fix: Fix slicelabels corruption when used with proto decoding

Alternative to https://github.com/prometheus/prometheus/pull/16957/

Signed-off-by: bwplotka <bwplotka@gmail.com>

# Conflicts:
#	model/textparse/protobufparse.go
#	model/textparse/protobufparse_test.go

.github/workflows/ci.yml

@@ -40,7 +40,7 @@ jobs:
       - uses: prometheus/promci@443c7fc2397e946bc9f5029e313a9c3441b9b86d # v0.4.7
       - uses: ./.github/promci/actions/setup_environment
       - run: go test --tags=dedupelabels ./...
-      - run: go test --tags=slicelabels -race ./cmd/prometheus ./prompb/io/prometheus/client
+      - run: go test --tags=slicelabels -race ./cmd/prometheus ./model/textparse ./prompb/...
       - run: go test --tags=forcedirectio -race ./tsdb/
       - run: GOARCH=386 go test ./...
       - uses: ./.github/promci/actions/check_proto

model/labels/labels_dedupelabels.go

@@ -775,6 +775,13 @@ func (b *ScratchBuilder) SetSymbolTable(s *SymbolTable) {
 	b.syms = s
 }
 
+// SetUnsafeAdd allows turning on/off the assumptions that added strings are unsafe
+// for reuse. ScratchBuilder implementations that do reuse strings, must clone
+// the strings.
+//
+// Dedupelabels implementation, does not reuse added strings, so this operation is noop.
+func (ScratchBuilder) SetUnsafeAdd(bool) {}
+
 func (b *ScratchBuilder) Reset() {
 	b.add = b.add[:0]
 	b.output = EmptyLabels()
@@ -786,12 +793,6 @@ func (b *ScratchBuilder) Add(name, value string) {
 	b.add = append(b.add, Label{Name: name, Value: value})
 }
 
-// UnsafeAddBytes adds a name/value pair, using []byte instead of string to reduce memory allocations.
-// The values must remain live until Labels() is called.
-func (b *ScratchBuilder) UnsafeAddBytes(name, value []byte) {
-	b.add = append(b.add, Label{Name: yoloString(name), Value: yoloString(value)})
-}
-
 // Sort the labels added so far by name.
 func (b *ScratchBuilder) Sort() {
 	slices.SortFunc(b.add, func(a, b Label) int { return strings.Compare(a.Name, b.Name) })

model/labels/labels_slicelabels.go

@@ -19,6 +19,7 @@ import (
 	"bytes"
 	"slices"
 	"strings"
+	"unique"
 	"unsafe"
 
 	"github.com/cespare/xxhash/v2"
@@ -437,7 +438,8 @@ func (b *Builder) Labels() Labels {
 
 // ScratchBuilder allows efficient construction of a Labels from scratch.
 type ScratchBuilder struct {
-	add Labels
+	add       Labels
+	unsafeAdd bool
 }
 
 // SymbolTable is no-op, just for api parity with dedupelabels.
@@ -466,6 +468,15 @@ func (*ScratchBuilder) SetSymbolTable(*SymbolTable) {
 	// no-op
 }
 
+// SetUnsafeAdd allows turning on/off the assumptions that added strings are unsafe
+// for reuse. ScratchBuilder implementations that do reuse strings, must clone
+// the strings.
+//
+// SliceLabels will clone all added strings when this option is true.
+func (b *ScratchBuilder) SetUnsafeAdd(unsafeAdd bool) {
+	b.unsafeAdd = unsafeAdd
+}
+
 func (b *ScratchBuilder) Reset() {
 	b.add = b.add[:0]
 }
@@ -473,16 +484,15 @@ func (b *ScratchBuilder) Reset() {
 // Add a name/value pair.
 // Note if you Add the same name twice you will get a duplicate label, which is invalid.
 func (b *ScratchBuilder) Add(name, value string) {
+	if b.unsafeAdd {
+		// Underlying label structure for slicelabels shares memory, so we need to
+		// copy it if the input is unsafe.
+		name = unique.Make(name).Value()
+		value = unique.Make(value).Value()
+	}
 	b.add = append(b.add, Label{Name: name, Value: value})
 }
 
