Store known span tags densely in TagMap by tag-id

dd-trace-core/src/main/java/datadog/trace/core/CoreTracer.java

@@ -37,6 +37,7 @@
 import datadog.trace.api.EndpointTracker;
 import datadog.trace.api.IdGenerationStrategy;
 import datadog.trace.api.InstrumenterConfig;
+import datadog.trace.api.KnownTagIds;
 import datadog.trace.api.Pair;
 import datadog.trace.api.TagMap;
 import datadog.trace.api.TraceConfig;
@@ -653,6 +654,14 @@ private CoreTracer(
     // preload this enum to avoid triggering classloading on the hot path
     TraceCollector.PublishState.values();
 
+    // Dense known-tag store (experimental, OFF by default): registering the KnownTags resolver
+    // flips the dense store live so known tags store without a per-tag Entry. Gated by a system
+    // property for A/B benchmarking; when off, keyOf stays a no-op and tag storage is byte-identical
+    // to today. Promote to a Config flag if this becomes a permanent rollout.
+    if (Boolean.getBoolean("dd.trace.dense.tags.enabled")) {
+      KnownTagIds.init();
+    }
+
     if (reportInTracerFlare) {
       TracerFlare.addReporter(this);
     }
@@ -2195,13 +2204,27 @@ protected static final DDSpanContext buildSpanContext(
       // By setting the tags on the context we apply decorators to any tags that have been set via
       // the builder. This is the order that the tags were added previously, but maybe the `tags`
       // set in the builder should come last, so that they override other tags.
-      context.setAllTags(mergedTracerTags, mergedTracerTagsNeedsIntercept);
+      //
+      // mergedTracerTags is trace-level shared state and the precedence floor (everything below
+      // overrides it). When it carries no interceptable tags, attach it as a read-through PARENT
+      // (shared by reference, no per-span copy) instead of copying its entries into the span. When
+      // it does need interception, fall back to copying (the interceptor's per-span side-effects
+      // can't be shared by reference).
+      if (mergedTracerTagsNeedsIntercept) {
+        context.setAllTags(mergedTracerTags, true);
+      } else {
+        context.parentTags(mergedTracerTags);
+      }
       context.setAllTags(tagLedger);
       context.setAllTags(coreTags, coreTagsNeedsIntercept);
       context.setAllTags(rootSpanTags, rootSpanTagsNeedsIntercept);
       context.setAllTags(contextualTags);
-      // remove version here since will be done later on the postProcessor.
-      // it will allow knowing if it will be set manually or not
+      // Version is added later by the postProcessor (InternalTagsAdder), only if not already set
+      // during the request. Config version is kept out of the trace-level bundle (see
+      // withTracerTags), so this removal now only wipes a version set via the span builder —
+      // keeping
+      // the existing semantics where a builder-set version is replaced by the config version. Under
+      // read-through this is a cheap local removal (version isn't in the parent, so no tombstone).
       context.removeTag(Tags.VERSION);
       return context;
     }
@@ -2432,6 +2455,13 @@ static TagMap withTracerTags(
       Map<String, ?> userSpanTags, Config config, TraceConfig traceConfig) {
     final TagMap result = TagMap.create(userSpanTags.size() + 5);
     result.putAll(userSpanTags);
+    // Version is conditionally managed by InternalTagsAdder (added only when service == DD_SERVICE
+    // and not set during the request), so keep it OUT of the trace-level bundle. This matters under
+    // read-through: the bundle becomes a shared parent, and a per-span removeTag(VERSION) on a key
+    // that lived in the parent would mint a per-span tombstone. With version excluded here, the
+    // per-span removeTag (retained, to wipe a builder-set version) is a cheap local op, never a
+    // tombstone. Behavior is unchanged: version was applied-then-removed at build today.
+    result.remove(Tags.VERSION);
     if (null != config) { // static
       if (!config.getEnv().isEmpty()) {
         result.set("env", config.getEnv());

dd-trace-core/src/main/java/datadog/trace/core/DDSpanContext.java

@@ -1045,6 +1045,21 @@ void setAllTags(final TagMap map, boolean needsIntercept) {
     }
   }
 
+  /**
+   * Attaches {@code parent} as a read-through parent of this span's tags instead of copying its
+   * entries in (level-split phase 1). The parent must be frozen and free of interceptable tags —
+   * the caller gates on {@code !needsIntercept}, since read-through bypasses the per-span
+   * interceptor side-effects that {@link #setAllTags(TagMap, boolean)} applies.
+   */
+  void parentTags(final TagMap parent) {
+    if (parent == null || parent.isEmpty()) {
+      return;
+    }
+    synchronized (unsafeTags) {
+      unsafeTags.withParent(parent);
+    }
+  }
+
   void setAllTags(final TagMap.Ledger ledger) {
     if (ledger == null) {
       return;

internal-api/build.gradle.kts

@@ -270,6 +270,7 @@ dependencies {
   api("com.datadoghq:dd-javac-plugin-client:0.2.2")
 
   testImplementation("org.snakeyaml:snakeyaml-engine:2.9")
+  testImplementation("org.openjdk.jol:jol-core:0.17") // StringIndexFootprintTest object-layout measurement
   testImplementation(project(":utils:test-utils"))
   testImplementation(libs.bundles.junit5)
   testImplementation("org.junit.vintage:junit-vintage-engine:${libs.versions.junit5.get()}")

