Amazon S3 Express One Zone

Definition

Amazon S3 Express One Zone is a high-performance, single-Availability-Zone storage class designed for workloads that need consistent single-digit millisecond first-byte latency. It uses a new bucket type — directory buckets — and delivers up to 10× faster access than S3 Standard for small-object reads. Released GA in late 2023, with significant price cuts in 2025 making it competitive for many production workloads.

Key characteristics

• Single AZ — Higher performance at the cost of multi-AZ durability. Pair with S3 Standard or replication for backup.
• Directory buckets — Use a flat directory namespace, not S3’s bucket-prefix model. Names take the form bucket-name--azid--x-s3.
• Latency — Single-digit millisecond first-byte for both reads and writes.
• Throughput — 2× more requests per second per prefix than S3 Standard at GA, with further gains in 2025.
• Pricing — Higher per-GB storage than Standard, but lower request pricing (50% off PUT and 75% off GET vs S3 Standard at 2025 price cuts).

When to use S3 Express One Zone

• ML training and feature stores — Random small-object reads during training epochs
• Real-time analytics scratch space — Spark and Trino intermediate shuffles
• Latency-sensitive log processing — Tail and reprocess windows
• HPC scratch — Genomics, simulation, financial modeling intermediate state

When not to use S3 Express One Zone

• Long-term storage — Use S3 Standard, S3 Standard-IA, S3 Glacier classes
• Multi-region or cross-AZ failover — Single-AZ means an AZ outage takes the bucket offline
• Compliance regimes requiring multi-AZ durability — Pair with replication if needed

Common mistakes

Mistake 1: Treating directory buckets like standard buckets. Different naming, different IAM patterns, and CloudFront does not integrate the same way.

Mistake 2: Forgetting to set lifecycle for stale objects. Storage is more expensive per GB — keep only the working set in Express One Zone and tier cold data out.

Mistake 3: Picking Express One Zone for a workload that does not stress S3 latency. If you are not seeing first-byte latency in your training-throughput profile, Standard is cheaper.

Related AWS Services

• Amazon SageMaker — Native integration for training data
• Amazon EMR / EMR Serverless — Spark scratch space
• Amazon Athena — Query results from Express One Zone
• Amazon FSx for Lustre — Alternative for full POSIX HPC workloads

Related FactualMinds Content

• AWS SageMaker ML Consulting
• AWS Data Analytics Services