[GCP] Dedup storage experiment (#363)

cmd/conformance/gcp/main.go

@@ -37,6 +37,7 @@ var (
 	listen            = flag.String("listen", ":2024", "Address:port to listen on")
 	spanner           = flag.String("spanner", "", "Spanner resource URI ('projects/.../...')")
 	signer            = flag.String("signer", "", "Note signer to use to sign checkpoints")
+	persistentDedup   = flag.Bool("gcp_dedup", false, "EXPERIMENTAL: Set to true to enable persistent dedupe storage")
 	additionalSigners = []string{}
 )
 
@@ -65,7 +66,18 @@ func main() {
 	if err != nil {
 		klog.Exitf("Failed to create new GCP storage: %v", err)
 	}
-	dedupeAdd := tessera.InMemoryDedupe(storage.Add, 256)
+
+	// Handle dedup configuration
+	addDelegate := storage.Add
+
+	// PersistentDedup is currently experimental, so there's no terraform or documentation yet!
+	if *persistentDedup {
+		addDelegate, err = gcp.NewDedupe(ctx, fmt.Sprintf("%s_dedup", *spanner), addDelegate)
+		if err != nil {
+			klog.Exitf("Failed to create new GCP dedupe: %v", err)
+		}
+	}
+	dedupeAdd := tessera.InMemoryDedupe(addDelegate, 256)
 
 	// Expose a HTTP handler for the conformance test writes.
 	// This should accept arbitrary bytes POSTed to /add, and return an ascii

dedupe.go

@@ -16,9 +16,10 @@ package tessera
 
 import (
 	"context"
+	"fmt"
 	"sync"
 
-	"github.com/hashicorp/golang-lru/v2/expirable"
+	lru "github.com/hashicorp/golang-lru/v2"
 )
 
 // InMemoryDedupe wraps an Add function to prevent duplicate entries being written to the underlying
@@ -35,30 +36,38 @@ import (
 // InMemoryDedupe. This allows recent duplicates to be deduplicated in memory, reducing the need to
 // make calls to a persistent storage.
 func InMemoryDedupe(delegate func(ctx context.Context, e *Entry) IndexFuture, size uint) func(context.Context, *Entry) IndexFuture {
+	c, err := lru.New[string, func() IndexFuture](int(size))
+	if err != nil {
+		panic(fmt.Errorf("lru.New(%d): %v", size, err))
+	}
 	dedupe := &inMemoryDedupe{
 		delegate: delegate,
-		cache:    expirable.NewLRU[string, IndexFuture](int(size), nil, 0),
+		cache:    c,
 	}
 	return dedupe.add
 }
 
 type inMemoryDedupe struct {
 	delegate func(ctx context.Context, e *Entry) IndexFuture
-	mu       sync.Mutex // cache is thread safe, but this mutex allows us to do conditional writes
-	cache    *expirable.LRU[string, IndexFuture]
+	cache    *lru.Cache[string, func() IndexFuture]
 }
 
 // Add adds the entry to the underlying delegate only if e hasn't been recently seen. In either case,
 // an IndexFuture will be returned that the client can use to get the sequence number of this entry.
 func (d *inMemoryDedupe) add(ctx context.Context, e *Entry) IndexFuture {
 	id := string(e.Identity())
-	d.mu.Lock()
-	defer d.mu.Unlock()
 
-	f, ok := d.cache.Get(id)
-	if !ok {
-		f = d.delegate(ctx, e)
-		d.cache.Add(id, f)
+	// However many calls with the same entry come in and are deduped, we should only call delegate
+	// once for each unique entry:
+	f := sync.OnceValue(func() IndexFuture {
+		return d.delegate(ctx, e)
+	})
+
+	// if we've seen this entry before, discard our f and replace
+	// with the one we created last time, otherwise store f against id.
+	if prev, ok, _ := d.cache.PeekOrAdd(id, f); ok {
+		f = prev
 	}
-	return f
+
+	return f()
 }

dedupe_test.go

@@ -60,13 +60,15 @@ func TestDedupe(t *testing.T) {
 			dedupeAdd := tessera.InMemoryDedupe(delegate, 256)
 
