perf(native): reduce processor flush overhead

Avoid allocating the active-span list until a native trace is actually being flushed. Complete-trace flushes can also reuse trace.started as the finished export list because all started spans are finished and the processor erases the trace by reassignment after export.

Partial flushes still build a finished-only array for the active-span case.

Author: bengl

packages/dd-trace/src/span_processor.js

@@ -244,17 +244,18 @@ class SpanProcessor {
 
   process (span) {
     const spanContext = span.context()
-    const active = []
     const trace = spanContext._trace
     const { flushMinSpans, DD_TRACE_ENABLED } = this._config
     const { started, finished } = trace
 
     if (trace.record === false) return
     if (DD_TRACE_ENABLED === false) {
-      this._erase(trace, active)
+      this._erase(trace, [])
       return
     }
-    if (started.length === finished.length || finished.length >= flushMinSpans) {
+    const allStartedFinished = started.length === finished.length
+    if (allStartedFinished || finished.length >= flushMinSpans) {
+      const active = []
       this.sample(span)
       this._gitMetadataTagger.tagGitMetadata(spanContext)
 
@@ -269,15 +270,15 @@ class SpanProcessor {
       // Pass raw spans to the native exporter; the WASM pipeline serializes
       // them. When native stats are enabled the concentrator handles stats
       // aggregation during flush_chunk.
-      const finishedSpansToExport = []
+      const finishedSpansToExport = allStartedFinished ? started : []
       const otelSemantics = this._config.DD_TRACE_OTEL_SEMANTICS_ENABLED
       let isFirstSpanInChunk = true
 
       for (const span of started) {
         if (span._duration === undefined) {
           active.push(span)
         } else {
-          finishedSpansToExport.push(span)
+          if (!allStartedFinished) finishedSpansToExport.push(span)
           const context = span.context()
 
           // OTLP trace metrics remain a JS-side stats feature. Build the same