Home › AWS › AWS data services › S3 Express One Zone
AWS data services · Storage and table formats · about 11 min read
Amazon S3 Express One Zone
The fastest S3 storage class: single-digit millisecond access from directory buckets kept in one Availability Zone you choose.
In plain words
What it is
Ordinary S3 is very durable and very cheap, but each request takes tens of milliseconds, because the data is spread across at least three Availability Zones (separate data centres in a Region). For most data engineering that is fine: a Spark job reads big files and the wait for the first byte is lost in the noise. For some jobs it is not fine. A job that opens millions of small files, or a model training run that reads the same images again and again, spends much of its time waiting.
S3 Express One Zone is a storage class built for those jobs. You pick one Availability Zone, ideally the same one as your compute, and create a directory bucket there. AWS says it gives consistent single-digit millisecond access, up to ten times faster than S3 Standard, with lower request prices. In return the data lives in only one zone. It is copied across several devices inside that zone and designed for 99.95 percent availability, but if the zone is lost, so is access to the data.
A concrete example: a quality team at a tyre plant trains an image model on two million photos of tread defects, each around 200 KB. Training reads the full set many times. Copying the photos into a directory bucket in the same zone as the GPU instances shortens each pass, because the GPUs stop waiting for files. The master copy stays in a normal S3 bucket; the Express copy is a fast working copy that can be rebuilt any time.
Directory buckets work differently from the general purpose buckets most people know. Their names include the zone, they have a real folder hierarchy, and access is checked through short sessions created with the CreateSession call. The AWS SDKs handle those sessions for you.
A workbench next to the machine: much quicker to reach than the warehouse, but it holds only what you are working on now, and a fire in that one room takes it all.
Big picture
Where it fits
The lit lane is where S3 Express One Zone works. Boxes on the left feed it; boxes on the right use what it produces.
On the site's reference platform, all tables (landing, bronze, silver, gold) live in general purpose S3 buckets, one per catalog, spread across zones, versioned, replicated to a second Region for recovery. S3 Express One Zone is not in that path and should not be. It has no versioning, no replication and no Object Lock, and Unity Catalog managed tables need the guarantees of normal S3.
Where it can fit is as a side room next to the processing lane. Copies of training data for ML, temporary shuffle or spill files for a self-managed Spark or EMR cluster, or a hot set of small lookup files read many thousands of times a minute. In each case the data is a copy or is short-lived, the compute runs in the same zone, and someone has measured that waiting on S3 is the bottleneck.
For most Databricks workloads, the Databricks disk cache on the cluster's local disks already gives fast repeated reads without a second storage system. Try that first. Databricks support for directory buckets as external locations or volumes may be limited, so test and check the current docs before you plan around it.
How it works
The ideas you need
| Idea | What it means |
|---|---|
| Availability Zone | One or more separate data centres inside an AWS Region. Express One Zone keeps data in one zone that you choose. |
| Directory bucket | The bucket type that holds Express One Zone data. It has a hierarchical folder structure and a name that includes the zone id. |
| Storage class | The S3 setting that decides where and how an object is stored and priced. EXPRESS_ONEZONE is the class used here. |
| CreateSession | An API call that returns temporary credentials scoped to one directory bucket. They expire after 5 minutes; the SDK renews them for you. |
| Zonal and Regional endpoints | Object reads and writes go to a zonal endpoint near the data; creating and deleting buckets goes to the Regional endpoint. |
| Gateway VPC endpoint | A free route from your VPC to S3 that avoids the internet and NAT. Directory buckets support it. |
| Co-location | Running compute in the same zone as the directory bucket, so reads do not cross zones and latency stays low. |
| Mountpoint for Amazon S3 | An open-source client that makes a bucket look like a local folder; it works with directory buckets and suits training jobs that read files. |
Choosing
When to use it, and when not
Use it when
- A training job reads the same set of small files many times and you have measured that GPUs or CPUs sit idle waiting for S3.
- A self-managed Spark, EMR or Ray cluster needs fast scratch space for intermediate files that can be recreated if lost.
- An interactive application must fetch many small objects with consistent low latency, and the compute already runs in one zone.
- Request costs dominate a workload of very many small reads and writes, and the data does not need to survive a zone failure.