-// UnsafeAddBytes adds a name/value pair, using []byte instead of string.
-// The default version of this function is unsafe, hence the name.
-// This version is safe - it copies the strings immediately - but we keep the same name so everything compiles.
-func (b *ScratchBuilder) UnsafeAddBytes(name, value []byte) {
-	b.add = append(b.add, Label{Name: string(name), Value: string(value)})
-}
-
 // Sort the labels added so far by name.
 func (b *ScratchBuilder) Sort() {
 	slices.SortFunc(b.add, func(a, b Label) int { return strings.Compare(a.Name, b.Name) })

model/labels/labels_slicelabels_test.go

@@ -15,6 +15,12 @@
 
 package labels
 
+import (
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
 var expectedSizeOfLabels = []uint64{ // Values must line up with testCaseLabels.
 	72,
 	0,
@@ -25,3 +31,55 @@ var expectedSizeOfLabels = []uint64{ // Values must line up with testCaseLabels.
 }
 
 var expectedByteSize = expectedSizeOfLabels // They are identical
+
+func TestScratchBuilderAdd_Strings(t *testing.T) {
+	t.Run("safe", func(t *testing.T) {
+		n := []byte("__name__")
+		v := []byte("metric1")
+
+		l := NewScratchBuilder(0)
+		l.Add(yoloString(n), yoloString(v))
+		ret := l.Labels()
+
+		// For slicelabels, in default mode strings are reused, so modifying the
+		// intput will cause `ret` labels to change too.
+		n[1] = byte('?')
+		v[2] = byte('?')
+
+		require.Empty(t, ret.Get("__name__"))
+		require.Equal(t, "me?ric1", ret.Get("_?name__"))
+	})
+	t.Run("unsafe", func(t *testing.T) {
+		n := []byte("__name__")
+		v := []byte("metric1")
+
+		l := NewScratchBuilder(0)
+		l.SetUnsafeAdd(true)
+		l.Add(yoloString(n), yoloString(v))
+		ret := l.Labels()
+
+		// Changing input strings should be now safe, because we marked adds as unsafe.
+		n[1] = byte('?')
+		v[2] = byte('?')
+
+		require.Equal(t, "metric1", ret.Get("__name__"))
+	})
+}
+
+/*
+	export bench=unsafe && go test -tags=slicelabels \
+	  -run '^$' -bench '^BenchmarkScratchBuilderUnsafeAdd' \
+	  -benchtime 5s -count 6 -cpu 2 -timeout 999m \
+	  | tee ${bench}.txt
+*/
+func BenchmarkScratchBuilderUnsafeAdd(b *testing.B) {
+	l := NewScratchBuilder(0)
+	l.SetUnsafeAdd(true)
+
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		l.Add("__name__", "metric1")
+		l.add = l.add[:0] // Reset slice so add can be repeated without side effects.
+	}
+}

model/labels/labels_stringlabels.go

@@ -618,12 +618,6 @@ func (b *ScratchBuilder) Add(name, value string) {
 	b.add = append(b.add, Label{Name: name, Value: value})
 }
 
-// UnsafeAddBytes adds a name/value pair using []byte instead of string to reduce memory allocations.
-// The values must remain live until Labels() is called.
-func (b *ScratchBuilder) UnsafeAddBytes(name, value []byte) {
-	b.add = append(b.add, Label{Name: yoloString(name), Value: yoloString(value)})
-}
-
 // Sort the labels added so far by name.
 func (b *ScratchBuilder) Sort() {
 	slices.SortFunc(b.add, func(a, b Label) int { return strings.Compare(a.Name, b.Name) })
@@ -680,6 +674,13 @@ func (*ScratchBuilder) SetSymbolTable(*SymbolTable) {
 	// no-op
 }
 
