Eliminate lambda indirection overhead in DynamoDB Enhanced Client

[RanVaknin]

Motivation and Context

The DynamoDB Enhanced Client v2 shows significant performance regression compared to v1 DynamoDB Mapper, with 32-98% slower operations across all operations. One theory was that lambda allocation overhead in ResolvedImmutableAttribute.attributeGetterMethod() is a key bottleneck affecting all operations in the serialization hot path.

The change refactors how attribute values are extracted during serialization by replacing an inline lambda expression with a method reference to a dedicated instance method. Profiling shows this reduces allocations of iterator and builder objects, with LinkedHashMap$LinkedEntryIterator allocations dropping 44% and DefaultDynamoDbExtensionContext$Builder allocations dropping 30%. The method reference approach appears to allow the JVM to optimize the call path more effectively, though the exact mechanism linking the code change to these specific allocation reductions is not fully clear from profiling data.

The optimization shows the strongest impact on small objects, with Get TINY operations improving 47% compared to 7% for Get HUGE, suggesting the allocation overhead represents a larger proportion of total processing time for smaller payloads.

Change

Before:

Function<T, AttributeValue> getAttributeValueWithTransform = item -> {
    R value = immutableAttribute.getter().apply(item);  // Lambda indirection
    return value == null ? nullAttributeValue() : attributeType.objectToAttributeValue(value);
};

After:

public Function<T, AttributeValue> attributeGetterMethod() {
    return this::getAttributeValue; 
}

AttributeValue getAttributeValue(T item) {
    R value = getter.apply(item);
    return value == null ? nullAttributeValue() : attributeType.objectToAttributeValue(value);
}

---

Results

running existing benchmarks for test/sdk-benchmarks/src/main/java/software/amazon/awssdk/benchmark/enhanced/dynamodb

Operation	Size	V1 (ops/s)	Before Fix	After Fix	Fix Improvement	Remaining Gap to V1
Delete	TINY	12,149,315	10,699,554	11,865,284	+11%	-2%
Delete	SMALL	12,198,550	10,826,259	11,094,164	+2%	-9%
Delete	HUGE	12,247,132	10,704,169	11,379,428	+6%	-7%
Get	TINY	7,243,909	4,561,782	6,720,756	+47%	-7%
Get	SMALL	4,920,953	3,015,931	3,583,527	+19%	-27%
Get	HUGE	554,562	289,412	311,014	+7%	-44%
Put	TINY	6,824,135	4,672,933	5,870,566	+26%	-14%
Put	SMALL	4,296,701	2,605,126	2,807,175	+8%	-35%
Put	HUGE	582,056	225,242	223,333	-1%	-62%
Update	TINY	4,539,823	2,288,870	2,632,113	+15%	-42%
Update	SMALL	2,862,813	242,741	242,506	+0%	-92%
Update	HUGE	551,769	54,629	54,374	+0%	-90%

JFR Profiling (Get TINY benchmark):

Allocation Impact:

Object Type	Before	After	Change	%
LinkedHashMap$LinkedEntryIterator	1,000	565	-435	-44%
DefaultDynamoDbExtensionContext$Builder	1,390	974	-416	-30%
DefaultDynamoDbExtensionContext	1,431	1,170	-261	-18%
EnhancedAttributeValue$InternalBuilder	871	678	-193	-22%
GetItemEnhancedResponse	320	170	-150	-47%
HashMap$Node[]	1,865	1,627	-238	-13%
LinkedHashMap	1,404	1,211	-193	-14%
HashMap	871	543	-328	-38%
EnhancedAttributeValue	664	836	+172	+26%
LinkedHashMap$Entry	590	466	-124	-21%

CPU Impact:

Metric	Before	After	Change	%
Total CPU Samples	2,128	2,037	-91	-4.3%

[sonarqubecloud[bot]]

Quality Gate failed

Failed conditions
C Reliability Rating on New Code (required ≥ A)

See analysis details on SonarQube Cloud

Catch issues before they fail your Quality Gate with our IDE extension SonarQube for IDE