 			// Add foo, bar, baz to prime the cache to make things interesting
-			dedupeAdd(ctx, tessera.NewEntry([]byte("foo")))
-			dedupeAdd(ctx, tessera.NewEntry([]byte("bar")))
-			dedupeAdd(ctx, tessera.NewEntry([]byte("baz")))
+			for _, s := range []string{"foo", "bar", "baz"} {
+				if _, err := dedupeAdd(ctx, tessera.NewEntry([]byte(s)))(); err != nil {
+					t.Fatalf("dedupeAdd(%q): %v", s, err)
+				}
+			}
 
 			idx, err := dedupeAdd(ctx, tessera.NewEntry([]byte(tC.newValue)))()
 			if err != nil {
-				t.Fatal(err)
+				t.Fatalf("dedupeAdd(%q): %v", tC.newValue, err)
 			}
 			if idx != tC.wantIdx {
 				t.Errorf("got != want (%d != %d)", idx, tC.wantIdx)

storage/gcp/gcp.go

@@ -37,11 +37,13 @@ import (
 	"io"
 	"net/http"
 	"os"
+	"sync/atomic"
 	"time"
 
 	"cloud.google.com/go/spanner"
 	"cloud.google.com/go/spanner/apiv1/spannerpb"
 	gcs "cloud.google.com/go/storage"
+	"github.com/globocom/go-buffer"
 	"github.com/google/go-cmp/cmp"
 	"github.com/transparency-dev/merkle/rfc6962"
 	tessera "github.com/transparency-dev/trillian-tessera"
@@ -783,3 +785,170 @@ func (s *gcsStorage) lastModified(ctx context.Context, obj string) (time.Time, e
 	}
 	return r.Attrs.LastModified, r.Close()
 }
+
+// NewDedupe returns wrapped Add func which will use Spanner to maintain a mapping of
+// previously seen entries and their assigned indices. Future calls with the same entry
+// will return the previously assigned index, as yet unseen entries will be passed to the provided
+// delegate function to have an index assigned.
+//
+// For performance reasons, the ID -> index associations returned by the delegate are buffered before
+// being flushed to Spanner. This can result in duplicates occuring in some circumstances, but in
+// general this should not be a problem.
+//
+// Note that the storage for this mapping is entirely separate and unconnected to the storage used for
+// maintaining the Merkle tree.
+//
+// This functionality is experimental!
+func NewDedupe(ctx context.Context, spannerDB string, delegate func(ctx context.Context, e *tessera.Entry) tessera.IndexFuture) (func(ctx context.Context, e *tessera.Entry) tessera.IndexFuture, error) {
+	/*
+	   Schema for reference:
+
+	   	CREATE TABLE IDSeq (
+	   	 id INT64 NOT NULL,
+	   	 h BYTES(MAX) NOT NULL,
+	   	 idx INT64 NOT NULL,
+	   	) PRIMARY KEY (id, h);
+	*/
+	dedupDB, err := spanner.NewClient(ctx, spannerDB)
+	if err != nil {
+		return nil, fmt.Errorf("failed to connect to Spanner: %v", err)
+	}
+
+	r := &dedupStorage{
+		ctx:      ctx,
+		dbPool:   dedupDB,
+		delegate: delegate,
+	}
+
+	// TODO(al): Make these configurable
+	r.buf = buffer.New(
+		buffer.WithSize(64),
+		buffer.WithFlushInterval(200*time.Millisecond),
+		buffer.WithFlusher(buffer.FlusherFunc(r.flush)),
+		buffer.WithPushTimeout(15*time.Second),
+	)
+	go func(ctx context.Context) {
+		t := time.NewTicker(time.Second)
+		for {
+			select {
+			case <-ctx.Done():
+				return
+			case <-t.C:
+				klog.V(1).Infof("DEDUP: # Writes %d, # Lookups %d, # DB hits %v, # buffer Push discards %d", r.numWrites.Load(), r.numLookups.Load(), r.numDBDedups.Load(), r.numPushErrs.Load())
+			}
+		}
+	}(ctx)
+	return r.add, nil
+}
+
+type dedupStorage struct {
+	ctx      context.Context
+	dbPool   *spanner.Client
+	delegate func(ctx context.Context, e *tessera.Entry) tessera.IndexFuture