internal-api/src/jmh/java/datadog/trace/api/DenseStoreAllocBenchmark.java

@@ -0,0 +1,122 @@
+package datadog.trace.api;
+
+import datadog.trace.bootstrap.instrumentation.api.Tags;
+import java.util.concurrent.TimeUnit;
+import org.openjdk.jmh.annotations.Benchmark;
+import org.openjdk.jmh.annotations.BenchmarkMode;
+import org.openjdk.jmh.annotations.Fork;
+import org.openjdk.jmh.annotations.Level;
+import org.openjdk.jmh.annotations.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+import org.openjdk.jmh.annotations.Param;
+import org.openjdk.jmh.annotations.Scope;
+import org.openjdk.jmh.annotations.Setup;
+import org.openjdk.jmh.annotations.State;
+import org.openjdk.jmh.annotations.Threads;
+import org.openjdk.jmh.annotations.Warmup;
+import org.openjdk.jmh.infra.Blackhole;
+
+/**
+ * Deterministic allocation A/B for the dense known-tag store, using the REAL {@link KnownTagIds}
+ * resolver (a {@code StringIndex} probe + a constant-returning {@code switch} — allocation-free,
+ * exactly like production). An earlier synthetic prefix resolver allocated in {@code keyOf}
+ * (substring) and {@code nameOf} (concat), contaminating the dense arm; this measures the store,
+ * not the resolver.
+ *
+ * <p>Models how a real span's tags route: {@code today} = all custom (what ships now — every tag
+ * buckets, since nothing is registered as known), {@code dense} = the same tag count with a
+ * realistic fraction routed to the dense store (real known tag names) and the rest custom. Run with
+ * {@code -prof gc}; the {@code gc.alloc.rate.norm} (B/op) delta at the same {@code tagCount} is
+ * what enabling the dense store does to a real span's per-build allocation.
+ */
+@State(Scope.Benchmark)
+@BenchmarkMode(Mode.Throughput)
+@OutputTimeUnit(TimeUnit.SECONDS)
+@Warmup(iterations = 2, time = 2)
+@Measurement(iterations = 3, time = 2)
+@Fork(1)
+@Threads(1)
+public class DenseStoreAllocBenchmark {
+
+  // Real stored (dense-routed) tag names — a realistic web/db span's known set.
+  static final String[] KNOWN =
+      new String[] {
+        DDTags.BASE_SERVICE,
+        Tags.VERSION,
+        Tags.COMPONENT,
+        Tags.SPAN_KIND,
+        Tags.HTTP_METHOD,
+        Tags.HTTP_ROUTE,
+        Tags.DB_TYPE,
+        Tags.DB_INSTANCE,
+        Tags.PEER_HOSTNAME,
+        Tags.DB_USER,
+        DDTags.LANGUAGE_TAG_KEY,
+        Tags.PEER_PORT,
+      };
+
+  // today = all custom (all bucket, what ships now); dense = ~70% known + custom (a real span);
+  // allKnown = 100% known (the trace-tier read-through parent's shape — exercises lazy buckets).
+  @Param({"today", "dense", "allKnown"})
+  String scenario;
+
+  @Param({"7", "12"})
+  int tagCount;
+
+  private String[] keys;
+  private String[] values;
+
+  @Setup(Level.Trial)
+  public void setup() {
+    KnownTagIds.init(); // registers the real (allocation-free) resolver
+    int knownCount;
+    if ("allKnown".equals(scenario)) {
+      knownCount = tagCount; // 100% known (<= KNOWN.length)
+    } else if ("dense".equals(scenario)) {
+      knownCount = (tagCount * 7) / 10; // ~70% known + custom
+    } else {
+      knownCount = 0; // today: all custom (all bucket)
+    }
+    this.keys = new String[tagCount];
+    this.values = new String[tagCount];
+    for (int i = 0; i < tagCount; i++) {
+      this.keys[i] = i < knownCount ? KNOWN[i] : "custom.tag." + i;
+      this.values[i] = "value-" + i;
+    }
+  }
+
+  @Benchmark
+  public TagMap buildMap() {
+    TagMap m = TagMap.create(16);
+    for (int i = 0; i < tagCount; i++) {
+      m.set(keys[i], values[i]);
+    }
+    return m;
+  }
+
+  @Benchmark
+  public void buildAndSerialize(Blackhole bh) {
+    TagMap m = TagMap.create(16);
+    for (int i = 0; i < tagCount; i++) {
+      m.set(keys[i], values[i]);
+    }
+    // forEach: the alloc-free flyweight emit for dense
+    m.forEach(reader -> bh.consume(reader.objectValue()));
+    bh.consume(m);
+  }
+
+  @Benchmark
+  public void buildAndSerializeViaIterator(Blackhole bh) {
+    TagMap m = TagMap.create(16);
+    for (int i = 0; i < tagCount; i++) {
+      m.set(keys[i], values[i]);
+    }
+    // models the REAL serializer's count pre-pass (TraceMapperV0_4:95): the EntryReader iterator
+    // materializes an Entry per dense tag -> should erase the dense alloc win.
+    for (TagMap.EntryReader reader : m) {
+      bh.consume(reader.objectValue());
+    }
+    bh.consume(m);
+  }
+}