TechPrufer
Get started
Datadog
Datadog logoDatadog · Feature trail

Java APM Class Matching Optimization

1 post · Aug 2026

What this trail covers

Datadog's APM Java agent has significantly improved its class-matching performance during startup. This involved optimizing the process of identifying which classes to instrument for observability. The team developed a novel approach by encoding a prefix trie, used for matching class name prefixes, directly as a JVM constant. This allows the trie to be loaded with a single bytecode instruction, eliminating the need for resource lookups and file parsing during the constrained `premain` phase.

All Datadog trails
  1. How we improved APM Java startup by encoding a prefix trie as a JVM constant | DatadogThe Datadog APM Java team optimized class matching during startup by encoding a prefix trie as a single JVM constant. This constant is loaded with a single bytecode instruction (`ldc`), eliminating the need for resource lookups and file parsing in the constrained `premain` phase. The trie is stored as a sequence of chars, where each char can hold control information and content. Each node in the trie is encoded with the number of branches, branch characters, branch values, and jump offsets.