Status

Current state: Under Discussion

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

Since its creation, the Kafka Connect Java API has provided support for offsets, which track the progress of both source and sink connectors. Tasks can use these offsets on startup to begin reading from the last-consumed point in their upstream system (i.e., Kafka for sink tasks and the external system for source tasks), instead of starting over at the beginning (causing duplicate data to be written) or at the end (causing data to be missed).

However, Kafka Connect does not yet provide an official API for cluster administrators to view or modify the offsets of connectors on the cluster. Some cases where this may be useful include:

• Resetting the offsets for a connector while iterating rapidly in a development environment (so that the connector does not have to be renamed each time)
• Viewing the in-memory offsets for source connectors on the cluster in order to recover from accidental deletion of the Connect source offsets topic (this is currently possible with the config topic, and can be a life-saver)
• Monitoring the progress of running connectors on a per-source-partition basis
• Skipping records that cause issues with a connector and cannot be addressed using existing error-handling features

Kafka Connect has also evolved over the years to support some use cases that make offset management complex. These include:

• Creating connectors that target a different Kafka cluster than the one that hosts the Kafka Connect cluster's internal topics (possible as of KIP-458: Connector Client Config Override Policy)
• Creating sink connectors with an overridden, user-specified consumer group ID (possible as of KIP-458: Connector Client Config Override Policy)
• Creating source connectors with a custom offsets topic (possible as of KIP-618: Exactly-Once Support for Source Connectors)

Any support for viewing and manipulating connector offsets added to Kafka Connect should accommodate all of these cases.

Public Interfaces

Endpoints

Several new endpoints will be added to the Kafka Connect REST API:

Response format

Queries for a connector's offsets will use one of these formats, depending on whether the offsets are for a source or a sink connector. Care is taken to keep the two formats symmetrical.

{
  "type": "SOURCE",
  "offsets": [
    {
      "partition": {
        // Connector-defined source partition
	  },
      "offset": {
        // Connector-defined source offset
      }
    }
  ]
}

{
  "type": "SINK",
  "offsets": [
    {
      "partition": {
        "topic": // Kafka topic
        "partition": // Kafka partition
	  },
      "offset": {
        "offset": // Kafka offset
      }
    }
  ]
}

Offsets endpoints behavior

The GET /connectors/{connector}/offsets  endpoint will be useful for examining the offsets of a currently-running connector. It will automatically deduce whether the connector is a source or sink based on the connector's configuration, and then return the offsets for the connector. The request will fail with a 404 response if the connector does not exist on the cluster.

The DELETE /connectors/connector/offsets  endpoint will be useful for resetting the offsets of stopped connectors. Requests will be rejected if the connector does not exist on the cluster (based on the config topic), if a rebalance is pending, or if the connector is not in the STOPPED  state (described below). If successful, the request will be met with a 204 ("no content") response and an empty response body. The request will fail with a 404 response if the connector does not exist on the cluster, and will fail with a 400 response if the connector does exist but is not in the correct state. This endpoint will be idempotent; multiple consecutive requests to reset offsets for the same connector with no new offsets produced in between those requests will all result in a 204 response (if they are successful).

A source offset will only be considered successfully deleted if the Connect worker is able to emit a tombstone to the offsets topic for its partition, and then read to the end of the offsets topic. A request to reset offsets for a source connector will only be considered successful if the worker is able to delete all known offsets for that connector, on both the worker's global offsets topic and (if one is used) the connector's dedicated offsets topic.

If exactly-once source support is enabled on a worker that receives a request to reset offsets for a source connector, it will fence out all previously-running tasks for the connector (if any exist, based on the presence of a task count record in the config topic) before resetting the offsets for the connector. Offsets will be reset transactionally for each topic that they exist in: a single transaction will be used to emit all tombstone records for the connector's dedicated offsets topic (if one is used) and another transaction will be used to emit all tombstone records for the worker's global offsets topic.

Requests to reset sink connector offsets will be satisfied by deleting the consumer group for that sink connector (as opposed to deleting all known offsets for that consumer group).

All offset reset requests will be forwarded to the leader of the cluster.

New target state: STOPPED

Background: target states today

Kafka Connect currently supports two "target states" for a connector: RUNNING  (the default), and PAUSED. The target state for a connector can be controlled using the REST API, using the PUT /connectors/connector/resume  endpoint for the RUNNING  state and the PUT /connectors/connector/pause  endpoint for the PAUSED  state.

When a connector is paused, its tasks continue to exist on the cluster. Many resources allocated for them, including Kafka and Kafka Connect resources such as clients, converters, and transformations, and connector-specific resources such as database connections, file descriptors, memory buffers, remain allocated. This can lead to confusing and sometimes suboptimal behavior when users pause connectors but notice that resources (especially database connections) are still allocated.

In addition, connectors and tasks are allocated across the cluster without regard to their target state during rebalance. This can lead to some skew in resource usage (such as network traffic and CPU utilization) across the cluster; in the worst case, all paused tasks are allocated to one set of workers, and all running tasks are allocated to a disjoint set of workers.