+// SetUnsafeAdd allows turning on/off the assumptions that added strings are unsafe
+// for reuse. ScratchBuilder implementations that do reuse strings, must clone
+// the strings.
+//
+// Stringlabels implementation, does not reuse added strings, so this operation is noop.
+func (ScratchBuilder) SetUnsafeAdd(bool) {}
+
 // SizeOfLabels returns the approximate space required for n copies of a label.
 func SizeOfLabels(name, value string, n uint64) uint64 {
 	return uint64(labelSize(&Label{Name: name, Value: value})) * n

model/textparse/protobufparse.go

@@ -94,10 +94,12 @@ type ProtobufParser struct {
 
 // NewProtobufParser returns a parser for the payload in the byte slice.
 func NewProtobufParser(b []byte, parseClassicHistograms, convertClassicHistogramsToNHCB, enableTypeAndUnitLabels bool, st *labels.SymbolTable) Parser {
+	builder := labels.NewScratchBuilderWithSymbolTable(st, 16)
+	builder.SetUnsafeAdd(true)
 	return &ProtobufParser{
 		dec:        dto.NewMetricStreamingDecoder(b),
 		entryBytes: &bytes.Buffer{},
-		builder:    labels.NewScratchBuilderWithSymbolTable(st, 16), // TODO(bwplotka): Try base builder.
+		builder:    builder,
 
 		state:                          EntryInvalid,
 		parseClassicHistograms:         parseClassicHistograms,
@@ -622,10 +624,7 @@ func (p *ProtobufParser) onSeriesOrHistogramUpdate() error {
 			Unit: p.dec.GetUnit(),
 		}
 		m.AddToLabels(&p.builder)
-		if err := p.dec.Label(schema.IgnoreOverriddenMetadataLabelsScratchBuilder{
+		if err := p.dec.Label(m.NewIgnoreOverriddenMetadataLabelScratchBuilder(&p.builder)); err != nil {
-			Overwrite:      m,
-			ScratchBuilder: &p.builder,
-		}); err != nil {
 			return err
 		}
 	} else {

model/textparse/protobufparse_test.go

@@ -16,10 +16,15 @@ package textparse
 import (
 	"bytes"
 	"encoding/binary"
+	"errors"
 	"fmt"
+	"io"
+	"math/rand"
+	"strings"
 	"testing"
 
 	"github.com/gogo/protobuf/proto"
+	"github.com/gogo/protobuf/types"
 	"github.com/prometheus/common/model"
 	"github.com/stretchr/testify/require"
 
@@ -27,6 +32,7 @@ import (
 	"github.com/prometheus/prometheus/model/histogram"
 	"github.com/prometheus/prometheus/model/labels"
 	dto "github.com/prometheus/prometheus/prompb/io/prometheus/client"
+	"github.com/prometheus/prometheus/util/pool"
 )
 
