Status

Current state: Under discussion

Discussion thread: here

JIRA:  Unable to render Jira issues macro, execution error.

PR (draft): https://github.com/apache/kafka/pull/7310

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

Motivation

The Connect framework uses an internal REST endpoint to relay task configurations from follower worker nodes to the leader. This endpoint is unique in that it is only meant to be invoked by Connect workers; every other endpoint is documented as part of the public REST API. This in turn leads to a problem when the Connect REST API is secured and authentication is required; every other kind of request will read user-supplied credentials for authentication and use those credentials when forwarding requests to the leader. However, since workers query this internal endpoint entirely on their own, they can't rely on authentication credentials that would be supplied in a user's REST request.

Successful requests this endpoint allow for arbitrary rewrite of task configurations, which is a significant security vulnerability that could lead to leaking of topic data, writing arbitrary data to topics, and other serious problems. As a result, it is imperative that access to this endpoint be restricted in order for any Kafka Connect cluster to be considered truly secure.

An ideal restriction of this endpoint would guarantee that requests made to it come exclusively from another work in the same cluster. Although mutual authentication via TLS, for example, may seem like a viable approach, this only accomplishes authentication and not authorization; that is, it verifies that the request comes from a trusted party with a given identity, but it does not make distinctions about whether party should be allowed to perform actions on the cluster. If mutual authentication is used for the Connect REST API, then this endpoint is still effectively unsecured since any user of the public REST API may also access this endpoint.

It should be noted that the goal here is not to completely secure any Kafka Connect cluster, but rather to patch an existing security hole for clusters that are already intended to be secure. A few examples of steps that should be taken in order to secure a Kafka Connect cluster include securing the public REST API (which can be done using a Connect REST extension), securing the worker group (which can be done with the use of ACLs on the Kafka broker), and securing the internal topic used by Connect to store configurations, statuses, and offsets for connectors (which can also be done with the use of ACLs on the Kafka broker). If any of these steps are not taken, the cluster is insecure anyways; therefore, relying on these precautions being in place in order to implement a fix for the problem posed by the internal REST endpoint used by Connect is acceptable.

However, it is not a goal of this KIP to limit the rollout of the new features based on whether the Connect cluster is already secure. With the pluggable nature of Kafka and Kafka Connect authorization, it would be difficult to know if all of the important Connect resources (worker group, internal topics, and Kafka Connect REST API to name a few) are actually secured. Additionally, the presence of other attack vectors shouldn't be justification for opening up a new one; it seems particularly dangerous in the event that someone accidentally misconfigures their Connect cluster and this endpoint ends up being exposed even though the user believes it to be protected.

Public Interfaces

There will be five new configurations added for distributed workers:

• internal.request.key.generation.algorithm
  • Purpose: the algorithm used to generate session keys
  • Type: string
  • Default: "HmacSHA256"
• internal.request.key.size
  • Purpose: the size of generated session keys, in bits
  • Type: int
  • Default: 256
• internal.request.key.rotation.interval.ms
  • Purpose: how often to force a rotation of the internal key used for request validation, or 0 if forced rotation should never occur
  • Type: long
  • Default: 3600000 (one hour)
• internal.request.signature.algorithm
  • Purpose: the algorithm to use to sign internal requests when sent from a follower worker to the leader
  • Type: string
  • Default: "HmacSHA256"
• internal.request.verification.algorithms
  • Purpose: a list of supported algorithms for verifying internal requests that are received by the leader from a follower
  • Type: list
  • Default: "HmacSHA256"

The default value for the connect.protocol configuration will also be altered from compatible to sessioned, so that the new request signing behavior proposed by this KIP will be enabled by default (once all workers in a cluster support it).

Additionally, although not part of the public API, the POST /connectors/<name>/tasks endpoint will be effectively disabled for public use. This endpoint should never be called by users, but since until now there hasn't been anything to prevent them from doing so, it should still be noted that anything that relies that endpoint will no longer work after these changes are made. The expected impact of this is low, however; the Connect framework (and the connectors it runs) handle the generation and storage of task configurations and there's no discernible reason for using that endpoint directly instead of going through the public Connect REST API.

Proposed Changes

A new Connect subprotocol, sessioned, will be implemented that will be identical to the cooperative incremental protocol but a higher protocol version number (2, instead of the current version for cooperative incremental rebalancing, which is 1). One downside of this approach is that the use of cooperative incremental assignments will be required in order to enable this new security behavior; however, given the lack of any serious complaints about the new rebalancing protocol thus far, this seems preferable to trying to enable this behavior across both assignment styles.

If the connect.protocol property is set to sessioned, the worker will advertise this new sessioned protocol to the Kafka group coordinator as a supported (and, currently, most preferable) protocol. If that protocol is then agreed on by the cluster during group coordination, a session key will be randomly generated by the leader and distributed to the cluster via the config topic. This key will be used by followers to sign requests to the internal endpoint, and verified by the leader to ensure that the request came from a current group member. It is imperative that inter-worker communication have some kind of transport layer security; otherwise, this session key will be leaked during rebalance to anyone who can eavesdrop on request traffic.

