Add per-operation self-tuning dense-store sizing (SizingHint)

dd-trace-core/build.gradle

@@ -134,6 +134,9 @@ jmh {
   if (project.hasProperty('jmh.profilers')) {
     profilers = project.property('jmh.profilers').tokenize(',')
   }
+  if (project.hasProperty('jmh.jvmArgs')) {
+    jvmArgsAppend = project.property('jmh.jvmArgs').tokenize(' ')
+  }
   if (project.hasProperty('testJvm')) {
     def testJvmSpec = new TestJvmSpec(project)
     jvm = testJvmSpec.javaTestLauncher.map { it.executablePath.asFile.absolutePath }

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

@@ -42,6 +42,8 @@
 import datadog.trace.api.InstrumenterConfig;
 import datadog.trace.api.KnownTags;
 import datadog.trace.api.Pair;
+import datadog.trace.api.SizingHint;
+import datadog.trace.api.SizingHintTable;
 import datadog.trace.api.TagMap;
 import datadog.trace.api.TraceConfig;
 import datadog.trace.api.civisibility.config.BazelMode;
@@ -659,7 +661,8 @@ private CoreTracer(
 
     // Dense known-tag store (experimental, OFF by default): registering the KnownTagCodec 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
+    // 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")) {
       KnownTags.init();
@@ -2189,6 +2192,14 @@ protected static final DDSpanContext buildSpanContext(
         requestContextDataIast = builderRequestContextDataIast;
       }
 
+      // Per-type dense-store sizing: an entry (local-root) span carries the trace-metadata /
+      // enriching tags a child doesn't, so pick the lane by whether we have a local parent. A
+      // resolved hint sizes the span's TagMap and self-tunes on finish; null (no/unkeyable
+      // operation
+      // name) falls back to the generic default capacity.
+      final boolean entrySpan = !(resolvedParentSpanContext instanceof DDSpanContext);
+      final SizingHint sizingHint = SizingHintTable.hintFor(operationName, entrySpan);
+
       // some attributes are inherited from the parent
       context =
           new DDSpanContext(
@@ -2217,7 +2228,8 @@ protected static final DDSpanContext buildSpanContext(
               tracer.profilingContextIntegration,
               tracer.injectBaggageAsTags,
               tracer.injectLinksAsTags,
-              mergedTracerTagsNeedsIntercept ? null : mergedTracerTags);
+              mergedTracerTagsNeedsIntercept ? null : mergedTracerTags,
+              sizingHint);
 
       // 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`

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

@@ -157,6 +157,7 @@ public boolean isFinished() {
   private void finishAndAddToTrace(final long durationNano) {
     // ensure a min duration of 1
     if (DURATION_NANO_UPDATER.compareAndSet(this, 0, Math.max(1, durationNano))) {
+      context.recordDenseSize();
       setLongRunningVersion(-this.longRunningVersion);
       SpanWrapper wrapper = getWrapper();
       if (wrapper != null) {

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

@@ -12,6 +12,7 @@
 import datadog.trace.api.DDTraceId;
 import datadog.trace.api.Functions;
 import datadog.trace.api.ProcessTags;
+import datadog.trace.api.SizingHint;
 import datadog.trace.api.TagMap;
 import datadog.trace.api.cache.DDCache;
 import datadog.trace.api.cache.DDCaches;
@@ -21,6 +22,7 @@
 import datadog.trace.api.gateway.RequestContext;
 import datadog.trace.api.gateway.RequestContextSlot;
 import datadog.trace.api.internal.TraceSegment;
+import datadog.trace.api.internal.VisibleForTesting;
 import datadog.trace.api.sampling.PrioritySampling;
 import datadog.trace.api.sampling.SamplingMechanism;
 import datadog.trace.bootstrap.instrumentation.api.AgentSpanContext;
@@ -138,6 +140,12 @@ public class DDSpanContext
    */
   private final TagMap unsafeTags;
 
