Missing Drools classloader in `canSkipClassLoaderByName` skip-list: `DynamicProjectClassLoader$DefaultInternalTypesClassLoader` causes severe startup/reload slowdown with the Drools executable-model canonical KieModule

[sushanbaskota]

Tracer Version(s)

1.66.0

Java Version(s)

25.0.3

JVM Vendor

Amazon Corretto

Bug Report

ClassLoaderMatchers.canSkipClassLoaderByName() already hardcodes a skip for
three Drools classloaders:

// drools
case "org.drools.core.rule.PackageClassLoader":
case "org.drools.wiring.dynamic.PackageClassLoader":
case "org.drools.core.rule.JavaDialectRuntimeData$PackageClassLoader":
  return true;

A fourth Drools-internal classloader used by the newer executable-model
canonical KieModule build path is not in that list:

org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader

Because it's missing, every class defined through this classloader falls
through into the full instrumentation-matching pipeline, including
ClassLoaderMatchers.buildHasClassMask(), which does an uncached
getResource() scan per class-name check across every active
instrumentation module. Drools creates a fresh instance of this
classloader for every KieBase build (once per rule, via
CanonicalKieModule.createKieBase → ProjectClassLoader.loadClass →
defineType), so the ClassLoaderValue cache backing hasClassMask()
never gets a hit — the expensive scan runs from scratch for every single
rule, every time a KieBase is (re)built (startup load of N rules, or any
runtime rule reload after a cache eviction).

At ~2000 rules this is severe: a single-threaded rule-load thread showed
~99% CPU time (elapsed=99.00s, cpu=76112.16ms) stuck entirely in this
path (full stack under Reproduction Code below).

Rule-load time with 8 concurrent loader threads did not improve at all vs.
1 thread (~264s either way) — evidence this was a per-classloader cache-miss
cost, not lock contention, since threads only ever contended on the
symptom (UrlJarFiles$Cache / URLClassPath locks under classloading),
not the root cause. Environment: Drools 8.44.2.Final, executable-model /
canonical KieModule path, ~2000 rules loaded via KieRepositoryImpl.addKieModule
→ CanonicalKieModule.createKieBase.

Confirmed fix / workaround: adding the missing classloader name to
dd.trace.classloaders.exclude / DD_TRACE_CLASSLOADERS_EXCLUDE (the
user-facing config already backed by the same skip-list check) resolves it
completely:

DD_TRACE_CLASSLOADERS_EXCLUDE=org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader

Before / after, same rule set, same machine:

Config	Rule-load time
8 threads, no exclude	~264,000 ms
1 thread, no exclude	did not complete in 8+ min (unbounded — cost scales per rule)
1 thread, with exclude	~193,000 ms
8 threads, with exclude	~54,000 ms

After the exclude, the same thread's stack shows it spending its time in
legitimate work only (S3 fetch + zip decompression of the rule's KJar), with
no Datadog instrumentation-matching frames at all.

Ask: please add
org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader
to the existing Drools block in canSkipClassLoaderByName
(dd-java-agent/agent-tooling/src/main/java/datadog/trace/agent/tooling/bytebuddy/matcher/ClassLoaderMatchers.java),
alongside the three classloaders already there, so this becomes the default
instead of requiring the manual config workaround:

// drools
case "org.drools.core.rule.PackageClassLoader":
case "org.drools.wiring.dynamic.PackageClassLoader":
case "org.drools.core.rule.JavaDialectRuntimeData$PackageClassLoader":
case "org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader": // add this
  return true;

Happy to open a PR with this one-line addition if that's preferred over a
maintainer making the change.

Expected Behavior

ClassLoaderMatchers.canSkipClassLoaderByName() should treat
org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader
the same as the three Drools classloaders already hardcoded in that method
(org.drools.core.rule.PackageClassLoader,
org.drools.wiring.dynamic.PackageClassLoader,
org.drools.core.rule.JavaDialectRuntimeData$PackageClassLoader) — i.e.
return true and skip the class immediately, without ever reaching
buildHasClassMask()'s getResource() scan.

Since Drools creates a fresh instance of this classloader per KieBase build,
the expected outcome is that repeated KieBase construction (startup rule
loading, or any runtime rule reload after a cache eviction) does not pay a
per-classloader, per-instrumentation-module getResource() cost on every
single build. Setting DD_TRACE_CLASSLOADERS_EXCLUDE to this classloader
name manually reproduces this expected behavior today (264s → 54s for ~2000
rules in our environment) — the ask is simply to make that the default,
matching how the other three Drools classloaders are already handled.

Reproduction Code

No isolated standalone reproducer (this surfaces at scale inside a large
Drools application, ~2000 rules), but the mechanism is straightforward to
reproduce in principle:

// with dd-java-agent attached, no DD_TRACE_CLASSLOADERS_EXCLUDE set:
for (int i = 0; i < 2000; i++) {
    // build a fresh executable-model KieBase per rule/module, e.g.:
    KieServices ks = KieServices.get();
    KieContainer kc = ks.newKieContainer(releaseId, ProjectClassLoader.class.getClassLoader());
    KieBase kb = kc.getKieBase(kBaseName); // triggers CanonicalKieModule.createKieBase
}

Each iteration creates a new DynamicProjectClassLoader$DefaultInternalTypesClassLoader
instance; with the agent attached and this classloader missing from the
skip-list, thread-dump sampling during the loop shows threads stuck in:

at java.lang.ClassLoader.getResource(...)
at org.drools.wiring.api.classloader.ProjectClassLoader.getResource(ProjectClassLoader.java:309)
at org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader.getResource(DynamicProjectClassLoader.java:145)
at datadog.trace.agent.tooling.bytebuddy.matcher.ClassLoaderMatchers.buildHasClassMask(ClassLoaderMatchers.java:175)
at datadog.trace.agent.tooling.bytebuddy.matcher.ClassLoaderMatchers$1.computeValue(ClassLoaderMatchers.java:150)
at datadog.instrument.utils.ClassLoaderValue.computeValueForKey(ClassLoaderValue.java:185)
at datadog.instrument.utils.ClassLoaderValue.getOtherValue(ClassLoaderValue.java:193)
at datadog.instrument.utils.ClassLoaderValue.get(ClassLoaderValue.java:92)
at datadog.trace.agent.tooling.bytebuddy.matcher.ClassLoaderMatchers.hasClassMask(ClassLoaderMatchers.java:160)
at datadog.trace.agent.tooling.bytebuddy.matcher.ClassLoaderMatchers.incompatibleClassLoader(ClassLoaderMatchers.java:70)
at datadog.trace.agent.tooling.bytebuddy.matcher.GlobalIgnoresMatcher.matches(GlobalIgnoresMatcher.java:39)
at net.bytebuddy.agent.builder.AgentBuilder$Default$ExecutingTransformer.doTransform(AgentBuilder.java:12799)
...
at org.drools.wiring.api.classloader.ProjectClassLoader.tryDefineType(ProjectClassLoader.java:193)
at org.drools.wiring.api.classloader.ProjectClassLoader.loadClass(ProjectClassLoader.java:130)
at org.drools.modelcompiler.CanonicalKieModule.createInstance(CanonicalKieModule.java:175)
at org.drools.modelcompiler.CanonicalKieModule.createKieBase(CanonicalKieModule.java:243)

Adding DD_TRACE_CLASSLOADERS_EXCLUDE=org.drools.wiring.dynamic.DynamicProjectClassLoader$DefaultInternalTypesClassLoader
removes this stack from every subsequent iteration.