feat(blooms)!: Index structured metadata into blooms (#14061)

Instead of indexing n-grams of the log line content, we index the plain values of the structured metadata keys and values. Resulting tokens are: * `name` * `chunkPrefix + name` * `value` * `chunkPrefix + value` * `name + '=' value` * `chunkPrefix + name + '=' + value` Indexed fields (metadata name) are also extracted into the series metadata. Indexed metadata values cannot be queried with the bloom gateways yet. This PR does not cleanup unused code used for ngram-tokenization, it is scope for a follow up. --- Signed-off-by: Christian Haudum <[email protected]>
grafana · Sep 10, 2024 · a2fbaa8 · a2fbaa8
1 parent 5f78784
commit a2fbaa8
Show file tree

Hide file tree

Showing 10 changed files with 264 additions and 177 deletions.
diff --git a/pkg/storage/bloom/v1/bloom_tokenizer.go b/pkg/storage/bloom/v1/bloom_tokenizer.go
@@ -2,7 +2,6 @@ package v1
 
 import (
 	"math"
-	"unsafe"
 
 	"github.com/go-kit/log"
 	"github.com/go-kit/log/level"
@@ -92,19 +91,15 @@ func estimatedCount(m uint, p float64) uint {
 
 // Populates a bloom filter(s) with the tokens from the given chunks.
 // Called once per series
-func (bt *BloomTokenizer) Populate(
-	blooms v2iter.SizedIterator[*Bloom],
-	chks v2iter.Iterator[ChunkRefWithIter],
-	ch chan *BloomCreation,
-) {
+func (bt *BloomTokenizer) Populate(blooms v2iter.SizedIterator[*Bloom], chks v2iter.Iterator[ChunkRefWithIter], ch chan *BloomCreation) {
 	clear(bt.cache) // MUST always clear the cache before starting a new series
 	var next bool
 
 	// All but the last bloom are considered full -- send back unaltered
 	for next = blooms.Next(); next && blooms.Remaining() > 0; next = blooms.Next() {
 		ch <- &BloomCreation{
-			Bloom:            blooms.At(),
-			SourceBytesAdded: 0,
+			Bloom: blooms.At(),
+			Info:  newIndexingInfo(),
 		}
 	}
 
@@ -118,35 +113,29 @@ func (bt *BloomTokenizer) Populate(
 		// We have the feeling that the empty blooms may be reused from old blocks.
 		// Here we log an error if we find an empty bloom.
 		if bloom.Count() == 0 {
-			level.Warn(bt.logger).Log(
-				"msg", "found existing empty bloom",
-			)
+			level.Warn(bt.logger).Log("msg", "found existing empty bloom")
 		}
 	} else {
 		bloom = NewBloom()
 	}
 
-	var bytesAdded int
+	info := newIndexingInfo()
 
 	for chks.Next() {
 		chk := chks.At()
 		itr := v2iter.NewPeekIter(chk.Itr)
 
 		for {
-			full, newBytes := bt.addChunkToBloom(
-				bloom,
-				chk.Ref,
-				itr,
-			)
-			bytesAdded += newBytes
+			full, chunkStats := bt.addChunkToBloom(bloom, chk.Ref, itr)
+			info = info.merge(chunkStats)
 
 			// If a bloom is full, the chunk wasn't completely added
 			// so we'll submit this bloom, start a new one, and continue indexing
 			if full {
-				bt.sendBloom(ch, bloom, bytesAdded)
+				bt.sendBloom(ch, bloom, info)
 
-				// start a new bloom + reset bytesAdded counter
-				bytesAdded = 0
+				// start a new bloom + reset stats
+				info = newIndexingInfo()
 				bloom = NewBloom()
 
 				// cache _MUST_ be cleared when a new bloom is created to ensure that all tokens from
@@ -161,21 +150,15 @@ func (bt *BloomTokenizer) Populate(
 
 	// TODO(salvacorts): Delete this once we solve the correctness bug
 	if bloom.Count() == 0 {
-		level.Warn(bt.logger).Log(
-			"msg", "resulting bloom is empty",
-		)
+		level.Warn(bt.logger).Log("msg", "resulting bloom is empty")
 	}
 
 	// Send the last bloom
-	bt.sendBloom(ch, bloom, bytesAdded)
+	bt.sendBloom(ch, bloom, info)
 	close(ch)
 }
 
-func (bt *BloomTokenizer) sendBloom(
-	ch chan<- *BloomCreation,
-	bloom *Bloom,
-	bytesAdded int,
-) {
+func (bt *BloomTokenizer) sendBloom(ch chan<- *BloomCreation, bloom *Bloom, info indexingInfo) {
 	fillRatio := bloom.ScalableBloomFilter.FillRatio()
 	bt.metrics.hammingWeightRatio.Observe(fillRatio)
 	bt.metrics.estimatedCount.Observe(
@@ -184,70 +167,57 @@ func (bt *BloomTokenizer) sendBloom(
 	bt.metrics.bloomSize.Observe(float64(bloom.ScalableBloomFilter.Capacity() / eightBits))
 	bt.metrics.bloomsTotal.Inc()
 	ch <- &BloomCreation{
-		Bloom:            bloom,
-		SourceBytesAdded: bytesAdded,
+		Bloom: bloom,
+		Info:  info,
 	}
 }
 
