Amazon S3 Express One Zone Complete Guide — Single-Digit Millisecond Latency Storage for ML/HPC [2026]

Amazon S3 Express One Zone delivers up to 10x performance and single-digit millisecond latency compared to S3 Standard. It is a purpose-built, high-performance storage class designed for latency-sensitive workloads including ML training data loading, HPC jobs, and real-time analytics. The performance gains come from a single-AZ architecture and a dedicated request processing path.

In 2025 AWS dramatically reduced S3 Express One Zone pricing, addressing the primary adoption barrier:

Multiple organizations have published measured results after migrating to S3 Express One Zone: ClickHouse: Query performance improved by 283% after switching to S3 Express One Zone, with the largest gains in workloads involving concurrent access to many small files. Pinterest: Data processing pipeline speed increased approximately 10x after migrating storage to S3 Express One Zone, significantly reducing ML model training cycle times. These results demonstrate that ROI is highest for workloads where storage I/O latency is the primary bottleneck.

The March 2026 update expanded CloudWatch metrics for S3 Express One Zone significantly: - Per-minute request counts: Enables burst detection and capacity planning - Data transfer (IN/OUT): Granular network cost analysis - Error rates (4xx/5xx): Application failure detection - Latency percentiles (P50/P95/P99): SLA monitoring and anomaly detection This makes it possible to immediately detect performance degradation in ML pipelines using S3 Express One Zone and trigger automated alerts.

Not every workload benefits from S3 Express One Zone. Consider alternatives for: - Long-term archival: S3 Glacier Instant Retrieval is far more cost-effective for infrequently accessed data - Multi-AZ redundancy required: Single-AZ design means data is inaccessible during an AZ outage - Cost-first workloads: Higher price than S3 Standard when latency is not critical - Large cold datasets: S3 Intelligent-Tiering is better for large volumes with unpredictable access patterns

S3 Express One Zone uses a dedicated bucket type called a "directory bucket". Unlike general-purpose buckets, directory buckets feature a hierarchical namespace optimized for fast prefix operations. Key characteristics: - Bucket names follow the format `<name>--<az-id>--x-s3` (e.g., `my-ml-bucket--use1-az4--x-s3`) - Fully compatible with the standard S3 API; existing SDKs work unchanged - Access via `CreateSession` API for temporary session-based credentials - Optimized for parallel, high-frequency small-object operations

Step 1: Find Your AZ ID and Create the Bucket

# Find AZ IDs in your region
aws ec2 describe-availability-zones \
  --region us-east-1 \
  --query 'AvailabilityZones[*].[ZoneName,ZoneId]'

# Create directory bucket
aws s3api create-bucket \
  --bucket my-ml-data--use1-az4--x-s3 \
  --region us-east-1

Step 2: IAM Policy

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "s3express:CreateSession",
        "s3express:GetObject",
        "s3express:PutObject"
      ],
      "Resource": "arn:aws:s3express:us-east-1:123456789012:bucket/my-ml-data--use1-az4--x-s3"
    }
  ]
}

Step 3: Access with boto3

import boto3

s3 = boto3.client('s3', region_name='us-east-1')

# Upload training data
s3.put_object(
    Bucket='my-ml-data--use1-az4--x-s3',
    Key='datasets/train/batch_001.parquet',
    Body=open('batch_001.parquet', 'rb').read()
)

# High-speed read
response = s3.get_object(
    Bucket='my-ml-data--use1-az4--x-s3',
    Key='datasets/train/batch_001.parquet'
)
data = response['Body'].read()

The most cost-effective design separates data by access temperature: - Hot data (active training, last 7 days): S3 Express One Zone - Warm data (evaluation, reuse, last 30 days): S3 Standard - Cold data (archive, 30+ days): S3 Glacier Instant Retrieval Configure S3 Lifecycle Policies to automatically move data between tiers, eliminating manual management overhead while keeping costs optimized.

Use the March 2026 CloudWatch metrics to set up production-grade alerting:

import boto3

cw = boto3.client('cloudwatch', region_name='us-east-1')

# Alert when P99 latency exceeds 10ms
cw.put_metric_alarm(
    AlarmName='S3Express-HighLatency',
    MetricName='FirstByteLatency',
    Namespace='AWS/S3Express',
    Statistic='p99',
    Dimensions=[
        {'Name': 'BucketName', 'Value': 'my-ml-data--use1-az4--x-s3'},
        {'Name': 'FilterId', 'Value': 'EntireBucket'}
    ],
    Period=60,
    EvaluationPeriods=5,
    Threshold=10.0,
    ComparisonOperator='GreaterThanThreshold',
    AlarmActions=['arn:aws:sns:us-east-1:123456789012:ml-ops-alerts']
)

Recommended monitoring targets: - P99 FirstByteLatency: threshold 10ms - 5xx error rate: threshold 0.1% - Per-minute request count: spike detection

Q1. Which AWS regions support S3 Express One Zone? Major regions including US East (N. Virginia), US West (Oregon), and Europe (Ireland). Check the AWS documentation for the latest list. Q2. Do existing S3 SDKs and APIs work with directory buckets? Yes. S3 Express One Zone uses S3-compatible APIs. Existing boto3, AWS CLI, and SDK code works without modification; only the bucket name format differs. Q3. What happens during an AZ outage? Data becomes temporarily inaccessible. Always configure replication to S3 Standard for important datasets to maintain access during AZ-level failures. Q4. Is there a throttling risk for large ML training jobs? Yes. Request rate limits apply. Implement exponential backoff and manage `CreateSession` connections carefully for large parallel training jobs. Q5. Can S3 Express One Zone be combined with Intelligent-Tiering? Not directly. Use lifecycle policies to move data from S3 Express One Zone to S3 Standard or Intelligent-Tiering buckets as it ages. Q6. Do per-minute CloudWatch metrics incur additional charges? Yes, CloudWatch custom metrics are billed per metric. Limit monitoring to the most critical metrics to manage costs. Q7. Does Amazon SageMaker support S3 Express One Zone natively? Yes. SageMaker natively supports directory buckets for training datasets, significantly reducing job startup and per-epoch data loading times.

Optimizing ML training pipelines with S3 Express One Zone, designing hybrid storage architectures, or setting up CloudWatch monitoring? Oflight provides AWS infrastructure design and implementation services tailored to ML and HPC workloads. Visit our Network & Infrastructure Services page to get started.