Status

Current state: Under Discussion

Discussion thread: here

JIRA: KAFKA-3492

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

Motivation

KIP-13 introduced client quotas in Kafka 0.9.0.0. Rate limits on producers and consumers are enforced to prevent clients saturating the network or monopolizing broker resources. The current implementation allocates quotas to client-ids. This works well in single user clusters or clusters that use PLAINTEXT where all users have the same identity. But since client-id is unauthenticated and can be set to any value by the client, multi-tenant secure installations require secure quotas to be enforced to guarantee fair allocation of resources and prevent denial-of-service.

With the introduction of security in Kafka 0.9, the identity of Kafka clients is the user principal.  User principal is an authenticated user or a grouping of unauthenticated users chosen by the broker using a configurable PrincipalBuilder and is currently used for ACLs. Client-id is a logical grouping of clients with a meaningful name that is used in client metrics and logs. Multi-user systems have a hierarchy - user owns zero or more clients. (user-principal, client-id) defines a safe group of clients. The shorter unsafe client-id is sufficient in client metrics and logs, but quotas should be allocated to safe groups to avoid clients of one user throttling clients of another user with the same client-id.

This KIP proposes the following changes to the existing implementation:

1. Quota configuration for user principal. This prevents users generating heavy traffic from monopolizing resources and impacting the performance of other users in a multi-tenant cluster.
2. Sub-quotas for clients of an authenticated user.  Like the current client-id implementation, this enables a user to rate-limit some producers or consumers to ensure that they don’t impact other more critical clients.  For instance, users may be able to rate-limit an auditing client running in the background, leaving resources always available for a critical event processing client.
3. Client-id quotas for users with unlimited quota. Existing quota configuration for client-ids will continue to be applied to users with unlimited quota, but these will be applied as per-user quota for clients that share both user principal and client-id.
4. Shared quotas for groups of clients that span multiple users will no longer be supported.

Public Interfaces

Quota Entity

Quotas are currently configured for client-ids. All clients with the same client-id are currently grouped together as a quota entity, enforcing one quota for all clients with the same client-id. This KIP proposes to define quotas for safe client groups which share the same user-principal and client-id. In a single user cluster, this retains the current semantics of client-id quotas.

Configuration Options

Two new configuration options will be added to specify default producer and consumer quotas for users. The existing default configuration options for client-id quotas will be applied only if default user quota is unlimited.

New properties

quota.user.producer.default, quota.user.consumer.default: Default quota for producers/consumers of users without a user quota override. This will be set to unlimited quota (Long.MaxValue) by default.

Changes to existing properties

quota.producer.default, quota.consumer.default: Default client-id producer/consumer quota  is currently applied to each unique client-id across all users. This will be modified to be a per-user quota for each unique client-id of each user. This client-id default will is applied only if default user quota is unlimited.

Default configuration

• All clients have unlimited quota by default
• If quota.user.producer.default, quota.user.consumer.default are set, these default quotas are allocated to each user principal.
• If quota.user.producer.default, quota.user.consumer.default are unlimited, quota.producer.default, quota.consumer.default are allocated to each unique client-id of each user.

Metrics

Quota related metrics are currently generated for client-ids and use the tag client-id. The metrics tag will be changed to quota-id and the value will include base-64 encoded user principal.

Sensor names are currently a sensor type concatenated with the client id value, eg. FetchThrottleTime-clientA. This will be modified to use quota-id instead of client-id. This is not a public interface change since sensor names are not reflected in JMX metrics.

Tools

kafka-configs.sh will be extended to support authenticated user quotas and sub-quotas for clients of a user.  A new entity type “users” will be added.  The key-value pairs supported for users will be:

• producer_byte_rate : The total rate limit for the user’s producers
• consumer_byte_rate : The total rate limit for the user’s consumers
• client_producer_byte_rates: Comma separated list of reserved sub-quotas for client-ids of the user (eg. clientA:10,clientB:20). Clients not listed share the remaining quota of the user.
• client_consumer_byte_rates: Comma separated list of reserved sub-quotas for client-ids of the use. (eg. clientA:30,clientB:40). Clients not listed share the remaining quota of the user.

The existing entity type "clients" will be retained for backward compatibility. But quotas set for clients are used only for users without a config override and only if default user quota is unlimited.

Proposed Changes

User Principal 

Authenticated user principal will be obtained from the Session object. Base64-encoded hex string version of the Principal will be used so that it can be used as a node name in Zookeeper and in metrics without placing any restrictions on the characters allowed in the principal.  For PLAINTEXT, the principal is "anonymous" by default and quotas will be applied for that principal. But principal can be overridden using a custom principal builder even for PLAINTEXT, enabling different user quotas, for example, for connections from different IP addresses.

Quota Configuration

Quotas are currently configured as the rate limits for producers and consumers based on their client-id. Default rate limits can be configured for clients without a config override.The same set of limits will be configurable each quota-id.

