[dougqh]

What Does This Do

Extracting a trace no longer builds a per-trace sampling state, whatever the headers. Three steps, one commit each (plus tests):

1. Shared states for simple decisions.
   • A decision with no ot member and no sampling rate is just a priority and a decision maker. Extracted priorities, manual keeps and forced keeps are almost always one of a few dozen such pairs.
   • States are immutable, so these now use preallocated shared states.
2. The tracestate lives on the propagation tags, not in the state.
   • Only extraction writes a trace's tracestate, before the trace is shared, so it needn't be published together with the decision.
   • A tracestate change only replaces the state when its ot member changes.
3. The inherited ot member lives on the trace.
   • The merge already builds a W3CTracestate holding the member unparsed. The trace now keeps that as its tracestate, and it parses the member on first read.
   • InheritedSamplingState becomes a shared marker that records only the decision and its pending checks.
   • The ot member is now read through PropagationTags.getOtelTraceState(SamplingState). A marker state throws if read directly, so a missed caller fails loudly instead of reading nothing.

Sampling states built per extraction (instrumented count):

scenario	before	after
Datadog-only	1	0
compound, tracestate without ot	2	0
compound sameVersion (our own injected headers, which carry ot)	2	0
compound withOtelMember	2	0

Motivation

TraceSegment work (APMLP-2049), phase 3: per-trace sampling state at extraction.

Allocation (PropagationScenariosBenchmark, JDK 17, -t 4, default sampler, -prof gc, B/op). These are 1-fork spot checks taken after each step; the same jar varies by about ±16 B from run to run.

step	leaf sameVersion	intermediate sameVersion
base (#13042)	1704.0 ± 0.1	3328.0 ± 0.2
shared states	1648.0 ± 0.1	3280.0 ± 0.1
tracestate on the tags	1648.0 ± 0.1	3264.0 ± 0.2
ot member on the trace	1608.0 ± 0.1	3208.0 ± 0.1

That's −96 and −120 B per extracted request. The withOtelMember arms move the same way but have ±20–90 B error.

No new per-trace cost: PTags stays at 88 B, because the tracestate field holds either the header value or the W3CTracestate. Shared inherited states are filled on first use. The cache of whether the ot member keeps its position, which W3C injection checks on every inject, moves to the W3CTracestate for inherited members.

Additional Notes

• Deliberately not done: replacing the preformatted dm and ksr values with primitives. Formatting ksr at export would allocate a string per trace, where the static cache allocates nothing for a single-rate service.
• Testing:
  • CompoundExtractionGoldenTest passes; full corpus: 0 of 10150 cases differ.
  • Propagation, sampling, OTLP and CoreSpanBuilderTest: 2992 pass.
  • Mutation checks on step 3: two were caught by existing tests. Two survived until the new tests in the last commit (a read under changed checks; locking the decision maker keeping pending checks).
  • dd-trace-core suite (4861 tests): only the known failures, the Docker-only TracerConnectionReliabilityTest and the order-dependent PendingTraceBufferTest flare dump.
• Throughput not measured; it's part of the stack's interleaved run.
• Rebase warning: Fix mismatched _dd.p.ksr under concurrent sampling decisions with different rates #12988 (merged to master) fixes the ksr cache race in PTagsFactory. This stack predates it, so a rebase onto master will conflict there.
• Stacked on Replace the per-trace sampling state lock with compare-and-set #13042.

🤖 Generated with Claude Code

[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	13.87 s	13.93 s	[-1.0%; +0.3%] (no difference)
startup:insecure-bank:tracing:Agent	12.85 s	12.97 s	[-1.6%; -0.1%] (maybe better)
startup:petclinic:appsec:Agent	17.25 s	17.04 s	[+0.5%; +2.0%] (maybe worse)
startup:petclinic:iast:Agent	17.06 s	16.98 s	[-0.3%; +1.3%] (no difference)
startup:petclinic:profiling:Agent	16.62 s	16.82 s	[-2.2%; -0.1%] (maybe better)
startup:petclinic:sca:Agent	17.09 s	16.99 s	[-0.2%; +1.4%] (no difference)
startup:petclinic:tracing:Agent	15.69 s	16.22 s	[-7.5%; +1.0%] (no difference)

Commit: 9d907af7 · 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.