+  // Per-operation sizing hint (from SizingHintTable, keyed by operation name) this span was sized
+  // from, if any. Held so the span can feed its final dense-store size back on finish (recordSize)
+  // -- the self-tuning loop that lets the reused hint converge to the operation's real known-tag
+  // high-water mark. Null when unsized.
+  private final SizingHint sizingHint;
+
   /** The service name is required, otherwise the span are dropped by the agent */
   private volatile String serviceName;
 
@@ -198,6 +206,8 @@ public class DDSpanContext
    */
   private volatile Map<String, Object> metaStruct = EMPTY_META_STRUCT;
 
+  @Deprecated
+  @VisibleForTesting
   public DDSpanContext(
       final DDTraceId traceId,
       final long spanId,
@@ -244,9 +254,12 @@ public DDSpanContext(
         ProfilingContextIntegration.NoOp.INSTANCE,
         true,
         true,
+        null,
         null);
   }
 
+  @Deprecated
+  @VisibleForTesting
   public DDSpanContext(
       final DDTraceId traceId,
       final long spanId,
@@ -295,10 +308,13 @@ public DDSpanContext(
         ProfilingContextIntegration.NoOp.INSTANCE,
         injectBaggageAsTags,
         injectLinksAsTags,
+        null,
         null);
   }
 
   /** Back-compat ctor (no read-through parent); delegates with a null parent. */
+  @Deprecated
+  @VisibleForTesting
   public DDSpanContext(
       final DDTraceId traceId,
       final long spanId,
@@ -351,6 +367,7 @@ public DDSpanContext(
         profilingContextIntegration,
         injectBaggageAsTags,
         injectLinksAsTags,
+        null,
         null);
   }
 
@@ -380,7 +397,8 @@ public DDSpanContext(
       final ProfilingContextIntegration profilingContextIntegration,
       final boolean injectBaggageAsTags,
       final boolean injectLinksAsTags,
-      final TagMap readThroughParent) {
+      final TagMap readThroughParent,
+      final SizingHint sizingHint) {
 
     assert traceCollector != null;
     this.traceCollector = traceCollector;
@@ -412,7 +430,8 @@ public DDSpanContext(
     this.unsafeTags =
         readThroughParent != null
             ? TagMap.createFromParent(readThroughParent)
-            : TagMap.create(capacity);
+            : sizingHint != null ? TagMap.create(sizingHint) : TagMap.create(capacity);
+    this.sizingHint = sizingHint;
 
     // must set this before setting the service and resource names below
     this.profilingContextIntegration = profilingContextIntegration;
@@ -963,6 +982,21 @@ public void setTag(final String tag, final String value) {
     }
   }
 
+  /**
+   * Feeds this span's final dense-store size back into the sizingHint it was created from (if any).
+   * Called once at span finish; lets a reused hint self-tune to the operation's observed known-tag
+   * high-water mark, so subsequent spans of the operation are sized correctly.
+   */
+  void recordDenseSize() {
+    if (sizingHint != null) {
+      // Intentionally unsynchronized: recordSize is benign-racy by design (monotonic-max on a plain
+      // int), and a stale knownCount read only ever costs one extra growth on a later span. The
+      // lock
+      // would add nothing but contention on the finish path.
+      unsafeTags.recordSize(sizingHint);
+    }
+  }
+
   public void setTag(TagMap.EntryReader entry) {
     if (entry == null) {
       return;

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

@@ -90,6 +90,9 @@ public class DenseStoreAllocBenchmark {
 
   private String[] keys;
   private String[] values;
+  // Per-operation sizing hint seeded to this scenario's known-tag count -- what SizingHintTable
+  // supplies.
+  private SizingHint sizingHint;
 
