Status

Current state: draft  [One of "Under Discussion", "Accepted", "Rejected"]

Discussion thread: here

JIRA: here KAFKA-3209

Released: <Kafka Version>0.10.2.0

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

Motivation

...

Motivation

This proposal is for adding a record transformation API to Kafka Connect as well as certain bundled transformations. At the same time, we should not extend Connect's area of focus beyond moving data between Kafka and other systems. We will only support simple 1:{0,1} transformations – i.e. map and filter operations.

The objective is to:

• Allow for lightweight updates to records.
  • Some transformations must be performed before the data hits Kafka (source) or

...

• • another

...

• • system (sink)

...

• • e.g. filtering certain types of events or sensitive information.
  • It's also useful for very light modifications that are easier to perform inline with the data import/export. It may be inconvenient to add stream processing into the mix for simple data massaging or control over routing.

• Benefit the growing connector ecosystem since some common options that are widely applicable can now be implemented once and reused. For example,
  • It is common for source connectors to allow configuring what format the topic name should follow based on some aspect of the source data, or in the case of sink connectors what 'bucket' (table, index etc.) a record should end up in based on the topic. This is configured in many different ways currently.
  • Some sink connectors allow inserting record metadata like the Kafka topic/partition/offset into the record key or value, while others do not. This information can get lost in translation if the functionality is absent and makes a connector less useful.
  • See the 'bundled transformations' section below for more examples.

Public Interfaces and Proposed Changes

Java API

// Existing base class for SourceRecord and SinkRecord, new self type parameter.
public abstract class ConnectRecord<R extends ConnectRecord<R>> {
 
    // ...
 
    // New abstract method:
       
    /** Generate a new record of the same type as itself, with the specified parameter values. **/
    public abstract R newRecord(String topic, Schema keySchema, Object key, Schema valueSchema, Object value, Long timestamp);
 
}

public interface Transformation<R extends ConnectRecord<R>> extends Configurable, Closeable {
 
    // via Configurable base interface:
    // void configure(Map<String, ?> configs);

    /**
     * Apply transformation to the {@code record} and return another record object (which may be {@code record} itself) or {@code null},
     * corresponding to a map or filter operation respectively. The implementation must be thread-safe.
     */
    R apply(R record);

    /** Configuration specification for this transformation. **/
    ConfigDef config();

    /** Signal that this transformation instance will no longer will be used. **/
    @Override
    void close();

}

Configuration

A transformation chain will be configured at the connector-level. The order of transformations is defined by the transforms config which represents a list of aliases. An alias in transforms implies that some additional keys are configurable:
- transforms.$alias.type – fully qualified class name for the transformation
- transforms.$alias.* – all other keys as defined in Transformation.config() are embedded with this prefix

Example:

transforms=tsRouter,insertKafkaCoordinates

transforms.tsRouter.type=org.apache.kafka.connect.transforms.TimestampRouter
transforms.tsRouter.topic.format=${topic}-${timestamp}
transforms.tsRouter.timestamp.format=yyyyMMdd
 
transforms.insertKafkaCoordinates.type=org.apache.kafka.connect.transforms.InsertInValue
transforms.insertKafkaCoordinates.topic=kafka_topic
transforms.insertKafkaCoordinates.partition=kafka_partition
transforms.insertKafkaCoordinates.offset=kafka_offset

Runtime changes

For source connectors, transformations are applied on the collection of SourceRecord retrieved from SourceTask.poll().

For sink connectors, transformations are applied on the collection of SinkRecord before being provided to SinkTask.put().

If the result of any Transformation.apply() in a chain is null, that record is discarded (not written to Kafka in the case of a source connector, or not provided to sink connector).

Bundled transformations

Criteria: SMTs that are shipped with Kafka Connect should be general enough to apply to many data sources & serialization formats. They should also be simple enough to not cause any additional library dependency to be introduced.

