Use Set hash in Distinct (2nd attempt)

CHANGELOG.md

@@ -12,6 +12,7 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 - Add `WithInstrumentationAttributeSet` option to `go.opentelemetry.io/otel/log`, `go.opentelemetry.io/otel/metric`, and `go.opentelemetry.io/otel/trace` packages.
   This provides a concurrent-safe and performant alternative to `WithInstrumentationAttributes` by accepting a pre-constructed `attribute.Set`. (#7287)
+- Greatly reduce the cost of recording metrics in `go.opentelemetry.io/otel/sdk/metric` using hashing for map keys. (#7175)
 
 ### Fixed
 
@@ -29,6 +30,7 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 - `WithInstrumentationAttributes` in `go.opentelemetry.io/otel/trace` synchronously de-duplicates the passed attributes instead of delegating it to the returned `TracerOption`. (#7266)
 - `WithInstrumentationAttributes` in `go.opentelemetry.io/otel/meter` synchronously de-duplicates the passed attributes instead of delegating it to the returned `MeterOption`. (#7266)
 - `WithInstrumentationAttributes` in `go.opentelemetry.io/otel/log` synchronously de-duplicates the passed attributes instead of delegating it to the returned `LoggerOption`. (#7266)
+- `Distinct` in `go.opentelemetry.io/otel/attribute` is no longer guaranteed to uniquely identify an attribute set. Collisions between `Distinct` values for different Sets are possible with extremely high cardinality (billions of series per instrument), but are highly unlikely. (#7175)
 
 <!-- Released section -->
 <!-- Don't change this section unless doing release -->

Makefile

@@ -146,11 +146,12 @@ build-tests/%:
 
 # Tests
 
-TEST_TARGETS := test-default test-bench test-short test-verbose test-race test-concurrent-safe
+TEST_TARGETS := test-default test-bench test-short test-verbose test-race test-concurrent-safe test-fuzz
 .PHONY: $(TEST_TARGETS) test
 test-default test-race: ARGS=-race
 test-bench:   ARGS=-run=xxxxxMatchNothingxxxxx -test.benchtime=1ms -bench=.
 test-short:   ARGS=-short
+test-fuzz:    ARGS=-fuzztime=10s -fuzz
 test-verbose: ARGS=-v -race
 test-concurrent-safe: ARGS=-run=ConcurrentSafe -count=100 -race
 test-concurrent-safe: TIMEOUT=120

attribute/hash.go

@@ -0,0 +1,92 @@
+// Copyright The OpenTelemetry Authors
+// SPDX-License-Identifier: Apache-2.0
+
+package attribute // import "go.opentelemetry.io/otel/attribute"
+
+import (
+	"fmt"
+	"reflect"
+
+	"go.opentelemetry.io/otel/attribute/internal/fnv"
+)
+
+// Type identifiers. These identifiers are hashed before the value of the
+// corresponding type. This is done to distinguish values that are hashed with
+// the same value representation (e.g. `int64(1)` and `true`, []int64{0} and
+// int64(0)).
+//
+// These are all 8 byte length strings converted to a uint64 representation. A
+// uint64 is used instead of the string directly as an optimization, it avoids
+// the for loop in [fnv] which adds minor overhead.
+const (
+	boolID         uint64 = 7953749933313450591 // "_boolean" (little endian)
+	int64ID        uint64 = 7592915492740740150 // "64_bit_i" (little endian)
+	float64ID      uint64 = 7376742710626956342 // "64_bit_f" (little endian)
+	stringID       uint64 = 6874584755375207263 // "_string_" (little endian)
+	boolSliceID    uint64 = 6875993255270243167 // "_[]bool_" (little endian)
+	int64SliceID   uint64 = 3762322556277578591 // "_[]int64" (little endian)
+	float64SliceID uint64 = 7308324551835016539 // "[]double" (little endian)
+	stringSliceID  uint64 = 7453010373645655387 // "[]string" (little endian)
+)
+
+// hashKVs returns a new FNV-1a hash of kvs.
+func hashKVs(kvs []KeyValue) fnv.Hash {
+	h := fnv.New()
+	for _, kv := range kvs {
+		h = hashKV(h, kv)
+	}
+	return h
+}
+
+// hashKV returns the FNV-1a hash of kv with h as the base.
+func hashKV(h fnv.Hash, kv KeyValue) fnv.Hash {
+	h = h.String(string(kv.Key))
+
+	switch kv.Value.Type() {
+	case BOOL:
+		h = h.Uint64(boolID)
+		h = h.Uint64(kv.Value.numeric)
+	case INT64:
+		h = h.Uint64(int64ID)
+		h = h.Uint64(kv.Value.numeric)
+	case FLOAT64:
+		h = h.Uint64(float64ID)
+		// Assumes numeric stored with math.Float64bits.
+		h = h.Uint64(kv.Value.numeric)
+	case STRING:
+		h = h.Uint64(stringID)
+		h = h.String(kv.Value.stringly)
+	case BOOLSLICE:
+		h = h.Uint64(boolSliceID)
+		rv := reflect.ValueOf(kv.Value.slice)
+		for i := 0; i < rv.Len(); i++ {
+			h = h.Bool(rv.Index(i).Bool())
+		}
+	case INT64SLICE:
+		h = h.Uint64(int64SliceID)
+		rv := reflect.ValueOf(kv.Value.slice)
+		for i := 0; i < rv.Len(); i++ {
+			h = h.Int64(rv.Index(i).Int())
+		}
+	case FLOAT64SLICE:
+		h = h.Uint64(float64SliceID)
+		rv := reflect.ValueOf(kv.Value.slice)
+		for i := 0; i < rv.Len(); i++ {
+			h = h.Float64(rv.Index(i).Float())
+		}
+	case STRINGSLICE:
+		h = h.Uint64(stringSliceID)
+		rv := reflect.ValueOf(kv.Value.slice)
+		for i := 0; i < rv.Len(); i++ {
+			h = h.String(rv.Index(i).String())
+		}
+	case INVALID:
+	default:
+		// Logging is an alternative, but using the internal logger here
+		// causes an import cycle so it is not done.
+		v := kv.Value.AsInterface()
+		msg := fmt.Sprintf("unknown value type: %[1]v (%[1]T)", v)
+		panic(msg)
+	}
+	return h
+}