Status

Current state: "Under Discussion"

Discussion thread: here

JIRA: KAFKA-4585

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

Motivation

Note: The discussion in this KIP applies to Java based (new) consumer only as the security feature is not supported by the old consumer.

From an authorization and ACL point of view, three operations (permission types) are defined for consumer groups: Describe, Read, All. By default, Read implies Describe, and All implies all the other operations.

Current consumer group related APIs and their minimum required permissions are listed in the following table:

DescribeGroup

The pattern we can see in this table is that a minimum Read permission is used for mutating APIs, whereas a minimum Describe permission is used for non-mutating APIs. One exception to this pattern is OffsetFetch, which is a non-mutating API, but requires a Read access. A Read access requirement for OffsetFetch is too restrictive, and unnecessary. Consider the following example by @ewencp in the corresponding JIRA's description: If we want to write a tool that only monitors offsets (no commits), we cannot achieve it with the current ACL settings. Because accessing the OffsetFetch API requires a Read permission; but a Read permission means we are also authorized to use the CommitOffset API (side note: for this tool to be able to read offsets of a group, it needs to have Describe access to the topics the group is consuming from. In other words, the tool will be able to see offsets of all topics (topic partitions) in the group it has Describe access to).

The other, and perhaps more compelling, incentive for this change is that the current ACL settings breaks a certain functionality (and this functionality seems to have been broken for a while). As mentioned in the above table the minimum required permission for DescribeGroup and OffsetFetch is Describe and Read, respectively. But implementation of the describe group command makes use of OffsetFetch API (version 0 and 1 pre-KIP-88, and version 2 post-KIP-88). Therefore, a user who is granted the current minimum requirement permission Describe for DescribeGroup still would not be able to run the describe group command and get the expected result. They would see something like this in the output:

Error: Executing consumer group command failed due to Not authorized to access group: Group authorization failed.

 If we make the change suggested below, the command runs successfully and reports the group offsets.

Proposed Changes

The change proposed by this KIP is very simple: to lower the minimum required permission of the OffsetFetch API from Read to Describe. These minimum required permissions are hard-coded in kafka.server.KafkaApis.scala inside each API handler method. For example, the part that enforces the minimum required permission for the OffsetFetch API currently looks like this:

if (!authorize(request.session, Read, new Resource(Group, offsetFetchRequest.groupId))) 
          offsetFetchRequest.getErrorResponse(requestThrottleMs, Errors.GROUP_AUTHORIZATION_FAILED)

And the proposal is to to modify it to:

if (!authorize(request.session, Describe, new Resource(Group, offsetFetchRequest.groupId))) 
          offsetFetchRequest.getErrorResponse(requestThrottleMs, Errors.GROUP_AUTHORIZATION_FAILED)

Additional Food for Thought

As I was experimenting with the ACLs for this KIP I made some observations that made me wonder about the rationale behind their current ACL settings; and whether they need adjustments too: 

1. Heartbeat does not seem to be a mutating action (no changes to the group or consumer seem to made when a heartbeat request is processed). If that is the case, should we also lower the minimum required permission of Heartbeat to Describe?
2. Users can see all groups in the cluster (using consumer group’s --list option) if they have Describe access to the cluster. Would it make sense to modify that experience and limit what is listed in the output to only those groups they have Describe access to? The reason is, almost anything else is accessible by a user only if the access is specifically granted (through ACL --add); and this scenario should not be an exception. The potential change would be updating the minimum required permission of ListGroup from Describe (Cluster) to Describe (Group). 

Compatibility, Deprecation, and Migration Plan

• A user that already has Read permission to a consumer group, with this change, would still be able to query the group like before (Read implies Describe). For this user the change is backward compatible.
• Consider a user with Describe access. The group Describe access implies access to DescribeGroup and FindCoordinator APIs; even though this user cannot make use of DescribeGroup, as explained above. Giving this user access to OffsetFetch API means fixing that broken experience.

In general, As a result of this change, Kafka admins may need to revisit the relevant ACLs and update them if necessary.

Rejected Alternatives