Beyond those being initially included with this KIP, transformations can be adopted for inclusion in future with JIRA/ML discussion to weigh the tradeoffs.

ValueToKey

Hoist{Key,Value}ToStruct

Extract{Key,Value}FromStruct

Patterns for implementing data transformations

• Data transformations could be applicable to the key or the value of the record. We will have *Key and *Value variants for these transformations that reuse the common functionality from a shared base class.

• Some common utilities for data transformations will shape up:

  • Cache the changes they make to Schema objects, possibly only preserving last-seen one as the likelihood of source data Schema changing is low.

  • Copying of Schema objects with the possible exclusion of some fields, which they are modifying. Likewise, copying of Struct object to another Struct having a different Schema with the exception of some fields, which they are modifying.

  • Where fields are being added and a field name specified in configuration, we will want a consistent way to convey if it should be created as a required or optional field. We can use a leading '!' or '?' character for this purpose if the user wants to make a different choice than the default determined by the transformation.
    

  • ConfigDef does not provide a Type.MAP, but for the time being we can piggyback on top of Type.LIST and represent maps as a list of key-value pairs separated by :.
  • Where field names are expected, in some cases we should allow for getting at nested fields by allowing a dotted syntax which is common in such usage (and accordingly, will need some utilities around accessing a field that may be nested).
  • There are escaping considerations to several such configs, so we will need utilities that that assume a consistent escaping style (e.g. backslashes).

Compatibility, Deprecation, and Migration Plan

There are no backwards compatibility concerns. Transformation is an additional layer at the edge of record exchange between the framework and connectors.

Test Plan

Unit tests for runtime changes and each bundled transformation, as well as system test exercising a few different transformation chains.

Rejected Alternatives

Transformation chains as top-level construct

The current proposal is to have transformation chains be configured in the connector config under the prefix transforms. An alternative would be to reference a transformation chain by a name in the connector configuration, with the transformation chain specification managed separately by Connect.

However, the surface area for such a change is much larger - we would need additional REST APIs for creating, updating and validating transformation chain configs. The current proposal does not prevent taking this direction down the line.

Not including any transformations with Connect

In the interest of providing a better out-of-the-box experience and avoiding duplication of effort in the ecosystem, we will be bundling certain transformations with Connect.

One concern here is that we should have a well-defined criteria for what belongs in Connect vs external dependencies, which was addressed

Currently, the "data transfer" logic is locked within a single function, thus Kafka Connect API is inflexible for making such modifications. This is more so when a Kafka Connect API is provided via third party where the user cannot modify the source. Custom modifications from user to user would require modifying the source of Connect API. 

Public Interfaces

For the transformers, we require to introduce an abstract class which others can extend to create their own custom Transformer. 

public abstract class Transformer<T1, T2> {
    public abstract T2 transform(T1 t1);
    public void initialize(Map<String, String> props) {}
}

Also, this would need a :

transformers=org.AbcTransformer,org2.XyzTransformer

property in the Kafka Connect config files to initialise the transformers, in the order specified in the above property. 

To accommodate the transformers, we can split the “functional” methods into separate parts from a user logical POV:

Proposed Changes

Besides breaking the functions push() and poll() of SinkTask and SourceTask respectively, the change introduced is the optional Transformers chain that users can introduce to operate on the message when being passed from one source to another, one of them being Kafka (as consistent with the Kafka Connect API.) 

Users can define and add their own 

Describe the new thing you want to do in appropriate detail. This may be fairly extensive and have large subsections of its own. Or it may be a few sentences. Use judgement based on the scope of the change.

Compatibility, Deprecation, and Migration Plan

• What impact (if any) will there be on existing users?
• If we are changing behavior how will we phase out the older behavior?
• If we need special migration tools, describe them here.
• When will we remove the existing behavior?

Test Plan

Describe in few sentences how the KIP will be tested. We are mostly interested in system tests (since unit-tests are specific to implementation details). How will we know that the implementation works as expected? How will we know nothing broke?

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.