
  5. AWS Graviton, read 10 Nov 2026
  6. Amazon EMR managed scaling, read 10 Nov 2026
  7. Amazon EMR instance types and Spot guidance, read 10 Nov 2026
  8. Amazon EMR instance fleets, read 10 Nov 2026
  9. Apache Iceberg Spark procedures, read 10 Nov 2026
  10. AWS price list, AWS Glue, us-east-1, read 10 Nov 2026
  11. AWS price list, Amazon S3, us-east-1, read 10 Nov 2026
  12. AWS Big Data Blog, EMR 7.12 and Iceberg benchmark, read 10 Nov 2026
  13. DualBird, Iceberg compaction benchmark blog, 2026-06-11, read 10 Nov 2026

Last reviewed 10 Nov 2026.