Avoid it when
- The data is a table on the medallion path. Bronze, silver and gold need multi-zone durability, versioning and replication that directory buckets do not offer.
- The data is the only copy of anything you cannot rebuild, such as landing files or audit logs.
- You need Object Lock, cross-Region replication, versioning or S3 server access logs; directory buckets do not support them.
- The workload reads large files in big sequential scans; ordinary S3 throughput is already good and the extra storage price buys nothing.
- Compute runs across several zones or is serverless in a zone you cannot choose, so cross-zone reads remove most of the speed gain.
Production grade
What goes wrong in production
| Problem | What breaks | What to do |
|---|---|---|
| Zone outage takes the data offline | Express One Zone stores data in one zone. If that zone has a problem, the bucket is unavailable and jobs that depend on it stop. | Keep the master copy in a general purpose bucket and treat Express data as a rebuildable cache; script the rebuild. |
| Compute in the wrong zone | A cluster starts in a different zone from the directory bucket. Every read crosses zones, latency rises and inter-zone data transfer is billed. | Pin the subnets for that workload to the bucket's zone and check the zone id, not the zone name, which differs between accounts. |
| Higher storage price forgotten | Storage per GB costs well above S3 Standard. A scratch bucket nobody empties becomes an expensive archive. | Use lifecycle expiration rules on directory buckets and a job that deletes scratch data at the end of each run. |
| Tools that assume general purpose buckets | Older clients, some connectors and some platform features expect versioning, ACLs or list behaviour that directory buckets do not provide, and fail in odd ways. | Test each engine and connector against a directory bucket first, including Databricks external locations, before committing. |
| Permissions model is different | Teams copy prefix-based bucket policies from general purpose buckets. Directory buckets set the access boundary at the bucket level only, and object access goes through s3express:CreateSession. | Grant CreateSession per bucket, split data that needs different access into separate directory buckets, and review policies with IAM Access Analyzer. |
| Weaker audit trail | S3 server access logs are not supported, so a team that relies on them for object-level audit loses it. | Turn on CloudTrail data events for the directory bucket where object-level audit matters. |
Alternatives
What to compare it with
| Alternative | Pick the alternative when | Pick S3 Express One Zone when |
|---|---|---|
| S3 Standard in a general purpose bucket | Almost every data platform table and file. Multi-zone durability, versioning, replication and Object Lock. | You have measured that request latency on many small files is your bottleneck. |
| Databricks disk cache | Databricks clusters re-read the same Parquet or Delta files; the cache keeps them on fast local disk with no extra storage system. | Data must be shared across many clusters or non-Databricks tools at low latency. |
| Amazon FSx for Lustre | A shared POSIX file system with very high throughput for HPC or ML that needs file locking and real file semantics. | Object access through the S3 API is enough and you want no file system to run. |
| Azure Premium Block Blob storage | You are on Azure and need low-latency object reads. | Your platform is on AWS. |
| Google Cloud Storage zonal buckets (Rapid) | You are on Google Cloud and want zonal, low-latency object storage next to compute. | Your platform is on AWS. |
Cost
How you pay
You pay a storage price per GB-month that is clearly higher than S3 Standard, and request prices that are lower than S3 Standard. There is also a per-GB charge for the part of each upload or download above a size threshold, so very large objects lose some of the request saving. There is no minimum storage duration and no minimum object size. Gateway VPC endpoints are free. Cross-zone data transfer is billed if compute sits in another zone. Check the S3 pricing page for your Region; AWS has cut Express prices before.
What drives the bill
- How many GB stay in the bucket, and for how long.
- Number of requests, which is where Express saves money on small-object workloads.
- Size of each request, because bytes above the threshold are billed per GB.
- Whether compute runs in the same zone; if not, inter-zone transfer adds a charge.
Cost traps
- Leaving training copies or scratch data in the bucket after the job finishes.
- Using Express for large sequential reads, where you pay the higher storage price for no gain.
- Running compute in another zone and paying inter-zone transfer on every read.
- Keeping a second full copy of a large dataset when only a hot subset needs to be fast.
Prices change; check the AWS pricing page for your region before deciding.
Security and governance
Keeping it safe
- Directory buckets are private by default; Block Public Access is on and cannot be changed, and ACLs are disabled.