 func metricFamiliesToProtobuf(t testing.TB, testMetricFamilies []string) *bytes.Buffer {
@@ -4349,3 +4355,187 @@ metric: <
 		})
 	}
 }
+
+func FuzzProtobufParser_Labels(f *testing.F) {
+	// Add to the "seed corpus" the values that are known to reproduce issues
+	// which this test has found in the past. These cases run during regular
+	// testing, as well as the first step of the fuzzing process.
+	f.Add(true, true, int64(123))
+	f.Add(true, false, int64(129))
+	f.Add(false, true, int64(159))
+	f.Add(false, true, int64(-127))
+	f.Fuzz(func(
+		t *testing.T,
+		parseClassicHistogram bool,
+		enableTypeAndUnitLabels bool,
+		randSeed int64,
+	) {
+		var (
+			r              = rand.New(rand.NewSource(randSeed))
+			buffers        = pool.New(1+r.Intn(128), 128+r.Intn(1024), 2, func(sz int) interface{} { return make([]byte, 0, sz) })
+			lastScrapeSize = 0
+			observedLabels []labels.Labels
+			st             = labels.NewSymbolTable()
+		)
+
+		for i := 0; i < 20; i++ { // run multiple iterations to encounter memory corruptions
+			// Get buffer from pool like in scrape.go
+			b := buffers.Get(lastScrapeSize).([]byte)
+			buf := bytes.NewBuffer(b)
+
+			// Generate some scraped data to parse
+			mf := generateFuzzMetricFamily(r)
+			protoBuf, err := proto.Marshal(mf)
+			require.NoError(t, err)
+			sizeBuf := make([]byte, binary.MaxVarintLen32)
+			sizeBufSize := binary.PutUvarint(sizeBuf, uint64(len(protoBuf)))
+			buf.Write(sizeBuf[:sizeBufSize])
+			buf.Write(protoBuf)
+
+			// Use protobuf parser to parse like in real usage
+			b = buf.Bytes()
+			p := NewProtobufParser(b, parseClassicHistogram, false, enableTypeAndUnitLabels, st)
+
+			for {
+				entry, err := p.Next()
+				if errors.Is(err, io.EOF) {
+					break
+				}
+				require.NoError(t, err)
+				switch entry {
+				case EntryHelp:
+					name, help := p.Help()
+					require.Equal(t, mf.Name, string(name))
+					require.Equal(t, mf.Help, string(help))
+				case EntryType:
+					name, _ := p.Type()
+					require.Equal(t, mf.Name, string(name))
+				case EntryUnit:
+					name, unit := p.Unit()
+					require.Equal(t, mf.Name, string(name))
+					require.Equal(t, mf.Unit, string(unit))
+				case EntrySeries, EntryHistogram:
+					var lbs labels.Labels
+					p.Labels(&lbs)
+					observedLabels = append(observedLabels, lbs)
+				}
+
+				// Get labels from exemplars
+				for {
+					var e exemplar.Exemplar
+					if !p.Exemplar(&e) {
+						break
+					}
+					observedLabels = append(observedLabels, e.Labels)
+				}
+			}
+
+			// Validate all labels seen so far remain valid. This can find memory corruption issues.
+			for _, l := range observedLabels {
+				require.True(t, l.IsValid(model.LegacyValidation), "encountered corrupted labels: %v", l)
+			}
+
+			lastScrapeSize = len(b)
+			buffers.Put(b)
+		}
+	})
+}
+
+func generateFuzzMetricFamily(
+	r *rand.Rand,
+) *dto.MetricFamily {
+	unit := generateValidLabelName(r)
+	metricName := fmt.Sprintf("%s_%s", generateValidMetricName(r), unit)
+	metricTypeProto := dto.MetricType(r.Intn(len(dto.MetricType_name)))
+	metricFamily := &dto.MetricFamily{
+		Name: metricName,
+		Help: generateHelp(r),
+		Type: metricTypeProto,
+		Unit: unit,
+	}
+	metricsCount := r.Intn(20)
+	for i := 0; i < metricsCount; i++ {
+		metric := dto.Metric{
+			Label: generateFuzzLabels(r),
+		}
+		switch metricTypeProto {
+		case dto.MetricType_GAUGE:
+			metric.Gauge = &dto.Gauge{Value: r.Float64()}
+		case dto.MetricType_COUNTER:
+			metric.Counter = &dto.Counter{Value: r.Float64()}
+		case dto.MetricType_SUMMARY:
+			metric.Summary = &dto.Summary{Quantile: []dto.Quantile{{Quantile: 0.5, Value: r.Float64()}}}
+		case dto.MetricType_HISTOGRAM:
+			metric.Histogram = &dto.Histogram{Exemplars: generateExemplars(r)}
+		}
+		metricFamily.Metric = append(metricFamily.Metric, metric)
+	}
+	return metricFamily
+}
+
+func generateExemplars(r *rand.Rand) []*dto.Exemplar {
+	exemplarsCount := r.Intn(5)
+	exemplars := make([]*dto.Exemplar, 0, exemplarsCount)
+	for i := 0; i < exemplarsCount; i++ {
+		exemplars = append(exemplars, &dto.Exemplar{
+			Label: generateFuzzLabels(r),
+			Value: r.Float64(),
+			Timestamp: &types.Timestamp{
+				Seconds: int64(r.Intn(1000000000)),
+				Nanos:   int32(r.Intn(1000000000)),
+			},
+		})
+	}
+	return exemplars
+}
+
+func generateFuzzLabels(r *rand.Rand) []dto.LabelPair {
+	labelsCount := r.Intn(10)
+	ls := make([]dto.LabelPair, 0, labelsCount)
+	for i := 0; i < labelsCount; i++ {
+		ls = append(ls, dto.LabelPair{
+			Name:  generateValidLabelName(r),
+			Value: generateValidLabelName(r),
+		})
+	}
+	return ls
+}
+
+func generateHelp(r *rand.Rand) string {
+	result := make([]string, 1+r.Intn(20))
+	for i := 0; i < len(result); i++ {
+		result[i] = generateValidLabelName(r)
+	}
+	return strings.Join(result, "_")
+}
+
+func generateValidLabelName(r *rand.Rand) string {
+	return generateString(r, validFirstRunes, validLabelNameRunes)
+}
+
+func generateValidMetricName(r *rand.Rand) string {
+	return generateString(r, validFirstRunes, validMetricNameRunes)
+}
+
+func generateString(r *rand.Rand, firstRunes, restRunes []rune) string {
+	result := make([]rune, 1+r.Intn(20))
+	for i := range result {
+		if i == 0 {
+			result[i] = firstRunes[r.Intn(len(firstRunes))]
+		} else {
+			result[i] = restRunes[r.Intn(len(restRunes))]
+		}
+	}
+	return string(result)
+}
+
+var (
+	validMetricNameRunes = []rune{
+		'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z',
+		'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M', 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z',
+		'0', '1', '2', '3', '4', '5', '6', '7', '8', '9',
+		'_', ':',
+	}
+	validLabelNameRunes = validMetricNameRunes[:len(validMetricNameRunes)-1] // skip the colon
+	validFirstRunes     = validMetricNameRunes[:52]                          // only the letters
+)

