feat(photos): separate object storage from metadata layer

Move physical file facts out of the photos table into an objects ledger:

- objects / photo_objects / variant_jobs tables (migration 000008)
- drop storage_key/thumb_key/preview_key/size/width/height/mime_type from
  photos (migration 000009); they are derived via joins to the original object
- album-free storage keys: objects/{id}/{role}.{ext}
- explicit GC via pending/ready/deleting states; objects deleted only when no
  photo references them
- background worker drives variant generation (durable variant_jobs) and GC
- copy/move are metadata-only; storage objects are shared between photos
- remove unused storage.DeletePrefix

Non-destructive to the external JSON API: size/dimensions/mime and URLs are
still returned, just sourced from the object layer.

OpenSpec change: media-object-layer

backend/cmd/server/main.go

@@ -69,6 +69,8 @@ func main() {
 	hub := sse.NewHub(cfg.SSE.ClientBuffer)
 	server := app.New(cfg, pool, store, hub)
 
+	go server.StartWorker(ctx)
+
 	httpServer := &http.Server{
 		Addr:              cfg.Server.Addr,
 		Handler:           server.Router(),

backend/internal/albums/store.go

@@ -44,11 +44,12 @@ func (s *Store) GetByID(ctx context.Context, id string) (*models.Album, error) {
 	var a models.Album
 	err := s.pool.QueryRow(ctx,
 		`SELECT a.id, a.owner_id, a.title, a.description, a.cover_photo_id, a.created_at, a.updated_at,
-		        COALESCE(p.thumb_key, ''),
+		        COALESCE(co.storage_key, ''),
 		        (SELECT count(*) FROM photos WHERE album_id = a.id),
 		        (SELECT count(*) FROM album_members WHERE album_id = a.id)
 		 FROM albums a
-		 LEFT JOIN photos p ON p.id = a.cover_photo_id
+		 LEFT JOIN photo_objects cpo ON cpo.photo_id = a.cover_photo_id AND cpo.role = 'thumb'
+		 LEFT JOIN objects co ON co.id = cpo.object_id
 		 WHERE a.id = $1`, id).
 		Scan(&a.ID, &a.OwnerID, &a.Title, &a.Description, &a.CoverPhotoID, &a.CreatedAt, &a.UpdatedAt, &a.CoverThumbKey, &a.PhotosCount, &a.MembersCount)
 	if err != nil {
@@ -63,12 +64,13 @@ func (s *Store) GetByID(ctx context.Context, id string) (*models.Album, error) {
 func (s *Store) ListForUser(ctx context.Context, userID string) ([]models.Album, error) {
 	rows, err := s.pool.Query(ctx,
 		`SELECT a.id, a.owner_id, a.title, a.description, a.cover_photo_id, a.created_at, a.updated_at,
-		        COALESCE(p.thumb_key, ''),
+		        COALESCE(co.storage_key, ''),
 		        (SELECT count(*) FROM photos WHERE album_id = a.id),
 		        (SELECT count(*) FROM album_members WHERE album_id = a.id)
 		 FROM albums a
 		 LEFT JOIN album_members m ON m.album_id = a.id AND m.user_id = $1
-		 LEFT JOIN photos p ON p.id = a.cover_photo_id
+		 LEFT JOIN photo_objects cpo ON cpo.photo_id = a.cover_photo_id AND cpo.role = 'thumb'
+		 LEFT JOIN objects co ON co.id = cpo.object_id
 		 WHERE a.owner_id = $1 OR m.user_id IS NOT NULL
 		 ORDER BY a.created_at DESC`, userID)
 	if err != nil {

backend/internal/app/app.go

@@ -1,6 +1,7 @@
 package app
 
 import (
+	"context"
 	"io/fs"
 	"net/http"
 	"strings"
@@ -37,6 +38,7 @@ type Server struct {
 	membersSvc *members.Service
 
 	photos *photos.Service
+	worker *photos.Worker
 
 	social *social.Service
 
@@ -61,6 +63,7 @@ func New(cfg *config.Config, pool *pgxpool.Pool, st *storage.Client, hub *sse.Hu
 
 	photosStore := photos.NewStore(pool)
 	photosSvc := photos.NewService(photosStore, albumsSvc, st, hub)
+	photosWorker := photos.NewWorker(photosSvc, cfg.Worker.Interval, cfg.Worker.Batch, cfg.Worker.PendingTTL)
 
 	socialStore := social.NewStore(pool)
 	socialSvc := social.NewService(socialStore, albumsSvc, photosStore, hub, notifSvc)
@@ -78,12 +81,20 @@ func New(cfg *config.Config, pool *pgxpool.Pool, st *storage.Client, hub *sse.Hu
 		auth:          authSvc,
 		membersSvc:    membersSvc,
 		photos:        photosSvc,
+		worker:        photosWorker,
 		social:        socialSvc,
 		notifications: notifSvc,
 		share:         shareSvc,
 	}
 }
 
