Amazon EMR Serverless Platform

Amazon EMR Serverless is AWS’s serverless deployment option for running Apache Spark without managing clusters. BenchBox integrates with EMR Serverless to execute benchmarks with automatic scaling and sub-second startup times.

Features

• Serverless - No clusters to manage, automatic scaling
• Fast startup - Sub-second cold starts with pre-initialized capacity
• Cost-effective - Pay only for vCPU-hours and memory-GB-hours used
• Integrated - Native S3 and Glue Data Catalog integration
• Spark optimization - Uses cloud-spark shared infrastructure for config tuning

Installation

uv add benchbox --extra cloud-spark-aws

This installs the AWS client dependency, boto3, via the cloud-spark-aws extra.

Prerequisites

1. AWS Account with EMR Serverless access
2. S3 Bucket for data staging and job scripts
3. IAM Execution Role with the following permissions:
   • emr-serverless:StartApplication, emr-serverless:GetApplication
   • emr-serverless:StartJobRun, emr-serverless:GetJobRun
   • s3:GetObject, s3:PutObject, s3:ListBucket
   • glue:GetDatabase, glue:CreateDatabase, glue:GetTable, glue:CreateTable
4. AWS Credentials configured via CLI, environment, or IAM role

Configuration

CLI Usage

The BenchBox CLI does not accept the EMR Serverless staging, role, or application settings, so run benchmarks through the Python API described below. The CLI can still preview the generated queries without connecting to AWS:

benchbox run --platform emr-serverless --benchmark tpch --dry-run ./preview

Adapter Options

These are EMRServerlessAdapter keyword arguments. Set application_id to use an existing application, or set create_application=True to create one.

Option	Default	Description
s3_staging_dir	required	S3 path for data staging
execution_role_arn	required	IAM role ARN for job execution
application_id	-	Existing EMR Serverless application ID
create_application	false	Create new application if ID not provided
application_name	-	Name for a new application
region	us-east-1	AWS region
database	benchbox	Glue Data Catalog database
release_label	emr-7.0.0	EMR release label
timeout_minutes	60	Job timeout in minutes

Python API

from benchbox.platforms.aws import EMRServerlessAdapter

adapter = EMRServerlessAdapter(
    application_id="00f12345abc67890",
    s3_staging_dir="s3://my-bucket/benchbox/",
    execution_role_arn="arn:aws:iam::123456789012:role/EMRRole",
    region="us-east-1",
    database="tpch_benchmark",
)

adapter = EMRServerlessAdapter(
    s3_staging_dir="s3://my-bucket/benchbox/",
    execution_role_arn="arn:aws:iam::123456789012:role/EMRRole",
    create_application=True,
    application_name="benchbox-tpch",
)

adapter.create_schema("tpch_sf1")

adapter.load_data(
    tables=["lineitem", "orders", "customer"],
    source_dir="/path/to/tpch/data",
)

result = adapter.execute_query("SELECT COUNT(*) FROM lineitem")

adapter.close()

The first adapter uses an existing application, and the second creates a new one. create_schema creates the database in the Glue Data Catalog. load_data uploads data to S3 and creates Glue tables. Each query is submitted as a job run. close cleans up and optionally stops the application.

Execution Model

EMR Serverless executes benchmarks as job runs within an application:

1. Application - Container for job runs with auto-start/stop
2. Job Submission - Query script uploaded to S3 and submitted
3. Automatic Scaling - Resources provisioned based on workload
4. Result Retrieval - Results read from S3 output location
5. Resource Tracking - vCPU-hours and memory-GB-hours logged

Pre-Initialized Capacity

For sub-second startup, configure pre-initialized workers:

adapter = EMRServerlessAdapter(
    application_id="00f12345abc67890",
    s3_staging_dir="s3://bucket/benchbox/",
    execution_role_arn="arn:aws:iam::123456789012:role/EMRRole",
    initial_capacity={
        "Driver": {
            "workerCount": 1,
            "workerConfiguration": {
                "cpu": "4 vCPU",
                "memory": "16 GB"
            }
        },
        "Executor": {
            "workerCount": 4,
            "workerConfiguration": {
                "cpu": "4 vCPU",
                "memory": "16 GB"
            }
        }
    },
)

Note: Pre-initialized capacity is charged even when idle.

Cost Estimation

Resource	Price
vCPU-hour	~$0.052624
Memory GB-hour	~$0.0057785

Scale Factor	Data Size	Est. Runtime	Est. Cost
0.01	~10 MB	~20 min	~$0.05
1.0	~1 GB	~1 hour	~$0.50
10.0	~10 GB	~3 hours	~$3.00
100.0	~100 GB	~8 hours	~$16.00

Estimates based on auto-scaling. Actual costs vary.

IAM Role Policy

Minimum required IAM policy for the execution role:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "s3:GetObject",
        "s3:PutObject",
        "s3:DeleteObject",
        "s3:ListBucket"
      ],
      "Resource": [
        "arn:aws:s3:::your-bucket",
        "arn:aws:s3:::your-bucket/*"
      ]
    },
    {
      "Effect": "Allow",
      "Action": [
        "glue:GetDatabase",
        "glue:CreateDatabase",
        "glue:GetTable",
        "glue:CreateTable",
        "glue:UpdateTable"
      ],
      "Resource": "*"
    },
    {
      "Effect": "Allow",
      "Action": [
        "logs:CreateLogGroup",
        "logs:CreateLogStream",
        "logs:PutLogEvents"
      ],
      "Resource": "arn:aws:logs:*:*:*"
    }
  ]
}

Also add a trust relationship for EMR Serverless:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Principal": {
        "Service": "emr-serverless.amazonaws.com"
      },
      "Action": "sts:AssumeRole"
    }
  ]
}

Troubleshooting

Common Issues

Application fails to start:

• Verify IAM role has correct trust relationship
• Check region has EMR Serverless available
• Review application configuration in console

Job fails immediately:

• Check S3 bucket permissions
• Verify execution role has required permissions
• Review job logs in S3 or CloudWatch

Slow startup:

• Configure pre-initialized capacity for warm workers
• Check application auto-start is enabled
• Review network configuration (VPC settings)

Query timeout:

• Increase timeout_minutes for large scale factors
• Check for data skew in queries
• Review Spark configuration

Comparison with Other Platforms

Aspect	EMR Serverless	AWS Glue	EMR on EC2
Cluster management	None	None	Full control
Startup time	Sub-second*	30-60s	Minutes
Billing	vCPU + memory hours	DPU-hours	Instance hours
Use case	Interactive	ETL batch	Long-running
Scaling	Automatic	Automatic	Manual/Auto

*With pre-initialized capacity

Related

• AWS Glue Platform - Serverless ETL with Glue
• Apache Spark Platform - Local/cluster Spark usage
• Cloud Platforms Overview - All cloud platforms