[reference] reactor-core-3.1 async blind regeneration for #11927

[jordan-wong]

[reference] reactor-core-3.1 blind regeneration for #11927

Reference output from the toolkit's automated instrumentation-authoring
workflow, run against the draft skill in #11927. Not intended to be merged
as-is.

Research context

This PR is part of the async-category research cycle: verifying whether
the current skill draft in #11927 generalizes to Category B
(context-propagation) libraries after the HTTP-cycle skill updates.
reactor-core-3.1 was chosen as the calibration target for non-RxJava
async libraries — Reactor's Publisher/Subscriber chain is a fundamentally
different context-propagation pattern than RxJava's Observer chain, so
this PR tests whether the skill's Cat B guidance generalizes.

Changes applied on top of eval output

Original eval branch inherited a stash-merge revert of #11852 during eval
branch setup, dropping DD_TRACE_ROBOLECTRIC_ENABLED plus 172 unrelated
files. Rebased onto current master by cherry-picking only in-scope paths
under dd-java-agent/instrumentation/reactor-core-3.1/.

Three follow-up fixes were applied on top of the rebased eval output to
unblock local CI-equivalent without hiding the substantive research
signal:

• spotless formatting (mechanical, ./gradlew spotlessApply)
• removed duplicate ReactorAsyncResultExtension — eval renamed the
  package reactor.core.* → reactorcore.* but did not remove the
  master's reactor.core.* version, causing runtime AsyncResultExtension
  registration conflict across all JVM inst test shards. Fixing this
  preserved the package-rename observation (still visible in the PR diff)
  without letting a mechanical duplicate obscure it.
• dropped Schedulers.elastic() from scheduler test — the generated
  scheduler-arg list included Schedulers.elastic() which was deprecated
  in Reactor 3.4 and removed in 3.5+, blocking :check (via aggregate
  compile) and :latestDepTest (via direct API drift). Removed the
  elastic row while keeping parallel and single for coverage of
  scheduler-hop trace propagation.

Original: 2e1d2b98d7 → rebased: dbf3d6b2c7 → fixes: 3f485e671e

Where to focus review

The reference output should be compared against the master implementation
to surface how well the current skill draft handles Reactor's
context-propagation model. Known observations for review, ordered by
significance:

1. Reactor context-propagation classes dropped (PRIMARY finding). The
   eval regenerated only 2 of the 7 master instrumentation classes for
   this module. Dropped from the eval output:

   • ReactorContextBridge — the helper that reads a user-provided span
     from Reactor's own reactor.util.context.Context under the
     dd.span key, adapts it to a Datadog Context, and drives
     context-store-based propagation (capture on subscribe, restore on
     signal, transfer to optimized subscribers).
   • BlockingPublisherInstrumentation — advice for Mono.block() /
     Flux.blockFirst() / blockLast() that activates the captured
     context on the blocking-thread side of the chain.
   • ContextWritingSubscriberInstrumentation — advice for
     .contextWrite(...) subscribers to capture the user-provided
     context at subscribe time.
   • CorePublisherInstrumentation — advice at the base publisher
     interface that hands off context to downstream subscribers.
   • OptimizableOperatorInstrumentation — advice for Reactor's
     internal fused-operator path so the captured context survives
     Reactor's operator-chain optimization.

   These are the classes that implement Reactor's library-native
   context-map propagation mechanism (users pass a span via
   .contextWrite(ctx -> ctx.put("dd.span", span))). Without them,
   the observer-wrapping pattern the eval did produce (TracingCoreSubscriber,
   FluxInstrumentation, MonoInstrumentation) only handles the
   Cat-B-generic "capture-restore active context at subscribe" case —
   it does NOT handle the Reactor-specific "user placed a span in the
   library's context map" case that Spring WebFlux, Spring Kafka
   @KafkaListener suspend fun, resilience4j-reactor, reactor-netty,
   and other Reactor-adjacent modules rely on.

   Concrete failure this causes: spring-messaging-4.0's
   KafkaBatchListenerCoroutineTest times out with timed out waiting for async listener to start because the parent spring.consume
   span never inherits through the Reactor-mediated coroutine chain.
   Similar failures in spring-webflux-6.0:bootTest. See CI failures
   under test_inst: [*, 2/8] and test_inst_latest: [*, 4/6] +
   [*, 6/6].

   Skill-update prescription (candidate for skill(apm-integrations): additional rules from recent HTTP-category PR reviews #11927): add prescriptive
   Rule First PR - Implements the Opentracing API #2 guidance for reactive-library regens — when regenerating an
   existing Cat B reactive module, enumerate every *Instrumentation.java
   in master and preserve or replace each one. For libraries with a
   first-class in-library context concept (Reactor's
   reactor.util.context.Context, Kotlin's CoroutineContext, JAX-RS
   ContainerRequestContext, etc.), also preserve any *ContextBridge-
   style helper that reads a well-known key from that library-native
   context and populates the toolkit-native ContextStore. Regen must
   preserve verbatim or produce equivalent semantics.