- Grant object access through s3express:CreateSession on a specific bucket; bucket management actions cannot be cross-account. The access boundary is the bucket, not a prefix or tag.
- Data is encrypted with SSE-S3 by default; SSE-KMS with your own key is supported. SSE-C and dual-layer DSSE-KMS are not.
- Reach the bucket from your VPC through a gateway VPC endpoint so traffic never touches the internet or a NAT gateway, matching the platform's no-public-IP rule.
- CloudTrail records bucket management as management events and object reads and writes as data events; server access logs are not available.
- Do not put untokenised personal data here: there is no versioning to support erasure checks and no Macie scan in the site's design for these buckets.
Operating it
What to watch, and the limits
| Watch | Why |
|---|---|
| BucketSizeBytes | Daily CloudWatch storage metric in the AWS/S3 namespace; rising size on a scratch bucket means clean-up is failing. |
| NumberOfObjects | Daily object count; a sudden jump shows a job writing far more small files than planned. |
| Job read time or GPU utilisation | The reason the bucket exists. If training or job time does not drop, the extra cost buys nothing. |
| CloudTrail data events for CreateSession, GetObject and PutObject | Who reads and writes the bucket, for audit and to spot unexpected clients. |
| Inter-zone data transfer in the Cost and Usage Report | Shows compute reading from the wrong zone. |
| Limit | Value and notes |
|---|---|
| Location | One Availability Zone per directory bucket, chosen at creation; no cross-zone redundancy |
| Availability | Designed for 99.95 percent within the zone, backed by the S3 Service Level Agreement |
| Session credentials | Temporary credentials from CreateSession expire after 5 minutes |
| Unsupported features | No S3 Versioning, Object Lock, replication or server access logs; no SSE-C; no MD5 checksums |
| Access control boundary | Bucket level only; no prefix or object-tag based permissions |
Limits and quotas checked 2026-11-11; many can be raised through Service Quotas.
Decisions
Situations and verdicts
| Situation | Choice | Verdict |
|---|---|---|
| Where to keep silver and gold tables for plant sensor readings | General purpose S3 buckets as Unity Catalog managed tables | Drop Express here: medallion tables need multi-zone durability, versioning and replication. |
| Tread-defect image model retrained weekly on two million small photos, GPUs idle 40 percent of the time waiting for reads | Copy the training set into a directory bucket in the GPU zone, read through Mountpoint | Pursue with care: measure before and after, keep the master copy in S3 Standard, expire the copy after training. |
| Databricks job re-reads the same gold files many times a day | Databricks disk cache on the cluster | Pursue: try the cache first; it needs no new storage system and no zone pinning. |
| Shared scratch space for a self-managed EMR cluster in one zone | Directory bucket with a lifecycle expiration rule | Pursue with care: only for data that can be recreated, with clean-up enforced. |
FAQ
Questions people ask
Is S3 Express One Zone faster than S3 Standard?
Yes, for request latency. AWS states consistent single-digit millisecond access and up to ten times faster access than S3 Standard. The gain is largest for many small reads from compute in the same zone; large sequential reads gain little.
Is S3 Express One Zone durable?
Data is stored on several devices in one Availability Zone, and lost redundancy is repaired quickly. It is not resilient to the loss of that zone, so keep any data you cannot rebuild somewhere else.
Can I put Delta or Iceberg tables in S3 Express One Zone?
Technically some engines can write there, but it is a poor fit for a governed lakehouse: no versioning, no replication, no Object Lock and one zone. Keep tables in general purpose buckets.
Does Databricks support directory buckets?
Support for directory buckets as Unity Catalog external locations or volumes may be limited and changes over time. Check the current Databricks docs and test before you rely on it. The Databricks disk cache is often the simpler answer.
How do I move data into a directory bucket?
Use the S3 console Import feature, which creates a Batch Operations copy job from a general purpose bucket, or copy with the CLI, SDK or a Spark job.
Sources
Checked against
- AWS docs: S3 Express One Zone
- AWS docs: Amazon S3 storage classes
- AWS docs: Differences for directory buckets
- AWS: Amazon S3 pricing
Checked 2026-11-11.
Related
Where it connects
Amazon S3 · S3 storage classes and lifecycle · Amazon SageMaker AI · Data platform on AWS and Databricks · Cost calculator · All AWS data services