Amazon S3 Express One Zone

Definition

Amazon S3 Express One Zone is a high-performance storage class built on directory buckets — a bucket type colocated in a single Availability Zone that delivers single-digit millisecond first-byte latency for reads and writes. It targets workloads that hammer S3 with small, random I/O: ML training epochs, Spark shuffle intermediates, feature stores, and HPC scratch. Express One Zone trades multi-AZ durability for speed and lower request pricing versus S3 Standard on hot paths; directory bucket names follow the pattern bucket-name--azid--x-s3.

Storage per GB-month costs more than Standard, but PUT/GET request charges are substantially lower after 2025 price reductions — total cost wins when request rate and latency sensitivity dominate, not when data sits cold for months. Pair Express One Zone with Standard or replication for anything that must survive an AZ outage.

When to use it

• ML training pipelines doing repeated random reads of shard files or manifests during each epoch.
• Spark, Trino, or EMR intermediate shuffle and spill directories where latency caps cluster utilization.
• Real-time analytics landing zones with tight tail-latency SLOs on small objects.
• HPC or genomics scratch space where jobs are AZ-local and duration is hours to days, not years.

When not to use it

• Long-term archive, compliance retention, or infrequently accessed backups — use Standard-IA or Glacier classes.
• Workloads requiring multi-AZ failover without a separate replication story — an AZ outage takes the bucket offline.
• Throughput-bound sequential reads of large objects where Standard S3 already saturates network — Express One Zone optimizes latency, not necessarily $/TB for cold data.

Tips

• Create directory buckets in the same AZ as your compute (SageMaker training jobs, EKS node groups, EMR clusters) to avoid cross-AZ charges and latency.
• Lifecycle cold prefixes out to Standard or Intelligent-Tiering after job completion — Express per-GB pricing punishes forgotten scratch data.
• Use the S3 Express auth model (session-based credentials for directory buckets) — legacy SigV4-only tooling may need SDK updates.
• Benchmark first-byte latency and requests/sec before migrating; if p99 latency is not on your critical path, Standard is cheaper.
• For training sets accessed once per run, stage in Express from Standard with a manifest copy job rather than storing canonical datasets only in single-AZ storage.

Gotchas

• Serious: Treating Express One Zone as durable multi-AZ storage — AZ failure means bucket unavailability until AWS restores the zone; keep authoritative copies elsewhere.
• Serious: CloudFront and some third-party tools do not treat directory buckets like general-purpose buckets — validate CDN and backup integrations before architecture sign-off.
• Regular: IAM policies for directory buckets differ from classic bucket ARNs — copy-paste Standard S3 policies fail with opaque AccessDenied.
• Regular: List operations on huge flat key spaces still cost time and money — partition keys thoughtfully even with directory semantics.
• Regular: Cross-AZ reads from compute in AZ-a to Express bucket in AZ-b add latency and data transfer — pin resources deliberately.

Official references

• S3 Express One Zone performance — latency and throughput expectations.
• Mountpoint for S3 — POSIX access patterns for S3 data sets.

Related FactualMinds content

• Amazon S3 — storage classes and security baseline
• AWS SageMaker ML Consulting
• AWS Data Analytics Services