[processor/dynamic_sampling] move buffered spans at decision time instead of copying

[MikeGoldsmith]

Description

Companion to #50026, applied to the decision path. assembleTrace deep-copied every buffered ResourceSpans into the output at decision time, and profiling showed that copy was ~52% of decision-path allocation bytes. The buffered spans are processor-owned (created fresh in ConsumeTraces) and the pending trace is discarded after the decision, so they can be moved instead. readIncomingSampling reads the same spans and already runs before assembly; a comment now pins that ordering and finishDecision nils the consumed slice.

Also fixes two benchmark issues found while measuring:

• The incoming-tracestate decide benchmark used an rv of 16 hex digits (the spec requires exactly 14), so it exercised the parse-failure path and understated real tracestate cost.
• BenchmarkDecide reused one pendingTrace across iterations, which breaks once spans are consumed; it now rebuilds the trace per iteration (untimed).

Benchmarks (Apple M4 Pro, mean of 5 runs, measured after the benchmark fixes so before/after are like-for-like):

BenchmarkDecide case	ns/op before	ns/op after	Δ	B/op before	B/op after	Δ	allocs/op before	allocs/op after	Δ
1span_catchall	1,931	1,436	-26%	1,272	648	-49%	24	13	-46%
10spans_catchall	4,523	3,102	-31%	7,392	3,320	-55%	105	49	-53%
10spans_catchall_incoming_tracestate	7,852	6,721	-14%	8,272	4,200	-49%	135	79	-41%
10spans_5ottl_rules	6,677	5,424	-19%	8,042	3,967	-51%	145	89	-39%
100spans_catchall	27,739	16,044	-42%	68,688	29,960	-56%	915	409	-55%
100spans_5ottl_rules	44,015	34,277	-22%	75,174	36,416	-52%	1,315	809	-38%
1000spans_catchall	264,149	150,837	-43%	680,791	296,364	-56%	9,015	4,009	-56%
10000spans_catchall	2,435,400	1,163,242	-52%	6,802,536	2,960,366	-56%	90,015	40,009	-56%
10000spans_5ottl_rules	4,084,604	2,673,498	-35%	7,444,723	3,602,399	-52%	130,018	80,011	-38%

Link to tracking issue

Refs #49311

Testing

• Full suite passes with the race detector; existing decide/eviction/late-span tests cover the consumed-spans semantics.
• Before/after benchmark comparison across all BenchmarkDecide cases (numbers above).

Documentation

Comments on assembleTrace and finishDecision documenting the move semantics and ordering constraint. Changelog entry included.

Authorship

• I, a human, wrote this pull request description myself.

[opentelemetry-pr-dashboard]

Pull request dashboard status

Merged · refreshed 2026-08-04 19:31 UTC

Status above doesn't look right?

• Anything look wrong? Report it with what you expected; it helps us improve the dashboard.