+// StartWorker runs the background reconciler until ctx is cancelled.
+func (s *Server) StartWorker(ctx context.Context) {
+	if s.worker != nil {
+		s.worker.Run(ctx)
+	}
+}
+
 func (s *Server) Router() *gin.Engine {
 	r := gin.New()
 	r.Use(gin.Recovery())

backend/internal/config/config.go

@@ -48,12 +48,19 @@ type SSE struct {
 	ClientBuffer      int
 }
 
+type Worker struct {
+	Interval   time.Duration
+	Batch      int
+	PendingTTL time.Duration
+}
+
 type Config struct {
 	Postgres Postgres
 	S3       S3
 	Auth     Auth
 	Server   Server
 	SSE      SSE
+	Worker   Worker
 }
 
 func Load() (*Config, error) {
@@ -92,6 +99,21 @@ func Load() (*Config, error) {
 		return nil, fmt.Errorf("invalid S3_USE_PATH_STYLE: %w", err)
 	}
 
+	workerInterval, err := time.ParseDuration(getEnv("WORKER_INTERVAL", "5s"))
+	if err != nil {
+		return nil, fmt.Errorf("invalid WORKER_INTERVAL: %w", err)
+	}
+
+	workerBatch, err := strconv.Atoi(getEnv("WORKER_BATCH", "10"))
+	if err != nil {
+		return nil, fmt.Errorf("invalid WORKER_BATCH: %w", err)
+	}
+
+	workerPendingTTL, err := time.ParseDuration(getEnv("WORKER_PENDING_TTL", "10m"))
+	if err != nil {
+		return nil, fmt.Errorf("invalid WORKER_PENDING_TTL: %w", err)
+	}
+
 	cfg := &Config{
 		Postgres: Postgres{
 			Host:     getEnv("PG_HOST", "pg.localnet"),
@@ -124,6 +146,11 @@ func Load() (*Config, error) {
 			HeartbeatInterval: heartbeat,
 			ClientBuffer:      bufSize,
 		},
+		Worker: Worker{
+			Interval:   workerInterval,
+			Batch:      workerBatch,
+			PendingTTL: workerPendingTTL,
+		},
 	}
 
 	if cfg.Auth.JWTSecret == "" {

backend/internal/photos/copy_test.go

@@ -0,0 +1,92 @@
+package photos
+
+import (
+	"context"
+	"testing"
+	"time"
+)
+
+func TestCopySharesStorageObjects(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	svc, cl := newTestService(t, pool)
+	user := createUser(t, ctx, pool)
+	albumA := createAlbum(t, ctx, pool, user)
+	albumB := createAlbum(t, ctx, pool, user)
+
+	p, err := svc.Upload(ctx, UploadInput{
+		AlbumID: albumA, UserID: user, Filename: "p.png", MimeType: "image/png", Data: makePNG(t),
+	})
+	if err != nil {
+		t.Fatalf("upload: %v", err)
+	}
+	t.Cleanup(func() { cleanupPhoto(t, ctx, pool, cl, p.ID) })
+
+	w := NewWorker(svc, time.Second, 10, time.Minute)
+	w.processVariantJobs(ctx)
+
+	srcIDs, err := svc.store.PhotoObjectIDs(ctx, p.ID)
+	if err != nil || len(srcIDs) != 3 {
+		t.Fatalf("source objects = %v err=%v", srcIDs, err)
+	}
+
+	clones, err := svc.CopyBatch(ctx, []string{p.ID}, user, albumB)
+	if err != nil || len(clones) != 1 {
+		t.Fatalf("copy: %v len=%d", err, len(clones))
+	}
+	clone := clones[0]
+	t.Cleanup(func() { cleanupPhoto(t, ctx, pool, cl, clone.ID) })
+
+	cloneIDs, err := svc.store.PhotoObjectIDs(ctx, clone.ID)
+	if err != nil || len(cloneIDs) != 3 {
+		t.Fatalf("clone objects = %v err=%v", cloneIDs, err)
+	}
+	if !sameSet(srcIDs, cloneIDs) {
+		t.Fatalf("clone does not share objects: src=%v clone=%v", srcIDs, cloneIDs)
+	}
+	for _, id := range cloneIDs {
+		if n, err := svc.store.ObjectRefCount(ctx, id); err != nil || n != 2 {
+			t.Fatalf("object %s ref count = %d err=%v, want 2", id, n, err)
+		}
+	}
+
+	// Move the source into the same album; the storage key and shared objects
+	// must be untouched.
+	beforeKey := p.StorageKey
+	moved, err := svc.MoveBatch(ctx, []string{p.ID}, user, albumB)
+	if err != nil || moved != 1 {
+		t.Fatalf("move: %v moved=%d", err, moved)
+	}
+	after, err := svc.store.GetByID(ctx, p.ID, "")
+	if err != nil {
+		t.Fatalf("get after move: %v", err)
+	}
+	if after.AlbumID != albumB {
+		t.Fatalf("photo not moved: album=%q", after.AlbumID)
+	}
+	if after.StorageKey != beforeKey {
+		t.Fatalf("move changed storage key: %q -> %q", beforeKey, after.StorageKey)
+	}
+	for _, id := range cloneIDs {
+		if n, err := svc.store.ObjectRefCount(ctx, id); err != nil || n != 2 {
+			t.Fatalf("object %s ref count after move = %d err=%v, want 2", id, n, err)
+		}
+	}
+}
+
+func sameSet(a, b []string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	m := make(map[string]int, len(a))
+	for _, v := range a {
+		m[v]++
+	}
+	for _, v := range b {
+		m[v]--
+		if m[v] < 0 {
+			return false
+		}
+	}
+	return true
+}