   @Setup(Level.Trial)
   public void setup() {
@@ -108,8 +111,11 @@ public void setup() {
       this.keys[i] = i < knownCount ? KNOWN[i] : "custom.tag." + i;
       this.values[i] = "value-" + i;
     }
+    // Size the dense store to the known-tag count (the "end state" per-operation sizing).
+    this.sizingHint = new SizingHint("bench", 0, Math.max(knownCount, 1));
   }
 
+  /** Current: generic default dense capacity (KNOWN_INIT_CAP=12) -- over-provisions small types. */
   @Benchmark
   public TagMap buildMap() {
     TagMap m = TagMap.create(16);
@@ -119,6 +125,16 @@ public TagMap buildMap() {
     return m;
   }
 
+  /** End state: dense store sized per-operation via a SizingHint -- no over-provision. */
+  @Benchmark
+  public TagMap buildMapSized() {
+    TagMap m = TagMap.create(sizingHint);
+    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);

internal-api/src/jmh/java/datadog/trace/util/FlatHashtableBenchmark.java

@@ -0,0 +1,97 @@
+package datadog.trace.util;
+
+import java.util.HashMap;
+import java.util.Map;
+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.Measurement;
+import org.openjdk.jmh.annotations.Mode;
+import org.openjdk.jmh.annotations.OutputTimeUnit;
+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;
+
+/**
+ * Directional: is the {@link FlatHashtable} hit-path lookup (what a span pays per create, all hits
+ * after warmup) cheap vs a {@link HashMap}? Concrete-typed {@code static final} helper so the
+ * static-poly specialization is in play. One op = one pass over the op-name set. Single-threaded,
+ * short — a rough signal, not a verdict (real numbers on the box).
+ */
+@State(Scope.Thread)
+@BenchmarkMode(Mode.AverageTime)
+@OutputTimeUnit(TimeUnit.NANOSECONDS)
+@Warmup(iterations = 3, time = 1)
+@Measurement(iterations = 3, time = 1)
+@Fork(1)
+@Threads(1)
+public class FlatHashtableBenchmark {
+
+  static final class StrHelper extends FlatHashtable.StringHelper<String> {
+    @Override
+    public boolean matches(String key, String value) {
+      return key == value || key.equals(value);
+    }
+
+    @Override
+    public String create(String key) {
+      return key; // store the key itself as the (self-identifying) entry
+    }
+  }
+
+  private static final StrHelper HELPER = new StrHelper();
+
+  private static final String[] KEYS = {
+    "servlet.request",
+    "database.query",
+    "http.request",
+    "grpc.client",
+    "kafka.produce",
+    "kafka.consume",
+    "jdbc.query",
+    "spring.handler",
+    "servlet.forward",
+    "okhttp.request"
+  };
+
+  private String[] table;
+  private Map<String, String> map;
+
+  @Setup
+  public void setup() {
+    table = FlatHashtable.create(String.class, KEYS.length);
+    map = new HashMap<>(KEYS.length * 2);
+    for (String k : KEYS) {
+      FlatHashtable.getOrCreate(table, k, HELPER);
+      map.put(k, k);
+    }
+  }
+
+  /** FlatHashtable all-hit lookups (concrete helper → specialized). */
+  @Benchmark
+  public void flatGet(Blackhole bh) {
+    for (String k : KEYS) {
+      bh.consume(FlatHashtable.get(table, k, HELPER));
+    }
+  }
+
+  /** Steady-state span-create shape: get-then-getOrCreate, all hits. */
+  @Benchmark
+  public void flatGetOrCreate(Blackhole bh) {
+    for (String k : KEYS) {
+      bh.consume(FlatHashtable.getOrCreate(table, k, HELPER));
+    }
+  }
+
+  /** Baseline: HashMap lookups over the same keys. */
+  @Benchmark
+  public void hashMapGet(Blackhole bh) {
+    for (String k : KEYS) {
+      bh.consume(map.get(k));
+    }
+  }
+}

