...

package org.apache.flink.streaming.api.operators;
 
/** The builder class for {@link OperatorAttributes}. */
@PublicEvolving
public class OperatorAttributesBuilder {     

	@Nullable private Boolean outputStreamRecordValueStored = null;
    @Nullable private Boolean inputStreamRecordStored = null;  

    public OperatorAttributesBuilder() {
    	...
    }

    /**
     * Set to truefalse if it is guaranteed that the operator will not store and access reference to the
     * StreamRecord#value that it has already emitted.
     */
    public OperatorAttributesBuilder setOutputStreamRecordValueStored(boolean outputStreamRecordValueStored) {...}

    /**
     * Set to false if it is guaranteed that the operator will not store and access reference to the
     * input StreamRecord instance.
     */
    public OperatorAttributesBuilder setInputStreamRecordStored(boolean inputStreamRecordStored) {...}
    
    /**
     * If any operator attribute is null, we will log it at DEBUG level and use the following
     * default values.
     * - outputStreamRecordValueStored defaults to true
     * - inputStreamRecordStored defaults to true
     */
	public OperatorAttributes build() {...}
}

...

@PublicEvolving
public interface StreamOperator<OUT> extends CheckpointListener, KeyContext, Serializable {
    ...
 
    default OperatorAttributes getOperatorAttributes() {
        return new OperatorAttributesBuilder().build();
    }
}
 
@PublicEvolving
public interface StreamOperatorFactory<OUT> extends Serializable {
    ...
 
    default OperatorAttributes getOperatorAttributes() {
        return new OperatorAttributesBuilder().build();
    }
}

3) Add the isObjectReuseCompliant isOutputValueStored method that returns false by default to the Function interface. And the RichFunction overwrite the isOutputValueStored method to return true .

@Public
public interface Function extends java.io.Serializable {
    /**
     * Returns truefalse if it is guaranteed that the function will not store and access
     * reference to the output value.
     */ 
    default boolean isOutputValueStored() {
        return true;
    }
}

@Public
public interface RichFunction extends Function {

    /**
     * RichFunction is able to put the values to the state backend so the method returns true by
     * default. For RichFunction that doesn't store output value to the state backend, it can return
     * false.
     */
    @Override
    default boolean isOutputValueStored() {
        return true;
    }
}

4) Update the description of pipeline.object-reuse  to mention that when it is false, Flink will decide whether to use object reuse based on the operator attributes.

Here is the updated description:

When it is true, objects that Flink internally uses for deserialization and passing data to user-code functions will be reused. When it is false, Flink will decide whether to use object reuse based on the operator attributes. Keep in mind that this can lead to bugs when the user-code function of an operation is not aware of this behavior.

Proposed Changes

1) Update OperatorChain to take advantage of the isOutputStreamRecordValueStored and isInputStreamRecordStored attribute.

...

• If pipeline.object-reuse is set to true, records emitted by this operator will be re-used.

• Otherwise, if isOutputStreamRecordValueStored() returns truefalse

  • If the downstream operator isInputStreamRecordStored() return false, records emitted by this operator will be re-used.

  • Otherwise, StreamRecord will be shallow-copied before being given to the next operator in the chain.

• Otherwise, records emitted by this operator will be deep-copied before being given to the next operator in the chain.

...

We will enumerate all operators/functions defined in Flink, identify those operators/functions which that do not store or access reference to the value of the output StreamRecord, and override their isOutputStreamRecordValueStored() method to return truefalse. Those operators which that do not store or  or access reference to the input StreamRecord should override the isInputStreamRecordStored() method to return false.

After implementing the proposed change, the operator/function developer will have the flexibility to avoid unnecessary deep-copy overhead by appropriately overriding the isOutputStreamRecordValueStored() method. This empowers Flink to achieve optimal performance with the default configuration.

POC and Analysis

We implement a POC and run the flink-benchmarks against the POC with global object-reuse disabled. We verify that, with the change proposed in the FLIP, many of the operators in the benchmark can enable object-reuse without code change. Only the custom operator and the AbstractUdfStreamOperator that contains a RichFunction cannot enable object-reuse to its output. In order to enable object-reuse of those operator/function, couple lines of code are needed to set the operator attributes accordingly. After that, all the operator in flink-benchmarks can enable object-reuse with the global object-reuse disable.

Compatibility, Deprecation, and Migration Plan

...