Skip to content
190 changes: 190 additions & 0 deletions hadoop-hdds/docs/content/design/ozone-sts.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,190 @@
---
title: AWS STS Design for Ozone S3
summary: STS Support in Ozone
date: 2026-10-30
jira: HDDS-13323
status: implementing
author: Ren Koike, Fabian Morgan
---
<!--
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. See accompanying LICENSE file.
-->

# AWS STS Design for Ozone S3

# 1. Introduction

S3 credentials used to communicate with Ozone S3 APIs are based on a Kerberos identity.

Historically, the Ozone community has had interest in a REST API capable of programmatically generating
temporary S3 credentials.

Amazon AWS has the [Security Token Service (STS)](https://docs.aws.amazon.com/STS/latest/APIReference/welcome.html) which
provides the ability to generate short-lived access to resources.

The primary scope of this document is to detail the initial implementation of STS within the Ozone ecosystem.

# 2. Why Use STS Tokens?

Providing short-lived access to various resources in Ozone is useful in scenarios such as Data Lake
solutions that want to aggregate data across multiple cloud providers.

# 3. How Ozone STS Works

The initial implementation of Ozone STS supports only the [AssumeRole](https://docs.aws.amazon.com/STS/latest/APIReference/API_AssumeRole.html)
Copy link
Contributor

@ChenSammi ChenSammi Nov 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks @fmorg-git for preparing the doc.

The AssumeRole has a few input parameters, some are optional. Are they all supported? If not, could you clarify what parameters are supported, and what are not.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated - I added the supported required and optional parameters

API from the AWS specification. A new STS endpoint `/sts` on port `9880` will be created to service STS requests in the S3 Gateway.

Furthermore, the initial implementation of Ozone STS focuses only on Apache Ranger for the first phase, as it aligns more
with IAM policies. Support for the Ozone Native Authorizer may be provided in a future phase. Consideration for the Ozone
Native Authorizer will be given when processing IAM policies as described below.

## 3.1 Capabilities

The Ozone STS implementation has the following capabilities:

- Create temporary credentials that last from a minimum of 15 minutes to a maximum of 12 hours. The
return value of the AssumeRole call will be temporary credentials consisting of 3 components:
- accessKeyId - a generated String identifier beginning with the sequence "ASIA"
- secretAccessKey - a generated String password
- sessionToken - an opaque String identifier
- The temporary credentials will have the permissions associated with a role. Furthermore, an
[AWS IAM Session Policy](https://docs.aws.amazon.com/IAM/latest/UserGuide/access_policies.html#policies_session) can
**optionally** be sent in the AssumeRole API call to limit the scope of the permissions further. If
an IAM policy is specified, the temporary credential will have the permissions comprising the intersection of the role permissions
and the IAM policy permissions. **Note:** If the IAM policy is specified and does not grant any permissions, then
the generated temporary credentials won't have any permissions and will essentially be useless.

## 3.2 Limitations in IAM Session Policy Support

The AWS IAM policy specification is vast and wide-ranging. The initial Ozone STS supports a limited
subset of its capabilities. The restrictions are outlined below:

- The only supported prefix in ResourceArn is `arn:aws:s3:::` - all others will be rejected. **Note**: a ResourceArn
Copy link
Contributor

@ChenSammi ChenSammi Nov 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There are both reject and silently ignored. Shall we unify the behavior to reject if operations and fields are not supported?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This design behavior came after (much) discussion with external teams. One external team will send S3 actions that Ozone doesn't support, and they don't have flexibility to change what they are sending, so that's one reason for the silent ignore. This external team also mentioned that some research indicates AWS also does not fail the AssumeRole request just because the inline session policy references unknown or unsupported actions, but rather it will fail when the temporary credentials are used, so this design is accordance with it. Another external team agreed that behavior is fine for actions, but does not work for Conditions, because one can have a Condition to restrict calls by sourceIp, and if we silently ignore this, the client may incorrectly think the temporary credentials are restricted for use by that IP address, so consensus was to reject the request for that scenario.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see. Can you update the doc about the background reason why some are rejected and some are silently ignored?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated

of `*` is supported as well.
- The only supported Condition operator is `StringEquals` - all others will be rejected.
- The only supported Condition attribute is `s3:prefix` - all others will be rejected.
- Only one Condition operator per Statement is supported - a Statement with more than one Condition will be rejected.
- The only supported Effect is `Allow` - all others will be rejected.
- If a (currently) unsupported S3 action is requested, such as `s3:GetAccelerateConfiguration`, it will be silently ignored.
Similarly, an invalid S3 action will be silently ignored.
Comment on lines +88 to +89
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Apart from being mentioned in the design doc, for a user, how can we make sure that this "silently ignored" is a known behaviour? Would user docs be the only way? (As this is a design doc, a simple user of ozone may not go through this)

