store: read postings directly into delta encoded format

Instead of allocating bytes for raw postings, let's read them directly
into diff varint format to save memory.

Signed-off-by: Giedrius Statkevičius <[email protected]>

pkg/store/bucket.go

@@ -2486,6 +2486,12 @@ func (r *bucketIndexReader) fetchExpandedPostingsFromCache(ctx context.Context,
 	return true, ps, nil
 }
 
+var bufioReaderPool = sync.Pool{
+	New: func() any {
+		return bufio.NewReader(nil)
+	},
+}
+
 // fetchPostings fill postings requested by posting groups.
 // It returns one posting for each key, in the same order.
 // If postings for given key is not fetched, entry at given index will be nil.
@@ -2570,52 +2576,60 @@ func (r *bucketIndexReader) fetchPostings(ctx context.Context, keys []labels.Lab
 		// We assume index does not have any ptrs that has 0 length.
 		length := int64(part.End) - start
 
+		brdr := bufioReaderPool.Get().(*bufio.Reader)
+		defer bufioReaderPool.Put(brdr)
+
 		// Fetch from object storage concurrently and update stats and posting list.
 		g.Go(func() error {
 			begin := time.Now()
 
-			b, err := r.block.readIndexRange(ctx, start, length)
+			partReader, err := r.block.bkt.GetRange(ctx, r.block.indexFilename(), start, length)
 			if err != nil {
 				return errors.Wrap(err, "read postings range")
 			}
-			fetchTime := time.Since(begin)
+			defer runutil.CloseWithLogOnErr(r.block.logger, partReader, "readIndexRange close range reader")
+			brdr.Reset(partReader)
+
+			rdr := newPostingsReaderBuilder(ctx, brdr, ptrs[i:j], start, length)
 
 			r.mtx.Lock()
 			r.stats.postingsFetchCount++
 			r.stats.postingsFetched += j - i
-			r.stats.PostingsFetchDurationSum += fetchTime
 			r.stats.PostingsFetchedSizeSum += units.Base2Bytes(int(length))
 			r.mtx.Unlock()
 
-			for _, p := range ptrs[i:j] {
-				// index-header can estimate endings, which means we need to resize the endings.
-				pBytes, err := resizePostings(b[p.ptr.Start-start : p.ptr.End-start])
+			for rdr.Next() {
+				diffVarintPostings, postingsCount, keyID := rdr.AtDiffVarint()
+
+				output[keyID] = newDiffVarintPostings(diffVarintPostings, nil)
+
+				startCompression := time.Now()
+				dataToCache, err := snappyStreamedEncode(int(postingsCount), diffVarintPostings)
 				if err != nil {
-					return err
+					r.mtx.Lock()
+					r.stats.cachedPostingsCompressionErrors += 1
+					r.mtx.Unlock()
+					return errors.Wrap(err, "encoding with snappy")
 				}
 
-				// Reencode postings before storing to cache. If that fails, we store original bytes.
-				// This can only fail, if postings data was somehow corrupted,
-				// and there is nothing we can do about it.
-				// Errors from corrupted postings will be reported when postings are used.
-				bep := newBigEndianPostings(pBytes[4:])
-				dataToCache, compressionTime, compressionErrors, compressedSize := r.encodePostingsToCache(bep, bep.length())
 				r.mtx.Lock()
-				// Return postings and fill LRU cache.
-				// Truncate first 4 bytes which are length of posting.
-				output[p.keyID] = newBigEndianPostings(pBytes[4:])
-
-				r.block.indexCache.StorePostings(r.block.meta.ULID, keys[p.keyID], dataToCache)
-
-				// If we just fetched it we still have to update the stats for touched postings.
 				r.stats.postingsTouched++
-				r.stats.PostingsTouchedSizeSum += units.Base2Bytes(len(pBytes))
+				r.stats.PostingsTouchedSizeSum += units.Base2Bytes(int(len(diffVarintPostings)))
 				r.stats.cachedPostingsCompressions += 1
-				r.stats.cachedPostingsCompressionErrors += compressionErrors
-				r.stats.CachedPostingsOriginalSizeSum += units.Base2Bytes(len(pBytes))
-				r.stats.CachedPostingsCompressedSizeSum += units.Base2Bytes(compressedSize)
-				r.stats.CachedPostingsCompressionTimeSum += compressionTime
+				r.stats.CachedPostingsOriginalSizeSum += units.Base2Bytes(len(diffVarintPostings))
+				r.stats.CachedPostingsCompressedSizeSum += units.Base2Bytes(len(dataToCache))
+				r.stats.CachedPostingsCompressionTimeSum += time.Since(startCompression)
 				r.mtx.Unlock()
+
+				r.block.indexCache.StorePostings(r.block.meta.ULID, keys[keyID], dataToCache)
+			}
+
+			r.mtx.Lock()
+			r.stats.PostingsFetchDurationSum += time.Since(begin)
+			r.mtx.Unlock()
+
+			if rdr.Error() != nil {
+				return errors.Wrap(err, "reading postings")
 			}
 			return nil
 		})
@@ -2665,21 +2679,6 @@ func (r *bucketIndexReader) encodePostingsToCache(p index.Postings, length int)
 	return dataToCache, compressionTime, compressionErrors, compressedSize
 }
 