prompb/io/prometheus/client/decoder.go

@@ -154,17 +154,22 @@ func (*MetricStreamingDecoder) GetLabel() {
 	panic("don't use GetLabel, use Label instead")
 }
 
-type scratchBuilder interface {
+// unsafeLabelAdder is an interface that expects unsafe label adds.
+// Typically, this means labels.ScratchBuilder with SetUnsafeAdd set to true.
+type unsafeLabelAdder interface {
 	Add(name, value string)
-	UnsafeAddBytes(name, value []byte)
 }
 
-// Label parses labels into labels scratch builder. Metric name is missing
+// Label parses labels into unsafeLabelAdder. Metric name is missing
 // given the protobuf metric model and has to be deduced from the metric family name.
-// TODO: The method name intentionally hide MetricStreamingDecoder.Metric.Label
+//
+// TODO: The Label method name intentionally hide MetricStreamingDecoder.Metric.Label
 // field to avoid direct use (it's not parsed). In future generator will generate
 // structs tailored for streaming decoding.
-func (m *MetricStreamingDecoder) Label(b scratchBuilder) error {
+//
+// Unsafe in this context means that bytes and strings are reused across iterations.
+// They are live only until the next NextMetric() or NextMetricFamily() call.
+func (m *MetricStreamingDecoder) Label(b unsafeLabelAdder) error {
 	for _, l := range m.labels {
 		if err := parseLabel(m.mData[l.start:l.end], b); err != nil {
 			return err
@@ -173,11 +178,9 @@ func (m *MetricStreamingDecoder) Label(b scratchBuilder) error {
 	return nil
 }
 
