This Confluence has been LDAP enabled, if you are an ASF Committer, please use your LDAP Credentials to login. Any problems file an INFRA jira ticket please.

Child pages
  • KIP-401: TransformerSupplier/ProcessorSupplier StateStore connecting
Skip to end of metadata
Go to start of metadata

Status

Current state: Under Discussion

Discussion threadhttps://lists.apache.org/thread.html/600996d83d485f2b8daf45037de64a60cebdfac9b234bf3449b6b753@%3Cdev.kafka.apache.org%3E

JIRA: https://issues.apache.org/jira/browse/KAFKA-7523

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

Motivation

When writing low-level Processors and Transformers that are stateful using kafka streams, often the processors (or transformers, I'll use "processors" to refer to both for brevity) want to "own" one or more state stores, the details of which are not important to the business logic of the application.  However, when incorporating these into a topology defined by the high level DSL, using KStream:process, you're forced to specify the state store names so the topology is wired up correctly.  This creates a clumsy pattern where the "owned" state store's name must be passed alongside the TransformerSupplier, when the supplier itself could just as easily supply that information on their own.

An example of the clumsiness:

String stateStoreName = "my-store";
StoreBuilder<KeyValueStore> storeBuilder =
        Stores.keyValueStoreBuilder(Stores.inMemoryKeyValueStore(stateStoreName), keySerde, valSerde);
topology.addStateStore(storeBuilder);
ProcessorSupplier processorSupplier = new MyStatefulProcessorSupplier(stateStoreName, val -> businessLogic(val));
builder.stream("input.topic")
        .map(...)
        .filter(...)
        .process(processorSupplier, stateStoreName);

Both the main topology definition (the chained, high-level DSL calls on StreamBuilderKStream, and KTable) and the internal implementation of MyStatefulProcessorSupplier need to know the state store name, when it should really only by MyStatefulProcessorSupplier that cares.  Additionally, topology.addStateStore(storeBuilder) and the creation of the StoreBuilder are required, all of which ought to be implicit when using MyStatefulProcessorSupplier.  Ultimately, because KStream:process requires store names as a separate argument, all of this "wiring" code is necessary alongside or nearby actual business logic.

Ideally, it would be reducible to something like:

builder.stream("input.topic")
        .map(...)
        .filter(...)
        .process(MyStatefulProcessorSupplier.make(val -> businessLogic(val)));


This allows for the same "reads top to bottom" type of clarity as when using Processors (and Transformers) as when using the high-level DSL.

Public Interfaces

Add an interface ConnectedStoreProvider that allows the implementor to specify state stores that should be connected to this processor/transformer (defaulting to no stores).

ConnectedStoreProvider

public interface ConnectedStoreProvider {
    default Set<StoreBuilder> stores() {
        return null;
    }
}


Change all Processor/TransformerSupplier  interfaces to extend from it:


TransformerSupplier

public interface TransformerSupplier<K, V, R> extends ConnectedStoreProvider {
    ...
}

ValueTransformerSupplier

public interface ValueTransformerSupplier<V, VR> extends ConnectedStoreProvider {
    ...
}

ValueTransformerWithKeySupplier

public interface ValueTransformerWithKeySupplier<K, V, VR> extends ConnectedStoreProvider {
    ...
}

ProcessorSupplier

public interface ProcessorSupplier<K, V> extends ConnectedStoreProvider {
    ...
}


Proposed Changes

The proposal is to enhance the ProcessorSupplier and TransformerSupplier interfaces by allowing them to provide information about what state stores they "own" when constructing a topology using StreamsBuilderKStream::process, KStream::transform, KStream::transformValues, and Topology::addProcessor.

The public interface changes above directly imply what needs to be changed in KStream:  The process etc methods would get state store names from the list of StoreBuilders that the supplier (which implements ConnectedStoreProvider ) provides, rather than the var args stateStoreNames.

The process method would add the StoreBuilders to the topology using builder.addStateStore() and connect the store to that processor, rather than requiring the user to do it themselves.  In order to solve the problem of addStateStore potentially being called twice for the same store (because more than one Supplier specifies it), the check for duplicate stores in addStateStores will be relaxed to allow for duplicates if the same StoreBuilder instance for the same store name.

Compatibility, Deprecation, and Migration Plan

Because the added interface methods are default with a reasonable default, those additions are backwards compatible.

A user may continue to "connect" stores to a processor by passing stateStoreNames when calling stream.process/transform(...) .  This may be used in combination with a Supplier  that provides its own state stores by implementing ConnectedStoreProvider::stores() .

If a StoreBuilder  that was manually added is also returned by a ConnectedStoreProvider , there is no issue since adding a state store will now be idempotent.

No migration tools are required since it's a relatively minor library change.

Alternatives

Have the added method on the Supplier interfaces only return store names, not builders

This solves the original issue only partially, but with perhaps less "API risk."  The String... stateStoreNames argument would no longer be needed on the KStream methods, but the user would still need to manually add the StoreBuilders to the Topology.  The downside is we don't achieve the full reduction of "wiring up" code required when building the topology (the user still needs to know to call topology.addStateStore()), but the upside is that the StoreBuilder is less coupled to the *Supplier.  I don't consider this upside significant, but perhaps there are other use cases I'm not considering.

Do nothing

This is a "quality of life" API improvement, and nothing more, so maybe it's unneeded churn.  I favor doing something (obviously) because I think that while small, this change can be a major usability improvement for the low-level API.


  • No labels