internal-api/src/main/java/datadog/trace/api/SizingHelper.java

@@ -0,0 +1,27 @@
+package datadog.trace.api;
+
+import datadog.trace.util.FlatHashtable;
+
+/**
+ * {@link FlatHashtable} policy for the per-operation {@link SizingHint} table: keys by operation
+ * name, entries are {@code SizingHint}s carrying that name plus its cached spread hash. Stateless —
+ * held by {@link SizingHintTable} as a concrete-typed {@code static final} singleton so {@code
+ * FlatHashtable.get}/{@code getOrCreate} specialize (devirtualize + inline) at the call site.
+ *
+ * <p>Extends {@link FlatHashtable.StringHelper}, which seals the spread {@code hash}; this class
+ * only supplies {@code matches} and {@code create}. Both use the inherited {@code hash} so the
+ * cached {@link SizingHint#labelHash} is always the same spread the probe used.
+ */
+final class SizingHelper extends FlatHashtable.StringHelper<SizingHint> {
+  @Override
+  public boolean matches(String key, SizingHint value) {
+    // int gate on the cached hash before equals; op-names are usually interned literals, so `==` is
+    // the common hit.
+    return value.labelHash == hash(key) && (key == value.label || key.equals(value.label));
+  }
+
+  @Override
+  public SizingHint create(String key) {
+    return new SizingHint(key, hash(key), SizingHintTable.SEED_SIZE);
+  }
+}

internal-api/src/main/java/datadog/trace/api/SizingHint.java

@@ -0,0 +1,49 @@
+package datadog.trace.api;
+
+/**
+ * Opaque per-operation dense-store sizing hint, and a self-contained {@link
+ * datadog.trace.util.FlatHashtable} slot: it carries everything the probe compares ({@link #label}
+ * + cached {@link #labelHash}) plus the tuned payload ({@link #size}). Holding key, hash, and value
+ * in ONE object behind ONE array slot is deliberate — entry publication is a single reference
+ * store, so a reader sees {@code null} or a complete entry (never a torn one), and the {@code
+ * final} identity fields are visible even under racy publication (JMM final-field guarantee). That
+ * sidesteps the memory-ordering / visibility problems parallel key/hash/value arrays would create,
+ * no volatile or atomics.
+ *
+ * <p>Opaque to everything outside {@code datadog.trace.api}: no public members. {@link TagMap}
+ * reads {@link #size} to size a fresh dense store and writes it back (monotonic-max) at a terminal
+ * point; {@code SizingHelper} mints and compares by {@link #label}/{@link #labelHash}. Callers only
+ * ever hold the reference.
+ *
+ * <p>{@link #labelHash} is supplied by the helper (a single spread source — {@code
+ * FlatHashtable.StringHelper.hash}) so the cached gate always matches the probe hash.
+ */
+public final class SizingHint {
+  // Identity: final => safely published under a racy single-reference store. `label` is the
+  // operation
+  // name (typically an interned literal, so the `==` fast-path usually hits). `labelHash` is the
+  // helper's spread hash, cached to gate `equals` with an int compare during probing.
+  final String label;
+  final int labelHash;
+
+  // Payload: the tuned dense-store size. Plain racy int, updated monotonic-max — a stale/lost read
+  // only mis-sizes an array (over/under-provision), never corrupts tag data, so no synchronization.
+  int size;
+
+  // When true, {@code size} is fixed and recordSize won't grow it. For the shared default /
+  // overflow
+  // hint (a HETEROGENEOUS catch-all for operation-less / over-budget spans): self-tuning it via
+  // monotonic-max would converge to the max across unlike sharers and over-provision the lean ones.
+  final boolean capped;
+
+  SizingHint(String label, int labelHash, int seedSize) {
+    this(label, labelHash, seedSize, false);
+  }
+
+  SizingHint(String label, int labelHash, int seedSize, boolean capped) {
+    this.label = label;
+    this.labelHash = labelHash;
+    this.size = seedSize;
+    this.capped = capped;
+  }
+}