[dougqh]

Stacked: master ← #13008 ← this PR ← #13009

What Does This Do

Local child spans no longer allocate their own PropagationTags. They create one lazily, only when a write targets the child itself: _dd.p.ts or _dd.p.debug set on that span.

Motivation

Every local child span built an empty PropagationTags through propagationTagsFactory.empty(), but nearly all trace-level state is read from the local root's instance: sampling, origin, force-keep, and inject. In a low-tags-per-span stress test, PropagationTags is about 10% of allocation.

Additional Notes

Output is unchanged by design. #11701 shared the root's instance with children, which also changed what children emit and mutate, and it failed system-tests (_dd.p.tid) in ways that were never explained. This PR doesn't share anything:

• A child holds no instance. Nothing on a child mutates the root's.
• OWN_PROPAGATION_TAGS serialization (OTLP, CI Visibility, LLMObs, file and list writers) builds a transient empty instance, the same way the constructor used to. _dd.p.tid and the propagation-disabled error tag come out identical.
• The agent mappers (v0.4, v0.5, v1) read ROOT_PROPAGATION_TAGS only on the first span of each chunk, which is unaffected.

ChildSpanPropagationTagsTest checks each child's serialized output against master's construction of an empty instance.

Known pre-existing behavior, kept as is: _dd.p.ts and _dd.p.debug set on a non-root span go to that span's own tags, so they are never injected and never reach the agent payload. That will be fixed in a follow-up PR with its own system-test run.

Allocation (TraceAssemblyBenchmark, JDK 17, @Threads(8), 2 forks × 5 iterations, -prof gc):

childCount	master B/op	branch B/op	Δ
1	2280.001 ± 0.001	2080.001 ± 0.001	−200 (−8.8%)
5	5176.222 ± 420.370	3888.002 ± 0.003	−1288 (−24.9%)
20	14684.005 ± 19.124	10672.006 ± 0.010	−4012 (−27.3%)

That's about 200 B per child span.

Throughput is pending. These runs were on a busy host, so throughput numbers are not meaningful yet. All pairs overlap within their error bars. A quiet run on JDK 8, 17 and 21 (TraceAssemblyBenchmark plus PropagationScenariosBenchmark, 5 forks) will be added before this is marked ready. The change adds a volatile read on each PropagationTags access and a CAS when a child first writes its own tags.

Contributes to APMLP-1515 / APMLP-2049 (TraceSegment).

🤖 Generated with Claude Code

[datadog-prod-us1-5]

🎯 Code Coverage (details)
• Patch Coverage: 97.06%
• Overall Coverage: 59.86% (+0.02%)

_{This comment will be updated automatically if new data arrives.
🔗 Commit SHA: 35909f3 | Docs | 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	14.76 s	14.69 s	[-0.5%; +1.5%] (no difference)
startup:insecure-bank:tracing:Agent	13.61 s	13.65 s	[-1.0%; +0.5%] (no difference)
startup:petclinic:appsec:Agent	17.09 s	16.98 s	[-0.1%; +1.3%] (no difference)
startup:petclinic:iast:Agent	17.00 s	17.04 s	[-1.0%; +0.6%] (no difference)
startup:petclinic:profiling:Agent	16.72 s	16.73 s	[-1.1%; +1.0%] (no difference)
startup:petclinic:sca:Agent	17.06 s	16.95 s	[-0.4%; +1.6%] (no difference)
startup:petclinic:tracing:Agent	16.23 s	16.00 s	[+0.4%; +2.4%] (maybe worse)

Commit: 35909f39 · 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.