-// parseLabel is essentially LabelPair.Unmarshal but directly adding into scratch builder
+// parseLabel is essentially LabelPair.Unmarshal but directly adding into unsafeLabelAdder.
-// via UnsafeAddBytes method to reuse strings.
+func parseLabel(dAtA []byte, b unsafeLabelAdder) error {
-func parseLabel(dAtA []byte, b scratchBuilder) error {
+	var unsafeName, unsafeValue string
-	var name, value []byte
-	var unsafeName string
 	l := len(dAtA)
 	iNdEx := 0
 	for iNdEx < l {
@@ -236,8 +239,7 @@ func parseLabel(dAtA []byte, b scratchBuilder) error {
 			if postIndex > l {
 				return io.ErrUnexpectedEOF
 			}
-			name = dAtA[iNdEx:postIndex]
+			unsafeName = yoloString(dAtA[iNdEx:postIndex])
-			unsafeName = yoloString(name)
 			if !model.UTF8Validation.IsValidLabelName(unsafeName) {
 				return fmt.Errorf("invalid label name: %s", unsafeName)
 			}
@@ -272,9 +274,9 @@ func parseLabel(dAtA []byte, b scratchBuilder) error {
 			if postIndex > l {
 				return io.ErrUnexpectedEOF
 			}
-			value = dAtA[iNdEx:postIndex]
+			unsafeValue = yoloString(dAtA[iNdEx:postIndex])
-			if !utf8.ValidString(yoloString(value)) {
+			if !utf8.ValidString(unsafeValue) {
-				return fmt.Errorf("invalid label value: %s", value)
+				return fmt.Errorf("invalid label value: %s", unsafeValue)
 			}
 			iNdEx = postIndex
 		default:
@@ -295,7 +297,7 @@ func parseLabel(dAtA []byte, b scratchBuilder) error {
 	if iNdEx > l {
 		return io.ErrUnexpectedEOF
 	}
-	b.UnsafeAddBytes(name, value)
+	b.Add(unsafeName, unsafeValue)
 	return nil
 }
 

prompb/io/prometheus/client/decoder_test.go

@@ -165,6 +165,8 @@ func TestMetricStreamingDecoder(t *testing.T) {
 
 		require.Equal(t, 1.546544e+06, d.Metric.GetCounter().GetValue())
 		b := labels.NewScratchBuilder(0)
+		b.SetUnsafeAdd(true)
+
 		require.NoError(t, d.Label(&b))
 		require.Equal(t, `{}`, b.Labels().String())
 	}
@@ -174,11 +176,14 @@ func TestMetricStreamingDecoder(t *testing.T) {
 	require.Equal(t, `{checksum="", path="github.com/prometheus/client_golang", version="(devel)"}`, firstMetricLset.String())
 }
 
+// Regression test against https://github.com/prometheus/prometheus/pull/16946
 func TestMetricStreamingDecoder_LabelsCorruption(t *testing.T) {
 	lastScrapeSize := 0
 	var allPreviousLabels []labels.Labels
 	buffers := pool.New(128, 1024, 2, func(sz int) any { return make([]byte, 0, sz) })
 	builder := labels.NewScratchBuilder(0)
+	builder.SetUnsafeAdd(true)
+
 	for _, labelsCount := range []int{1, 2, 3, 5, 8, 5, 3, 2, 1} {
 		// Get buffer from pool like in scrape.go
 		b := buffers.Get(lastScrapeSize).([]byte)
@@ -201,9 +206,10 @@ func TestMetricStreamingDecoder_LabelsCorruption(t *testing.T) {
 		require.NoError(t, d.NextMetricFamily())
 		require.NoError(t, d.NextMetric())
 
-		// Get the labels
+		// Get the labels. Decode is reusing strings when adding labels. We
+		// test if scratchBuilder with unsafeAdd set to true handles that.
 		builder.Reset()
-		require.NoError(t, d.Label(&builder)) // <- this uses unsafe strings to create labels
+		require.NoError(t, d.Label(&builder))
 		lbs := builder.Labels()
 		allPreviousLabels = append(allPreviousLabels, lbs)
 

schema/labels.go

@@ -139,18 +139,22 @@ func (m Metadata) SetToLabels(b *labels.Builder) {
 	b.Set(metricUnit, m.Unit)
 }
 