Periodically (with frequency dictated by the internal.request.key.rotation.interval.ms property), the leader will compute a new session key and distribute it to the cluster.

The default algorithm used to sign requests will be HmacSHA256; this algorithm is guaranteed to be supported on all implementations of the Java Platform (source). However, users will be able to configure their cluster to use other algorithms with the internal.request.signature.algorithm property if, for example, the default is not suitable for compliance with an existing security standard.

Similarly, the default algorithm used to generate request keys will also be HmacSHA256; again, this algorithm is guaranteed to be supported on all implementations of the Java Platform (source). And again, users will be able to configure their cluster to use other algorithms or keys of a different size with the internal.request.key.generation.algorithm and internal.request.key.size properties, respectively.

Each signed request will include two headers:

• X-Connect-Authorization: the signature of the request body (base 64 encoded)
• X-Connect-Request-Signature-Algorithm: the key algorithm used to sign the request

When a request is received by the leader, the request signature algorithm described by the X-Connect-Request-Signature-Algorithm header will be used to sign the request body and the resulting signature will be checked against the contents of the X-Connect-Authorization header. If the contents do not match, or the request signature algorithm is not in the list of permitted algorithms controlled by the internal.request.verification.algorithms property, the request will be rejected.

The leader will only accept requests signed with the most current key. This should not cause any major problems; if a follower attempts to make a request with an expired key (which should be quite rare and only occur if the request is made by a follower that is not fully caught up to the end of the config topic), the initial request will fail, but will be subsequently retried after a backoff period. This backoff period should leave sufficient room for the rebalance to complete. One potential downside is that, should this occur, an error-level log message of "Failed to reconfigure connector's tasks, retrying after backoff: " followed by a stack trace will be generated. This can be mitigated by altering the log message or the generated exception to include a note that this may not be an issue if key rotation is enabled, and/or logging an info-level log message after successfully completing task reconfiguration that potentially includes a note that any above error messages related to task reconfiguration may be safely disregarded.

Compatibility, Deprecation, and Migration Plan

Backwards compatibility

All of the proposed configurations here have default values, making them backwards compatible.

Reverting an upgrade

The group coordination protocol will be used to ensure that all workers in a cluster support verification of internal requests before this behavior is enabled; therefore, a rolling upgrade of the cluster will be possible. In line with the regression plan for KIP-415: Incremental Cooperative Rebalancing in Kafka Connect, if it is desirable to disable this behavior for some reason, the connect.protocol configuration can be set to compatible or default for one (or more) workers, and it will automatically be disabled.

Migrating to a new request signature algorithm

If a new signature algorithm should be used, a rolling upgrade will be possible with the following steps (assuming a new algorithm of HmacSHA489):

1. Add HmacSHA489 to the internal.key.verification.algorithms list for each worker, and restart them one-by-one
2. Change the internal.key.signature.algorithm property for each worker to HmacSHA489, and restart them one-by-one
3. (Optional) Remove the old algorithm from the internal.key.verification.algorithms list for each worker, and restart them one-by-one

Rejected Alternatives

Configurable inter-worker headers

Summary: A new worker configuration would be added that would control auth headers used by workers when making requests to the internal endpoint.

Rejected because: The additional complexity of another required configuration would be negative for users; security already isn't simple to implement with Kafka Connect, and requiring just one more thing for them to add should be avoided if possible. Also, the use of static headers isn't guaranteed to cover all potential auth mechanisms, and would require manual rotation by reconfiguring the worker.

Replace endpoint with Kafka topic

Summary: The REST endpoint could be removed entirely and replaced with a Kafka topic. Either an existing internal Connect topic (such as the configs topic) could be used, or a new topic could be added to handle all non-forwarded follower-to-leader communication.

Rejected because: Achieving consensus in a Connect cluster about whether to begin engaging in this new topic-based protocol would require either reworking the Connect group coordination protocol or installing several new configurations and a multi-stage rolling upgrade in order to enable it. Requiring new configurations and a multi-stage rolling upgrade for the default use case of a simple version bump for a cluster would be a much worse user experience, and if the group coordination protocol is going to be reworked, we might as well just use the group coordination protocol to distribute keys instead. Additionally, the added complexity of switch from a synchronous to an asynchronous means of communication for relaying task configurations to the leader would complicate the implementation enough that reworking the group coordination protocol might even be a simpler approach with smaller changes required.

Distribute session key during rebalance

Summary: Instead of distributing a session key via the config topic, include the session key as part of the worker assignment handed out during rebalance. Periodically force a rebalance in order to rotate session keys.

Rejected because: The implementation complexity of adding a session key to the rebalance protocol would be quite high, and the additional API would complicate the code base significantly. Additionally, there are few, if any advantages, compared to distributing the keys via the config topic.