-// addChunkToBloom adds the tokens from the given chunk to the given bloom.
-// It continues until the chunk is exhausted or the bloom is full.
-// NB(owen-d): We ensure the invariant that each line is indexed entirely into at least one bloom.
-// This includes both raw ngrams and chunk-prefixed ngrams and is why we use a peeking iterator --
-// so we can advance the iterator only after we're sure the bloom has accepted the line.
-// This is because the _line_ is the atom in Loki's data model and a query must either match (or not) an individual line.
-// Therefore, we index entire lines into a bloom to ensure a lookups are accurate.
-func (bt *BloomTokenizer) addChunkToBloom(bloom *Bloom, ref ChunkRef, entryIter v2iter.PeekIterator[push.Entry]) (full bool, bytesAdded int) {
+func prefixForChunkRef(chk ChunkRef) []byte {
+	enc := encoding.EncWith(make([]byte, 0, 20))
+	enc.PutBE64(uint64(chk.From))    // 8 bytes
+	enc.PutBE64(uint64(chk.Through)) // 8 bytes
+	enc.PutBE32(chk.Checksum)        // 4 bytes
+	return enc.Get()
+}
+
+// addChunkToBloom adds the values from structured metadata from the entries of the given chunk to the given bloom.
+// addChunkToBloom returns true if the bloom has been completely filled, and may not have consumed the entire iterator.
+// addChunkToBloom must be called multiple times until returning false with new blooms until the iterator has been fully consumed.
+func (bt *BloomTokenizer) addChunkToBloom(bloom *Bloom, ref ChunkRef, entryIter v2iter.PeekIterator[push.Entry]) (bool, indexingInfo) {
 	var (
-		tokenBuf, prefixLn = prefixedToken(bt.lineTokenizer.N(), ref, nil)
-		tokens             int
-		successfulInserts  int
-		cachedInserts      int
-		collisionInserts   int
-		chunkBytes         int
-		linesAdded         int
+		tokens            int
+		successfulInserts int
+		cachedInserts     int
+		collisionInserts  int
+		linesAdded        int
+
+		collision bool
 	)
 
+	// return values
+	full, info := false, newIndexingInfo()
+
+	tokenizer := NewStructuredMetadataTokenizer(string(prefixForChunkRef(ref)))
+
 	// We use a peeking iterator to avoid advancing the iterator until we're sure the bloom has accepted the line.
-outer:
 	for entry, ok := entryIter.Peek(); ok; entry, ok = entryIter.Peek() {
-		line := entry.Line
-		chunkBytes += len(line)
-
-		tokenItrs := []v2iter.Iterator[[]byte]{
-			// two iterators, one for the raw tokens and one for the chunk prefixed tokens.
-			// Warning: the underlying line tokenizer (used in both iterators) uses the same buffer for tokens.
-			// They are NOT SAFE for concurrent use.
-			NewPrefixedTokenIter(tokenBuf, prefixLn, bt.lineTokenizer.Tokens(line)),
-			bt.lineTokenizer.Tokens(line),
-		}
+		for _, kv := range entry.StructuredMetadata {
+			info.sourceBytes += len(kv.Name) + len(kv.Value)
+			info.indexedFields.Add(Field(kv.Name))
 
-		for _, itr := range tokenItrs {
-			for itr.Next() {
-				tok := itr.At()
+			tokenItr := tokenizer.Tokens(kv)
+			for tokenItr.Next() {
+				tok := tokenItr.At()
 				tokens++
 
-				// TODO[owen-d]: [n]byte this
-				// To avoid allocations, an unsafe string can be used to check ownership in cache.
-				str := unsafe.String(unsafe.SliceData(tok), len(tok))
 				// A cache is used ahead of the SBF, as it cuts out the costly operations of scaling bloom filters
-				if _, found := bt.cache[str]; found {
+				if _, found := bt.cache[tok]; found {
 					cachedInserts++
 					continue
 				}
 
 				// maxBloomSize is in bytes, but blooms operate at the bit level; adjust
-				var collision bool
-				collision, full = bloom.ScalableBloomFilter.TestAndAddWithMaxSize(tok, bt.maxBloomSize*eightBits)
-
-				if full {
-					// edge case: one line maxed out the bloom size -- retrying is futile
-					// (and will loop endlessly), so we'll just skip indexing it
-					if linesAdded == 0 {
-						_ = entryIter.Next()
-					}
-
-					break outer
-				}
+				collision, full = bloom.ScalableBloomFilter.TestAndAddWithMaxSize([]byte(tok), bt.maxBloomSize*eightBits)
 
 				if collision {
 					collisionInserts++
@@ -257,8 +227,7 @@ outer:
 
 				// only register the key in the cache if it was successfully added to the bloom
 				// as can prevent us from trying subsequent copies
-				str = string(tok)
-				bt.cache[str] = nil
+				bt.cache[tok] = nil
 				if len(bt.cache) >= cacheSize { // While crude, this has proven efficient in performance testing.  This speaks to the similarity in log lines near each other
 					clear(bt.cache)
 				}
@@ -268,14 +237,19 @@ outer:
 		// Only advance the iterator once we're sure the bloom has accepted the line
 		linesAdded++
 		_ = entryIter.Next()
+
+		// Only break out of the loop if the bloom filter is full after indexing all structured metadata of an entry.
+		if full {
+			break
+		}
 	}
 
 	// update metrics after each chunk added for more consistent reporting
 	bt.metrics.tokensTotal.Add(float64(tokens))
 	bt.metrics.insertsTotal.WithLabelValues(collisionTypeFalse).Add(float64(successfulInserts))
 	bt.metrics.insertsTotal.WithLabelValues(collisionTypeCache).Add(float64(cachedInserts))
 	bt.metrics.insertsTotal.WithLabelValues(collisionTypeTrue).Add(float64(collisionInserts))
-	bt.metrics.sourceBytesAdded.Add(float64(chunkBytes))
+	bt.metrics.sourceBytesAdded.Add(float64(info.sourceBytes))
 
-	return full, chunkBytes
+	return full, info
 }