2. Package rename (Rule First PR - Implements the Opentracing API #2, regen-preservation). The eval reorganized
   the package layout from datadog.trace.instrumentation.reactor.core
   (master) to datadog.trace.instrumentation.reactorcore (eval). The
   skill should preserve master's package layout on regen; it did not.
   Side-effect: breaks a fully-qualified class reference in
   graal-native-image-20.0's NativeImageGeneratorRunnerInstrumentation
   (line 158) that listed datadog.trace.instrumentation.reactor.core.ReactorAsyncResultExtension
   as build_time.

3. Version-sensitive tests in the wrong source set. The scheduler
   test that broke latestDepTest was generated in src/test/java/
   (which compiles against Reactor 3.1) rather than in a
   src/latestDepTest/ variant (which would compile against the latest
   published version). Master handles this by splitting Reactor
   version-sensitive tests into a dedicated latestDepTest source set
   (see master's src/latestDepTest/groovy/ReactorCoreTest.groovy).
   The skill draft does not currently emit multiple test source sets
   for API-drift-sensitive libraries; illustrating this gap is the
   underlying reason the Schedulers.elastic() failure occurred.

CI status

Local CI-equivalent (./gradlew :check :muzzle :latestDepTest --no-daemon):
all green after the three fixes above.

Remote CI is intentionally left red to preserve the primary research
signal (observation #1). The 12 failing shards
(test_inst: [17/21/25/tip, 2/8], test_inst_latest: [17/21/25/tip, 4/6],
test_inst_latest: [17/21/25/tip, 6/6]) map deterministically via
dd-trace-java's slot-hash to spring-messaging-4.0 and spring-webflux-6.0
tests that depend on ReactorContextBridge for coroutine / reactive
context propagation. Fixing them would require restoring the 5 dropped
master instrumentation classes, which would hide the finding.

Base: master @ 05671ce3b0

[datadog-datadog-us1-prod]

🎯 Code Coverage (details)
• Patch Coverage: 100.00%
• Overall Coverage: 57.17% (+0.00%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 3f485e6 | Docs | Datadog PR Page | Give us feedback!}

[dd-octo-sts]

🟢 Java Benchmark SLOs — All performance SLOs passed

Suite	Status
Startup	🟢 pass

SLO thresholds are defined here based on automatically generated metrics. A warning is raised when results are within 5% of the threshold.

PR vs. master results

Scenario	Candidate	master	Δ (95% CI of mean)
startup:insecure-bank:iast:Agent	14.06 s	13.97 s	[-0.2%; +1.5%] (no difference)
startup:insecure-bank:tracing:Agent	13.01 s	13.05 s	[-1.0%; +0.4%] (no difference)
startup:petclinic:appsec:Agent	16.95 s	16.76 s	[+0.2%; +2.0%] (maybe worse)
startup:petclinic:iast:Agent	16.92 s	16.96 s	[-1.1%; +0.5%] (no difference)
startup:petclinic:profiling:Agent	16.66 s	16.93 s	[-2.4%; -0.7%] (maybe better)
startup:petclinic:sca:Agent	16.96 s	16.82 s	[-0.1%; +1.7%] (no difference)
startup:petclinic:tracing:Agent	16.18 s	16.10 s	[-0.6%; +1.6%] (no difference)

Commit: 3f485e67 · CI Pipeline · Benchmarking Platform UI

---

Load and DaCapo benchmarks can be triggered manually in the GitLab pipeline. Results will appear in the Benchmarking Platform UI after completion.

[jordan-wong]

Rebased 2026-07-14 — branch force-pushed onto current master to fix the shared config-inversion-linter failure (same root cause as PR #11939).

Root cause (not a toolkit bug): the original eval branch inherited a stash-merge revert of #11852 during branch setup. See #11939 comment for the full diagnostic.

Fix: cherry-picked only the in-scope dd-java-agent/instrumentation/reactor-core-3.1/ paths from the original terminal SHA (2e1d2b98d7) onto fresh master (05671ce3b0). All reactor metadata entries were already present on master, so no metadata change is needed.

Result: 24 files changed (all under reactor-core-3.1/), no sibling contamination.

Note for reviewers: the eval reorganized the package layout from datadog.trace.instrumentation.reactor.core (master) to datadog.trace.instrumentation.reactorcore (eval output). This is a Rule #2 (regen-preservation) concern worth flagging — the master module's package layout should typically be preserved verbatim on regen.

Original commit: 2e1d2b98d7 → rebased: dbf3d6b2c7

The 10 CI failures beyond the shared metadata bug are being triaged separately.

[mcculls]

these are for various annotation driven tests

[mcculls]

Main thing it wasn't able to recreate is the hand-off via ReactorContextBridge - but TBH that is quite a leap for AI, and that only came about after several human-led revisions of the instrumentation.

It is using context stores and wrapping - but the places it is doing that could lead to increased memory use. Overall I wouldn't say this is the same quality - but I'd be interested to compare the traces it produces against the current approach.