+
+	numLookups  atomic.Uint64
+	numWrites   atomic.Uint64
+	numDBDedups atomic.Uint64
+	numPushErrs atomic.Uint64
+
+	buf *buffer.Buffer
+}
+
+// index returns the index (if any) previously associated with the provided hash
+func (d *dedupStorage) index(ctx context.Context, h []byte) (*uint64, error) {
+	d.numLookups.Add(1)
+	var idx int64
+	if row, err := d.dbPool.Single().ReadRow(ctx, "IDSeq", spanner.Key{0, h}, []string{"idx"}); err != nil {
+		if c := spanner.ErrCode(err); c == codes.NotFound {
+			return nil, nil
+		}
+		return nil, err
+	} else {
+		if err := row.Column(0, &idx); err != nil {
+			return nil, fmt.Errorf("failed to read dedup index: %v", err)
+		}
+		idx := uint64(idx)
+		d.numDBDedups.Add(1)
+		return &idx, nil
+	}
+}
+
+// storeMappings stores the associations between the keys and IDs in a non-atomic fashion
+// (i.e. it does not store all or none in a transactional sense).
+//
+// Returns an error if one or more mappings cannot be stored.
+func (d *dedupStorage) storeMappings(ctx context.Context, entries []dedupeMapping) error {
+	m := make([]*spanner.MutationGroup, 0, len(entries))
+	for _, e := range entries {
+		m = append(m, &spanner.MutationGroup{
+			Mutations: []*spanner.Mutation{spanner.Insert("IDSeq", []string{"id", "h", "idx"}, []interface{}{0, e.ID, int64(e.Idx)})},
+		})
+	}
+
+	i := d.dbPool.BatchWrite(ctx, m)
+	return i.Do(func(r *spannerpb.BatchWriteResponse) error {
+		s := r.GetStatus()
+		if c := codes.Code(s.Code); c != codes.OK && c != codes.AlreadyExists {
+			return fmt.Errorf("failed to write dedup record: %v (%v)", s.GetMessage(), c)
+		}
+		return nil
+	})
+}
+
+// dedupeMapping represents an ID -> index mapping.
+type dedupeMapping struct {
+	ID  []byte
+	Idx uint64
+}
+
+// add adds the entry to the underlying delegate only if e isn't already known. In either case,
+// an IndexFuture will be returned that the client can use to get the sequence number of this entry.
+func (d *dedupStorage) add(ctx context.Context, e *tessera.Entry) tessera.IndexFuture {
+	idx, err := d.index(ctx, e.Identity())
+	if err != nil {
+		return func() (uint64, error) { return 0, err }
+	}
+	if idx != nil {
+		return func() (uint64, error) { return *idx, nil }
+	}
+
+	i, err := d.delegate(ctx, e)()
+	if err != nil {
+		return func() (uint64, error) { return 0, err }
+	}
+
+	err = d.enqueueMapping(ctx, e.Identity(), i)
+	return func() (uint64, error) {
+		return i, err
+	}
+}
+
+// enqueueMapping buffers the provided ID -> index mapping ready to be flushed to storage.
+func (d *dedupStorage) enqueueMapping(_ context.Context, h []byte, idx uint64) error {
+	err := d.buf.Push(dedupeMapping{ID: h, Idx: idx})
+	if err != nil {
+		d.numPushErrs.Add(1)
+		// This means there's pressure flushing dedup writes out, so discard this write.
+		if err != buffer.ErrTimeout {
+			return err
+		}
+	}
+	return nil
+}
+
+// flush writes enqueued mappings to storage.
+func (d *dedupStorage) flush(items []interface{}) {
+	entries := make([]dedupeMapping, len(items))
+	for i := range items {
+		entries[i] = items[i].(dedupeMapping)
+	}
+
+	ctx, c := context.WithTimeout(d.ctx, 15*time.Second)
+	defer c()
+
+	if err := d.storeMappings(ctx, entries); err != nil {
+		klog.Infof("Failed to flush dedup entries: %v", err)
+		return
+	}
+	d.numWrites.Add(uint64(len(entries)))
+}