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, I believe this behavior will need to be included in user docs. Also, if IIRC from previous discussions with an external team, the silently ignored behavior is what happens in AWS as well, at least in the AssumeRole call. When the token is used, it should fail though. Please see #9223 (comment)

- Supported wildcard expansions in Actions are: `s3:*`, `s3:Get*`, `s3:Put*`, `s3:List*`,
`s3:Create*`, and `s3:Delete*`.
- If using OzoneNativeAuthorizer, bucket wildcards (ex. ResourceArn `arn:aws:s3:::*`, `arn:aws:s3:::bucket*` or `*`) will be rejected,

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Given STS is only available with Ranger authorizer (see #3 above), I suggest removing lines 93 to 97, to avoid potential misinterpretation that STS is supported with Ozone native authorizer as well.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated

as wildcards are not supported for buckets for the OzoneNativeAuthorizer.
However, certain object wildcards be accepted. For example, ResourceArn `arn:aws:s3:::myBucket/*` and `arn:aws:s3:::myBucket/folder/logs/*`
will be accepted but `arn:aws:s3:::myBucket/file*.txt`
will not be accepted. Again, these restrictions correspond to the capabilities of the OzoneNativeAuthorizer.

A sample IAM policy that allows read access to all objects in the `example-bucket` bucket is shown below:
```JSON
{
"Version": "2012-10-17",
"Statement": [
{
"Effect": "Allow",
"Action": "s3:GetObject",
"Resource": "arn:aws:s3:::example-bucket/*"
}
]
}

```

## 3.3 SessionToken Format

As mentioned above, one of the return values from the AssumeRole call will be the sessionToken. To support not
storing temporary credentials server-side in Ozone, the sessionToken will comprise various components needed to validate
subsequent S3 calls that use the token. The sessionToken will have the following information encoded:

- The originalAccessKeyId - this is the Kerberos identity of the user that created the sessionToken via the AssumeRole call.
Copy link
Contributor

@ChenSammi ChenSammi Nov 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why originalAccessKeyId is needed in sessionToken, In which step it will be used?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated to add verbiage about why originalAccessKeyId is needed. I also updated the flow section to indicate where originalAccessKeyId is used.

This identity is included in the sessionToken because S3 API calls (such as PutObject) require a Kerberos identity, but the temporary credentials don't have a Kerberos identity associated to them, therefore the Kerberos identity of the user that created the token will be used in these cases.

When the temporary credentials are used to make S3 API calls, this Kerberos identity (in conjunction with the role permissions and
optional session policy) will be used to authorize the call.
- The roleArn - the role used in the original AssumeRole call
- The encrypted secretAccessKey - this will be used to validate the AWS signature when the temporary credentials are used
to make S3 API calls
- (Optional) sessionPolicy - when using the RangerOzoneAuthorizer, if Ranger successfully authorizes the AssumeRole call,
Copy link
Contributor

@ChenSammi ChenSammi Nov 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

When both roleArn and sessionPolicy exist, the final decision made by Ranger is the intersection of roleArn and sessionPolicy?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, that's correct - I added additional verbiage to hopefully make that more clear

it will return a String representing the resources (i.e. buckets, keys, etc.) and permissions (i.e. ACLType) that the token
has been granted access to. This sessionPolicy will be included in the sessionToken sent back to the client so it can be sent to Ranger to
authorize subsequent S3 API calls that use the sessionToken.
- HMAC-SHA256 signature - used to ensure the sessionToken was created by Ozone and was not altered since it was created.
- The expiration time of the token (via `ShortLivedTokenIdentifier#getExpiry()`)
- The UUID of the secret key used to sign the sessionToken and encrypt the secretAccessKey (via `ShortLivedTokenIdentifier#getSecretKeyId()`)

## 3.5 STS Token Revocation

In the rare event temporary credentials need to be revoked (ex. for security reasons), a table in RocksDB will be created
Copy link
Contributor

@ChenSammi ChenSammi Nov 10, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Will OM be involved in the assumeRole API call and following S3 API calls using the sessionToken authorization check? Is it OM RocksDB here used, or other RocksDB?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes, OM is involved in the assumeRole API and S3 API calls (OM is currently involved in S3 calls with permanent credentials). Yes, OM RocksDB is used - I added some verbiage around that.

to store revoked tokens, and a command-line utility will be created to add tokens to the table. A background cleaner service
will be created to run every 3 hours to delete revoked tokens that have been in the table for more than 12 hours. The
input parameter for the command-line utility will be the accessKeyId of the temporary token - this value is returned in
plain text as a result of the AssumeRole call (mentioned above). In this way, specific STS tokens can be revoked as opposed
to all tokens. Furthermore, AWS doesn't have a standard API to revoke tokens therefore we are creating our own system.