-func resizePostings(b []byte) ([]byte, error) {
-	d := encoding.Decbuf{B: b}
-	n := d.Be32int()
-	if d.Err() != nil {
-		return nil, errors.Wrap(d.Err(), "read postings list")
-	}
-
-	// 4 for postings number of entries, then 4, foreach each big endian posting.
-	size := 4 + n*4
-	if len(b) < size {
-		return nil, encoding.ErrInvalidSize
-	}
-	return b[:size], nil
-}
-
 // bigEndianPostings implements the Postings interface over a byte stream of
 // big endian numbers.
 type bigEndianPostings struct {

pkg/store/postings.go

@@ -0,0 +1,142 @@
+// Copyright (c) The Thanos Authors.
+// Licensed under the Apache License 2.0.
+
+package store
+
+import (
+	"bufio"
+	"context"
+	"encoding/binary"
+	"fmt"
+	"io"
+
+	"github.com/pkg/errors"
+)
+
+type postingsReaderBuilder struct {
+	e       error
+	readBuf []byte
+
+	r        *bufio.Reader
+	postings []postingPtr
+
+	lastOffset int64
+	pi         int
+
+	start, length         int64
+	cur                   []byte
+	keyID                 int
+	repeatFor             int
+	numberOfPostingsInCur uint64
+	uvarintEncodeBuf      []byte
+	ctx                   context.Context
+}
+
+// newPostingsReaderBuilder is a builder that reads directly from the index
+// and builds a diff varint encoded []byte that could be later used directly.
+func newPostingsReaderBuilder(ctx context.Context, r *bufio.Reader, postings []postingPtr, start, length int64) *postingsReaderBuilder {
+	prb := &postingsReaderBuilder{
+		r:                r,
+		readBuf:          make([]byte, 4),
+		start:            start,
+		length:           length,
+		postings:         postings,
+		uvarintEncodeBuf: make([]byte, binary.MaxVarintLen64),
+		ctx:              ctx,
+	}
+
+	return prb
+}
+
+func getInt32(r io.Reader, buf []byte) (uint32, error) {
+	read, err := r.Read(buf)
+	if err != nil {
+		return 0, errors.Wrap(err, "reading")
+	}
+	if read != 4 {
+		return 0, fmt.Errorf("read got %d bytes instead of 4", read)
+	}
+	return binary.BigEndian.Uint32(buf), nil
+}
+
+func (r *postingsReaderBuilder) Next() bool {
+	if r.ctx.Err() != nil {
+		r.e = r.ctx.Err()
+		return false
+	}
+	if r.repeatFor > 0 {
+		r.keyID = r.postings[r.pi-r.repeatFor].keyID
+		r.repeatFor--
+		return true
+	}
+	if r.pi >= len(r.postings) {
+		return false
+	}
+	if r.Error() != nil {
+		return false
+	}
+	from := r.postings[r.pi].ptr.Start - r.start
+
+	if from-r.lastOffset < 0 {
+		panic("would have skipped negative bytes")
+	}
+
+	_, err := r.r.Discard(int(from - r.lastOffset))
+	if err != nil {
+		return false
+	}
+	r.lastOffset += from - r.lastOffset
+
+	postingsCount, err := getInt32(r.r, r.readBuf[:])
+	if err != nil {
+		r.e = err
+		return false
+	}
+	r.lastOffset += 4
+
+	// Assume 1.25 bytes per compressed posting.
+	r.cur = make([]byte, 0, int(float64(postingsCount)*1.25))
+
+	prev := uint32(0)
+
+	for i := 0; i < int(postingsCount); i++ {
+		posting, err := getInt32(r.r, r.readBuf[:])
+		if err != nil {
+			r.e = err
+			return false
+		}
+		r.lastOffset += 4
+
+		uvarintSize := binary.PutUvarint(r.uvarintEncodeBuf, uint64(posting-prev))
+		r.cur = append(r.cur, r.uvarintEncodeBuf[:uvarintSize]...)
+		prev = posting
+	}
+	r.numberOfPostingsInCur = uint64(postingsCount)
+
+	r.keyID = r.postings[r.pi].keyID
+	r.pi++
+	for {
+		if r.pi >= len(r.postings) {
+			break
+		}
+
+		if r.postings[r.pi].ptr.Start == r.postings[r.pi-1].ptr.Start &&
+			r.postings[r.pi].ptr.End == r.postings[r.pi-1].ptr.End {
+			r.repeatFor++
+			r.pi++
+			continue
+		}
+
+		break
+	}
+
+	return true
+}
+
+func (r *postingsReaderBuilder) Error() error {
+	return r.e
+}
+
+func (r *postingsReaderBuilder) AtDiffVarint() ([]byte, uint64, int) {
+	return r.cur, r.numberOfPostingsInCur, r.keyID
+}

pkg/store/postings_codec.go

@@ -339,3 +339,27 @@ func (it *diffVarintPostings) Seek(x storage.SeriesRef) bool {
 func (it *diffVarintPostings) Err() error {
 	return it.buf.Err()
 }
+
+func snappyStreamedEncode(postingsLength int, diffVarintPostings []byte) ([]byte, error) {
+	compressedBuf := bytes.NewBuffer(make([]byte, 0, estimateSnappyStreamSize(postingsLength)))
+	if n, err := compressedBuf.WriteString(codecHeaderStreamedSnappy); err != nil {
+		return nil, fmt.Errorf("writing streamed snappy header")
+	} else if n != len(codecHeaderStreamedSnappy) {
+		return nil, fmt.Errorf("short-write streamed snappy header")
+	}
+
+	sw, err := extsnappy.Compressor.Compress(compressedBuf)
+	if err != nil {
+		return nil, fmt.Errorf("creating snappy compressor: %w", err)
+	}
+
+	_, err = sw.Write(diffVarintPostings)
+	if err != nil {
+		return nil, err
+	}
+	if err := sw.Close(); err != nil {
+		return nil, errors.Wrap(err, "closing snappy stream writer")
+	}
+
+	return compressedBuf.Bytes(), nil
+}