A new STOPPED state

A new target state will be added for connectors: STOPPED. The semantics for a connector that becomes stopped will be:

• The connector config remains present in the config topic of the cluster (if running in distributed mode), unmodified
• The connector config remains visible in the REST API
• All tasks for the connector are shut down completely. If running in distributed mode, a set of empty tasks is published to the config topic for the connector
• If running in distributed mode, as a result of the empty set of task configs being published to the config topic, a rebalance will be triggered, and no tasks for the connector will be allocated across the cluster
• As a further result of that rebalance, any information on the connector provided by the REST API will show it with an empty set of tasks
• The Connector, if running, will be shut down (by invoking Connector::stop and deallocating all Kafka Connect- and Kafka-related resources allocated for it by Kafka Connect)
• The Connector  will still appear in the status portion of the REST API, with a state of STOPPED. This will take place even if the connector was in the FAILED state before the request to stop it, or if it failed during shutdown in response to a request to stop
• If running in distributed mode, the Connector  will still be assigned to a worker during each rebalance
• The Connector will not be started (by invoking Connector::start), and it will not be able to generate new task configurations (by invoking ConnectorContext::requestTaskReconfiguration)

When a stopped connector is resumed or paused (there is no difference between the two transitions), the semantics will be:

• The Connector  is started on the worker that it is assigned to, and allowed to generate a set of task configurations
• In standalone mode, tasks are immediately brought up based on that set of configurations
• In distributed mode, if that set of configurations is non-empty, it is written to the config topic, and a rebalance ensues
• All further behavior is identical to any other cases where a new set of task configurations for a connector is generated (including a connector being reconfigured via the REST API, or a connector automatically generating a new set of configurations based on a monitoring thread that polls the external system for changes)

Compatibility, Deprecation, and Migration Plan

Offsets endpoints

This feature is fully backwards-compatible with existing Kafka Connect releases. Migration should occur automatically whenever a Kafka Connect cluster is upgraded to a version that supports this feature.

STOPPED target state

Diligent readers will note that the addition of a new target state creates problems with cluster downgrades and rolling upgrades. If a connector is stopped, and a worker running an older version of Kafka Connect either joins or currently exists in the cluster, that worker may not know how to handle the new record in the config topic that includes the request to stop the connector.

With some careful work, we can actually gracefully degrade in this scenario instead of confusing the worker running the older version of Kafka Connect.

Instead of publishing a "naive" record to the config topic with contents like this:

{"state": "STOPPED"}

We can "fool" older workers into treating STOPPED  requests as PAUSE  requests by emitting records with this format:

{
	"state": "PAUSED",
	"state.v2": "STOPPED"
}

Older workers will inspect the state  field of the record value, see that it is PAUSED, and pause the parts of the connector that they are assigned.

Newer workers will first inspect the state.v2 field and, if it is found, use that as the new target state. If no state.v2  field is found, they will fall back on the state  field.

Test Plan

Integration and/or system tests will be added for these cases:

Offsets tests:

• Cannot view the offsets of a nonexistent connector
• Can view the offsets of a running sink and source connector, and verify that that those offsets reflect an expected level of progress for each connector (i.e., they are greater than or equal to a certain value depending on how the connectors are configured and how long they have been running)
• Cannot reset the offsets of a nonexistent, paused, or running connector
• Can reset the offsets of a stopped sink and source connector
• Second follow-up requests to reset offsets for those connectors after their offsets have already been reset once are also successful
• Can view/reset the offsets for a sink connector that uses an overridden consumer group ID
• Can view/reset the offsets for a source connector that uses a custom offsets topic
• Can view/reset the offsets for a sink and source connector that targets a different Kafka cluster than the one used for the internal topics of the Kafka Connect cluster

Stopped state tests:

• Can stop a running connector
• Can stop a paused connector
• Cannot see task configuration or status information in the REST API for a stopped connector
• Can resume a stopped connector
• Can pause a stopped connector
• Stopping a failed connector updates its state to STOPPED  in the REST API
• Stopping a connector that fails during shutdown after receiving a stop request updates its state to STOPPED  in the REST API
• Can resume a connector after its Connector  has failed both before and during shutdown after receiving a stop request

Future work

Overwriting offsets

Summary: We may add finer-grained support for modifying connector offsets beyond resetting all of the offsets for a connector. This could include resetting offsets for a subset of partitions, or writing new offsets for partitions.

Delayed because: We may refine our error-handling features instead of augmenting offsets support this way, and the cost of implementing this feature is not high enough to justify is inclusion in the first draft of offsets support.

Automatically delete offsets with connectors

Summary: Add a mechanism (such as an includeOffsets  URL query parameter) to allow offsets to be reset at the same time a connector is deleted.

Delayed because: The cost of implementing this feature is not high enough to justify its inclusion in the first draft of offsets support, though it may be a nice quality-of-life improvement later on. The difficulty here is finding a way to ensure that the connector and all of its tasks are stopped before its offsets are reset, since the existing endpoint to delete connectors does not synchronously wait for the connector to actually be deleted (only for the deletion request to be persisted in the config topic). One option might be to add some kind of offsets reset request record to the config topic directly after the tombstone for the connector config.

Rejected Alternatives

Source-only support

Summary: Only provide support for source connector offsets instead of both source and sink connector offsets, since there is existing CLI tooling for sink connector offsets provided already.

Rejected because: The additional cost of supporting sink connector offsets, at least with the APIs proposed here, is negligible, and it will be easier for Connect cluster administrators to interact with a single unified API for both source and sink connectors.

Resetting offsets for active connectors

Summary: Allow offsets to be reset for currently-running connectors.

Rejected because: This would be significantly harder to implement, and the benefits are negligible compared to the workaround of stopping the connector, resetting its offsets, and then recreating it.