## 3.6 Prerequisites

A user must be configured with a Kerberos identity in Ozone and the S3 `getSecret` command
must be called to issue permanent S3 credentials. With these credentials, the AssumeRole API call can be made, but additional
steps below are needed for the call to be successfully authorized.
When using RangerOzoneAuthorizer, a role must be configured in Ranger UI for each role the AssumeRole API
can be used with. The user making the AssumeRole call must be in this role, and this role must have access

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The user making the AssumeRole call must be in this role

This is not accurate.

I suggest replacing the sentence from 166 to 168 with the following:

Further, Apache Ranger policies should be in place to grant the user permission to assume the role.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

updated

to the (new) `ASSUME_ROLE` permission in Ranger, in addition to the volumes/buckets/keys and other permissions (such as `READ`,
`LIST`, etc.) that are required.

### 3.6.1 Additions to RangerOzoneAuthorizer

The `IAccessAuthorizer` interface that both the RangerOzoneAuthorizer and OzoneNativeAuthorizer implement, will have a
new method:

```java
default String generateAssumeRoleSessionPolicy(AssumeRoleRequest assumeRoleRequest) throws OMException {
throw new OMException("The generateAssumeRoleSessionPolicy call is not supported", NOT_SUPPORTED_OPERATION);
}
```

When using RangerOzoneAuthorizer, the AssumeRole API call must invoke this method to ensure the caller is authorized to create
temporary credentials, given the criteria in the AssumeRoleRequest. The AssumeRoleRequest input parameter will have the
components:
- `String` host - hostname of caller
- `InetAddress` ip - IP address of caller
- `UserGroupInformation` ugi - the user making the call
- `String` targetRoleName - what role is being assumed
- `Set<AssumeRoleRequest.OzoneGrant>` grants - further limiting the scope of the role according to the grants

The grants parameter is optional, and would only be present if the AssumeRole API call had an IAM session policy JSON
parameter supplied. A conversion utility, `IamSessionPolicyResolver` will process the IAM policy and convert it to a
`Set<AssumeRoleRequest.OzoneGrant>`, in effect translating from S3 nomenclature for resources and actions to Ozone nomenclature of
`IOzoneObj` and `ACLType`. Ranger would use all of this information to determine if the AssumeRole call should be
successfully authorized, and if so, it will return a String representation of the granted permissions and paths.

The format of this String is entirely up to the Ranger team. What is required from the Ozone side is to supply this String to Ranger when any
subsequent S3 API calls are made that use STS tokens. In order to achieve this, the sessionPolicy String from Ranger will
Comment on lines +191 to +192
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can we get a Ranger PMC member to sign off on this doc as well to indicate we will be working together on this integration?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I sent Slack messages and an email asking for Ranger input.

be included in the sessionToken response to the AssumeRole API call (as mentioned above), and Ozone will supply this String
to Ranger whenever STS tokens are used on S3 API calls via a new `RequestContext.sessionPolicy` field in the
`IAccessAuthorizer#checkAccess(IOzoneObj, RequestContext)` call.

## 3.7 Overall Flow

The following section outlines the overall flow when using STS in Ozone:

- An authorized user for AssumeRole API calls must be configured in Ozone, and if using RangerOzoneAuthorizer, the role
created in Ranger as per the Prerequisites above.
- This authorized user (having permanent S3 credentials) makes the AssumeRole STS call to Ozone.
- If successful, Ozone responds with the temporary credentials.
- A client makes S3 API calls with the temporary credentials for up to as long as the credentials last.
- When Ozone receives an S3 api call using temporary credentials, it will perform the following checks:
- Ensure the token is not expired
- Ensure the token is not revoked
- Validate the HMAC-SHA256 signature in the sessionToken
- Decrypt the secretAccessKey from the sessionToken and validate the AWS signature
- Authorize the call with either RangerOzoneAuthorizer or OzoneNativeAuthorizer

Assuming all these checks pass, the S3 API call will be invoked.