Status

Current state: "Accepted"

Discussion thread: here

JIRA: here

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Kafka authorizes access to resources like topics, consumer groups etc. by way of ACLs. The current supported semantic of resource name in ACL definition is either full resource name or special wildcard '*', which matches everything.

Kafka should support a way of defining bulk ACLs instead of specifying individual ACLs.
Example use cases:

• Principal “user2” has access to all topics that start with “com.company.product1.”.
• Principal “user1” has access to all consumer groups that start with “com.company.client1.”.

This support will greatly simplify ACL operational story in a multi-tenant environment.

Public Interfaces

Summary of changes

• Add new field 'ResourceNameType' to Resource, and ResourceFilter classes to distinguish between literal and prefixed resource names.
  • ResourceNameType is an enum to support more types in the future.
  • ResourceNameType will include an 'ANY' type for use in filters
  • ResourceNameType will default to literal, to maintain backwards compatibility with existing clients.

• The SimpleAclAuthorizer will be enhanced to support prefixed resource names.
  
  • The contract of existing CRUD operations on the SimpleAclAuthorizer will remain the same: they will only return ACLs for the resources that exact matches the ones passed in.
  • The CRUD operations will accept literal and prefixed resources.
  • The authorize(...) method, which calls getAcls(resource) and getAcls('*') to get all the matching ACLs today, will instead now look for all matching literal and prefixed ACLs.

• Changes to command line tool class https://github.com/apache/kafka/blob/trunk/core/src/main/scala/kafka/admin/AclCommand.scala
  • Expose a '--resource-name-type' flag, which can be set to literal, prefixed or any, e.g.
          bin/kafka-acls.sh --authorizer-properties zookeeper.connect=localhost:2181 --add --allow-principal User:Bob  --operation Read --group my-app-* --resource-name-type prefixed
  • The flag will default to 'literal', meaning the commands will return only the ACLs they previously returned, even if prefixed ACLs exists for the resource.
  • Users can now set the flag to 'all' to retrieve/delete all ACLs affecting the supplied ACLs resource.
    
    
• Changes to the admin client to support prefixed ACLs.
  • New schema version for CreateAclsRequest / DeleteAclsRequest / DescribeAclsRequest, and associated responses, which will have a new byte field in schemas to distinguish literals vs prefixed resource names.
  • The CreateAclRequest will only accept prefixed or literal ResourceNameTypes
  • The DeleteAclRequest and DescribeAclRequests will accept prefixed or literal, and will return delete/describe only ACLs exactly matching the resources. i.e. no pattern matching will be performed
  • The DeleteAclRequest and DescribeAclRequests will also accept a ResourceNameType of 'all', which will delete/describe all ACLs affecting the resource, including the literal wildcard ACL '*', if present.

AdminClient changes

The examples below should help demonstrate the purposed functionality:

Exissting functionality

Each call retrieves ACLs stored in one path in ZK:
adminClient.describeAcls(... TOPIC, "foobar" ...) -> would return only ACLs from '/kafka-acls/foobar' path
adminClient.describeAcls(... TOPIC, "*" ...) -> would return only ACLs from '/kafka-acls/*' path

New functionality

Using legacy constructors

Legacy constructors default resourceNameType to Literal and maintains existing contract:
adminClient.describeAcls(... TOPIC, "foobar" ...) -> still returns only ACLs from '/kafka-acls/foobar' path
adminClient.describeAcls(... TOPIC, "*" ...) -> still returns only ACLs from '/kafka-acls/*' path

Using constructors with explicit resourceNameType

The user needs to know, up front, which prefixed resource paths exist to be able to query them:
adminClient.describeAcls(... TOPIC, "foobar", Literal ...)-> will return only ACLs from '/kafka-acls/foobar' path
adminClient.describeAcls(... TOPIC, "*", Literal ...) -> will return only ACLs from '/kafka-acls/*' path
adminClient.describeAcls(... TOPIC, "foo", Prefixed ...) -> will return only ACLs from '/kafka-prefixed-acls/foo' path

Using 'ANY' resource name type

This allows user to discover all the ACLs affecting a specific resource:
adminClient.describeAcls(... TOPIC, "foobar", Any ...) -> will return all ACLs affecting topic "foobar", including any prefixed and wildcard ACLs.

The return value from describeAcls will contain the resourceNameType field for each ACL, so the user can determine if it is literal or prefixed.

Acls on prefixed resource paths are never returned to earlier clients. Nor can older clients delete ACLs on prefixed resource paths.

Proposed Changes

Solution

The proposal is to enhance the SimpleAclAuthorizer to support prefixed ACLs.
This means that it will be possible to create ACLs of type: User:clientA has READ access on topic prefixed with 'orgA' from 'hostA', i.e clientA has READ access to all topics that start with `orgA` from hostA.
The concept of prefixed ACLs will be applicable only to resource names.

Storage model

Currently, ACLs are stored on ZK under path /kafka-acl/<resource-type>/<resource-name>.

For example:
ACLs for topic topicName will be stored under /kafka-acl/Topic/topicName.
ACLs for consumer group groupId will be stored under /kafka-acl/Group/groupId.

An example ACL definition looks like:

$ get /kafka-acl/Topic/topicName
{"version":1,"acls":[{"principal":"User:clientA","permissionType":"Allow","operation":"Read","host":"*"},{"principal":"User:clientA","permissionType":"Allow","operation":"Write","host":"*"},{"principal":"clientB","permissionType":"Allow","operation":"Write","host":"host1"}]}

Current supported resource names are either full resource names like topicName or a special wildcard '*'.

$ get /kafka-acl/Topic/*
{"version":1,"acls":[{"principal":"User:clientA","permissionType":"Allow","operation":"Read","host":"*"}]}
which means that clientA has read access to all topics from all hosts.

The challenge here is that there can be both literal and prefixed resource paths with the same name. Some resources like consumer groups don't have any defined naming convention, so can include any characters in their name. It is therefore not possible to prepend/append a 'special character' to prefixed names to distinguish them from literal ones.

We extend the same storage model to store prefixed ACLs in a different location 'kafka-prefixed-acl'. Changes will first be stored at 'kafka-prefixed-acl-changes'. Literal resources, including the wildcard resource '*', will continue to be stored in their original location.

$ get /kafka-prefixed-acl/Topic/teamA
{"version":1,"acls":[{"principal":"User:clientA","permissionType":"Allow","operation":"Read","host":"*"}]}

ACLs write path

Write to a new location 'kafka-prefixed-acl'.

$ get /kafka-prefixed-acl/Topic/orgName
{"version":1,"acls":[{"principal":"User:clientA","permissionType":"Allow","operation":"Read","host":"*"}]}

ACLs read path

On read path, we look for all matching ACLs when authorize() is called.

Access will be allowed if there is at least one ALLOW matching acl and no DENY matching ACL (current behaviour is maintained). Note that the length of the prefix doesn't play any role here.

Compatibility, Deprecation, and Migration Plan

On downgrade, any prefixed ACLs will be ignored because they are in separate path. This means that any prefixed ACLs will be treated as if they were never added. This is fine for ALLOW ACLs, but might have security implications if DENY ACLs are ignored.

Constructors for Resource and ResourceFilter that don't take a ResourceNameType will be deprecated in favour of those that do.

Care has been taken to ensure legacy clients can neither add, list or remove prefixed ACLs. Clients wishing to use prefixed ACLs will need to upgrade their (admin)clients.

Rejected Alternatives

• Use escaping to identify prefixed vs literals (won't work on ZK). We decided to use a separate path for prefixed ACLs