-// IgnoreOverriddenMetadataLabelsScratchBuilder is a wrapper over labels scratch builder
+// NewIgnoreOverriddenMetadataLabelScratchBuilder creates IgnoreOverriddenMetadataLabelScratchBuilder.
-// that ignores label additions that would collide with non-empty Overwrite Metadata fields.
+func (m Metadata) NewIgnoreOverriddenMetadataLabelScratchBuilder(b *labels.ScratchBuilder) *IgnoreOverriddenMetadataLabelScratchBuilder {
-type IgnoreOverriddenMetadataLabelsScratchBuilder struct {
+	return &IgnoreOverriddenMetadataLabelScratchBuilder{ScratchBuilder: b, overwrite: m}
-	*labels.ScratchBuilder
+}
 
-	Overwrite Metadata
+// IgnoreOverriddenMetadataLabelScratchBuilder is a wrapper over labels.ScratchBuilder
+// that ignores label additions that would collide with non-empty Overwrite Metadata fields.
+type IgnoreOverriddenMetadataLabelScratchBuilder struct {
+	*labels.ScratchBuilder
+	overwrite Metadata
 }
 
 // Add a name/value pair, unless it would collide with the non-empty Overwrite Metadata
 // field. Note if you Add the same name twice you will get a duplicate label, which is invalid.
-func (b IgnoreOverriddenMetadataLabelsScratchBuilder) Add(name, value string) {
+func (b IgnoreOverriddenMetadataLabelScratchBuilder) Add(name, value string) {
-	if !b.Overwrite.IsEmptyFor(name) {
+	if !b.overwrite.IsEmptyFor(name) {
 		return
 	}
 	b.ScratchBuilder.Add(name, value)

schema/labels_test.go

@@ -142,7 +142,7 @@ func TestIgnoreOverriddenMetadataLabelsScratchBuilder(t *testing.T) {
 		t.Run(fmt.Sprintf("meta=%#v", tcase.highPrioMeta), func(t *testing.T) {
 			lb := labels.NewScratchBuilder(0)
 			tcase.highPrioMeta.AddToLabels(&lb)
-			wrapped := &IgnoreOverriddenMetadataLabelsScratchBuilder{ScratchBuilder: &lb, Overwrite: tcase.highPrioMeta}
+			wrapped := tcase.highPrioMeta.NewIgnoreOverriddenMetadataLabelScratchBuilder(&lb)
 			incomingLabels.Range(func(l labels.Label) {
 				wrapped.Add(l.Name, l.Value)
 			})

tsdb/record/record.go

@@ -19,6 +19,7 @@ import (
 	"errors"
 	"fmt"
 	"math"
+	"unsafe"
 
 	"github.com/prometheus/common/model"
 
@@ -204,8 +205,10 @@ type Decoder struct {
 	builder labels.ScratchBuilder
 }
 
-func NewDecoder(*labels.SymbolTable) Decoder { // FIXME remove t
+func NewDecoder(*labels.SymbolTable) Decoder { // FIXME remove t (or use scratch builder with symbols)
-	return Decoder{builder: labels.NewScratchBuilder(0)}
+	b := labels.NewScratchBuilder(0)
+	b.SetUnsafeAdd(true)
+	return Decoder{builder: b}
 }
 
 // Type returns the type of the record.
@@ -289,6 +292,10 @@ func (*Decoder) Metadata(rec []byte, metadata []RefMetadata) ([]RefMetadata, err
 	return metadata, nil
 }
 
+func yoloString(b []byte) string {
+	return unsafe.String(unsafe.SliceData(b), len(b))
+}
+
 // DecodeLabels decodes one set of labels from buf.
 func (d *Decoder) DecodeLabels(dec *encoding.Decbuf) labels.Labels {
 	d.builder.Reset()
@@ -296,7 +303,7 @@ func (d *Decoder) DecodeLabels(dec *encoding.Decbuf) labels.Labels {
 	for range nLabels {
 		lName := dec.UvarintBytes()
 		lValue := dec.UvarintBytes()
-		d.builder.UnsafeAddBytes(lName, lValue)
+		d.builder.Add(yoloString(lName), yoloString(lValue))
 	}
 	return d.builder.Labels()
 }