// Quotas for user1 (without sub-quotas). Zookeeper persistence path /users/<encoded-user1>
{
    "version":1,
    "config": {
        "total" : {"producer_byte_rate":"1024","consumer_byte_rate":"2048"}
    }
}
// Quotas for user2 (with sub-quotas). Zookeeper persistence path /users/<encoded-user2>
{
    "version":1,
    "config": {
    "total" : {"producer_byte_rate":"4096","consumer_byte_rate":"8192"},
    "clients": {
       "clientA" : {"producer_byte_rate":"10","consumer_byte_rate":"20"},
       "clientB" : {"producer_byte_rate":"30","consumer_byte_rate":"40"}
    }
} 
// Quotas for client-id clientA of users without config override if default user quota is unlimited. Zookeeper persistence path /clients/clientA
{
    "version":1,
    "config": {
        "total" : {"producer_byte_rate":"100","consumer_byte_rate":"200"}
    }
}

In the sample configuration above:

1. Total rate limits for all clients with user principal user1 is (1024, 2048).
2. Total rate limits for all clients with user principal user2 is (4096, 8192).
   
   • The rate limits for clients with user principal user2 AND client-id clientA is (10, 20).
   • Clients of user2 with client-id other than clientA and clientB share the remaining quota (4056, 8132).
3. Total rate limits for all clients of user3  is (quota.user.producer.default, quota.user.consumer.default) configured in server.properties, since no config override is specified.
4. If default user quota is unlimited, clients of user3 use client-id quota configuration. For example quota for client-id clientA of user3 is (100, 200). And quota for client-id clientB of user3 without a client-id override is (quota.producer.default, quota.consumer.default)
   • In a single-user cluster, this provides the same semantics as the current client-id implementation
   • In a multi-user cluster,  quotas are now per-user, treating clientA of user4 as a different group from clientA of user2.

Quota Identifier

Quota configuration and metrics currently use client-id as the unique key, enforcing one quota for all clients with the same client-id. This will be replaced with a new quota-id that includes user principal. Each quota-id is associated with a pair of producer and consumer rate limits which may be config overrides or the default quota.

• quota-id is the concatenation of base64-encoded user principal and client-id. Clients-ids without a sub-quota override share the user's quota and hence use the encoded user principal as quota-id.
• In the example (non-encoded user principal is used here for readability):
  • All clients of user1 share the quota-id user1
  • clientA of user2 uses the quota-id user2clientA
  • clientC of user2 uses the quota-id user2 since it does not have a client quota override, sharing a quota with other clients of user2.
  • clientA of user3 uses the quota-id user3clientA

Quota Persistence in Zookeeper

Client-id based quota configuration overrides will continue be stored under /config/clients, but these will be applied only to clients of users without a quota override and only if default user quota is unlimited. Quota configuration overrides for user principals will be stored under /config/users and these will include any sub-quota overrides for clients of each user. Note that Base64-encoded hex version of the user principal will be used as node name under /config/users to cope with Zookeeper naming restrictions. The non-encoded user principal will be stored as a property to make it easy to identify the actual user associated with the path.

Tools

kafka-configs.sh will be extended to support a new entity type "users". Quota configuration for users will be provided as key-value pairs to be consistent with other configuration options. Hence no new command line arguments will be added to the tool. The tool will parse the key-value pairs specifying total user quota and possibly some client sub-quotas, validate these and convert them to the equivalent JSON for persistence in Zookeeper. The existing entity “clients” will continue to be supported to set client-id quotas for users with unlimited quota.

Compatibility, Deprecation, and Migration Plan

No client API changes are necessary to work with the new implementation. Existing client-id quota configuration will be processed from Zookeeper and since users have unlimited quota as default,  existing client-id quotas will be applied to all clients with the client-id. In a PLAINTEXT or single-user cluster, this provides the same semantics as now, but in a multi-user cluster, quotas will be applied per-user rather than across users.  For multi-user clusters that are currently using client-id rate limits, new per-user rate limits should be configured first before upgrading the cluster if shared rate limits are being used for clients of different users.

Rejected Alternatives

Broker configuration property to choose either client-id or user quota

Single user clusters require only client-id based quotas and some secure clusters may require only overall user quotas.  A simple switch between these two provides a simpler configuration and implementation.  But a server option that changes the semantics of quota scan be confusing and it is hard to configure a cluster without actually knowing which quota mode is being used. Since hierarchical quotas are useful in multi-user clusters,  it is better to handle both client-id  quotas and user quotas with semantics suitable to both secure and plaintext clusters.

Add quota-id to Kafka protocol to define a new grouping

Both client-id and user principal are groupings of clients currently used for other purposes. User-principals are used for ACLs and client-ids are used for client metrics and logs. It may be useful to define a quota group that doesn't have to align with existing uses of user principal or client-id. But this require addition of a new concept of quota-id into the Kafka protocol that is authenticated. Use of existing ids simplifies configuration and keeps the definitions consistent.