Status

Current state: "Under Discussion" Accepted

Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]
JIRA: here [Change the link from KAFKA-1 to your own ticket] DISCUSS, DISCUSS+VOTE

JIRA:

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

...

Today, we can use SCRAM authentication for Kafka Brokers when the cluster uses ZooKeeper (ZK) for the quorum servers. This is possible to bootstrap by first setting up the ZK servers and setting the inter-broker communication password by directly updating them on the ZK before the Kafka Brokers are started. See Configuring SCRAM for details. We wish to implement some similar mechanism for storing the Kafka Broker authentication credentials for SCRAM when the cluster uses KRaft for the quorum servers. We want these credentials to be stored before the Kafka cluster starts for the first time.

Public Interfaces

Briefly list any new interfaces that will be introduced as part of this proposal or any existing interfaces that will be removed or changed. The purpose of this section is to concisely call out the public contract that will come along with this feature.

A public interface is any change to the following:

...

Binary log format

...

The network protocol and api behavior

...

Any class in the public packages under clientsConfiguration, especially client configuration

• org/apache/kafka/common/serialization

• org/apache/kafka/common

• org/apache/kafka/common/errors

• org/apache/kafka/clients/producer

• org/apache/kafka/clients/consumer (eventually, once stable)

...

...

Command line tools and arguments

...

Proposed Changes

We will update the kafka-storage tool to take SCRAM credentials and store them when we format each node.

The kafka-storage tool is used to initialize the storage space for each broker and controller. One of the files created is the bootstrap.checkpoint which contains a set of ApiMessageAndVersion records that are needed for the bootstrap process of the cluster. I am proposing we add an interface to the kafka-storage tool so we can add the needed SCRAM credential updates, UserScramCredentialsRecord which is an ApiMessageAndVersion record, when the storage is formatted. With this, when the initial start of the cluster happens, each broker will have a copy of the SCRAM server side credential needed so they can communicate with each other.

I propose we add an option --add-

...

scram to the kafka-storage tool that will add an ApiMessageAndVersion records directly into the __cluster_metadata topic. (See KIP-801 for details on __cluster_metadata ). The option can be applied multiple times to the format command so that multiple SCRAM records can be added to bootstrap the cluster. Below is the new updates kafka-storage command.

./bin/kafka-storage.sh format -h 
usage: kafka-storage format [-h] --config CONFIG --cluster-id CLUSTER_ID [--add-metadatascram METADATASCRAM_CREDENTIAL] [--release-version RELEASE_VERSION] [--ignore-formatted] 

optional arguments:
  -h, --help            show this help message and exit
  --config CONFIG, -c CONFIG 
                        The Kafka configuration file to use. 
  --cluster-id CLUSTER_ID, -t CLUSTER_ID
                        The cluster ID to use. 
  --add-metadatascram METADATASCRAM_CREDENTIAL, -AS METADATASCRAM_CREDENTIAL 
                        Some METADATA MessageA SCRAM_CREDENTIAL to add to the __cluster_metadata log for this node e.g. 
                        'UserScramCredentialsRecord={"Name":"alice","Mechanism":1,"Salt":'SCRAM-SHA-256=[user=alice,password=alice-secret]'
						'SCRAM-SHA-512=[user=alice,iterations=8192,salt="MWx2NHBkbnc0ZndxN25vdGN4bTB5eTFrN3E=","SaltedPasswordsaltedpassword=":"mT0yyUUxnlJaC99HXgRTSYlbuqa4FSGtJCJfTMvjYCE=","Iterations":8192}]' 
  --release-version RELEASE_VERSION, -r RELEASE_VERSION
                        A KRaft release version to use for the initial metadata version. 
  --ignore-formatted, -g

 I I propose the METADATA SCRAM_CREDENTIAL argument will contain the type of the ApiMessageAndVersion record to be added followed by a JSON set of name value pairs to populate the record. Initially I will only add support for UserScramCredentialsRecord which will require arguments of a Name (which is implicitly type user), a mechanism which is 1 for a key value pair where the key is one of the SCRAM mechanisms supported, either SCRAM-SHA-256 or 2 for SCRAM-SHA-512, and the Salt, a Password or SaltedPassword, and the Iteration count. The salt and iteration count are not optional unlike in the kafka-config command used to talk with ZK. We want the record data to match exactly on each node and normally when a salt is not specified a random salt is chosen. Since each node is initialized individually this would result in records with different salts which we don’t wantvalue is a set of key value pairs of parameters to populate the UserScramCredentialsRecord. The SCRAM_CREDENTIAL argument is very similar to the argument passed to the kafka-config tool for configuring SCRAM in a ZK cluster. See Configuring SCRAM for details. 

The subarguments for the SCRAM_CREDENTIAL require a "user" key and either a "password" key or a "saltedpassword" key. If using the "saltedpassword" key you must also supply an "iteration" key and a "salt" key. The "iteration" and "salt" key are otherwise optional. However if they are not supplied, "iteration" count will default to 4096 and the "salt" will be randomly generated. The value for "salt" and "saltedpassword" is base64 encoding of binary data.

I propose to also add support for the argument parsing to include taking a file of arguments. This is a standard Argparse4j feature and will make it easier to bootstrap brokers and controllers with multiple SCRAM_CREDENTIALS. I propose to use the '@' character to proceed the filename argument which contains additional arguments. See the Argparse4j fromfileprefix manual entry for details.

Compatibility, Deprecation, and Migration Plan

These changes are additions only so there isn't a backwards compatibility issueto the kafka-storage tool only. There are no backwards compatibility issues with this change.

Test Plan

These changes are to the bootstrap of a cluster and have no impact of existing clusters at all. 

Rejected Alternatives

...

1. The original proposal was for adding raw records to the __cluster_metadata. These records would be describe as a single key value where the key is the name of the ApiMessageAndVersion record and the value a JSON encoding of the fields for that record. An example would be the following for a SCRAM credential record.

   Code Block
   UserScramCredentialsRecord={"name":"alice","mechanism":1,"salt":"MWx2NHBkbnc0ZndxN25vdGN4bTB5eTFrN3E=","SaltedPassword":"mT0yyUUxnlJaC99HXgRTSYlbuqa4FSGtJCJfTMvjYCE=","iterations":8192}'

   This proposal is not very customer friendly as the customer now needs to know what the fields of the underlying  UserScramCredentialsRecord are. It also has the issue that the underlying record format could change in the future requiring the command line to change. It is desired that even if the record format changes, the argument parsing shouldn't be affected.

2. Update kafka-config to take a format directory option and use the same arguments for altering SCRAM credentials to add them to the __cluster_metadata topic for bootstrap. The issues with this is that it requires multiple commands to format each node in a cluster. It also has the problem of adding a whole new block of code to kafka-config just to handle the bootstrap.checkpoint file and it would need logic to understand if the bootstrap had completed. 
3. Update kafka-storage to append records to bootstrap.checkpoint with multiple invocations of the tool. This would allow the use of the same command line arguments from kafka-config to be used. It was deemed a requirement that a single invocation of kafka-storage format all the records for bootstrap.