backend/internal/photos/helpers_test.go

@@ -0,0 +1,149 @@
+package photos
+
+import (
+	"bytes"
+	"context"
+	"image"
+	"image/color"
+	"image/png"
+	"os"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"shared-album/internal/auth"
+	"shared-album/internal/config"
+	"shared-album/internal/models"
+	"shared-album/internal/roles"
+	"shared-album/internal/sse"
+	"shared-album/internal/storage"
+)
+
+func testPool(t *testing.T) *pgxpool.Pool {
+	t.Helper()
+	dsn := os.Getenv("TEST_DATABASE_URL")
+	if dsn == "" {
+		t.Skip("TEST_DATABASE_URL not set")
+	}
+	pool, err := pgxpool.New(context.Background(), dsn)
+	if err != nil {
+		t.Skipf("cannot connect: %v", err)
+	}
+	if err := pool.Ping(context.Background()); err != nil {
+		t.Skipf("cannot ping: %v", err)
+	}
+	t.Cleanup(pool.Close)
+	return pool
+}
+
+func testEnvOr(key, fallback string) string {
+	if v := os.Getenv(key); v != "" {
+		return v
+	}
+	return fallback
+}
+
+type stubAlbums struct{}
+
+func (stubAlbums) Access(ctx context.Context, albumID, userID string) (*models.Album, string, error) {
+	return &models.Album{ID: albumID}, roles.Owner, nil
+}
+
+func createUser(t *testing.T, ctx context.Context, pool *pgxpool.Pool) string {
+	t.Helper()
+	id := auth.NewID()
+	if _, err := pool.Exec(ctx,
+		`INSERT INTO users (id, email, password_hash, name) VALUES ($1, $2, 'x', 't')`,
+		id, id+"@test.local"); err != nil {
+		t.Fatalf("create user: %v", err)
+	}
+	t.Cleanup(func() { _, _ = pool.Exec(ctx, `DELETE FROM users WHERE id = $1`, id) })
+	return id
+}
+
+func createAlbum(t *testing.T, ctx context.Context, pool *pgxpool.Pool, ownerID string) string {
+	t.Helper()
+	id := auth.NewID()
+	if _, err := pool.Exec(ctx,
+		`INSERT INTO albums (id, owner_id, title) VALUES ($1, $2, 't')`, id, ownerID); err != nil {
+		t.Fatalf("create album: %v", err)
+	}
+	t.Cleanup(func() { _, _ = pool.Exec(ctx, `DELETE FROM albums WHERE id = $1`, id) })
+	return id
+}
+
+func createPhoto(t *testing.T, ctx context.Context, pool *pgxpool.Pool, uploaderID string) string {
+	t.Helper()
+	id := auth.NewID()
+	if _, err := pool.Exec(ctx,
+		`INSERT INTO photos (id, album_id, uploader_id) VALUES ($1, NULL, $2)`, id, uploaderID); err != nil {
+		t.Fatalf("create photo: %v", err)
+	}
+	return id
+}
+
+func makePNG(t *testing.T) []byte {
+	t.Helper()
+	img := image.NewRGBA(image.Rect(0, 0, 4, 4))
+	for x := 0; x < 4; x++ {
+		for y := 0; y < 4; y++ {
+			img.Set(x, y, color.RGBA{R: uint8(x * 40), G: uint8(y * 40), B: 128, A: 255})
+		}
+	}
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		t.Fatalf("encode png: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// newTestService builds a service backed by the real dev storage. It skips when
+// S3 env is absent so unit runs stay hermetic.
+func newTestService(t *testing.T, pool *pgxpool.Pool) (*Service, *storage.Client) {
+	t.Helper()
+	endpoint := os.Getenv("S3_ENDPOINT")
+	if endpoint == "" {
+		t.Skip("S3_ENDPOINT not set")
+	}
+	cl, err := storage.New(context.Background(), config.S3{
+		Endpoint:     endpoint,
+		Region:       testEnvOr("S3_REGION", "garage"),
+		AccessKey:    os.Getenv("S3_ACCESS_KEY"),
+		SecretKey:    os.Getenv("S3_SECRET_KEY"),
+		Bucket:       testEnvOr("S3_BUCKET", "galo-data"),
+		UsePathStyle: testEnvOr("S3_USE_PATH_STYLE", "true") != "false",
+		PresignTTL:   time.Hour,
+	})
+	if err != nil {
+		t.Skipf("storage init: %v", err)
+	}
+	return NewService(NewStore(pool), stubAlbums{}, cl, sse.NewHub(1)), cl
+}
+
+// cleanupPhoto removes a photo, its objects (rows and blobs) and bindings.
+func cleanupPhoto(t *testing.T, ctx context.Context, pool *pgxpool.Pool, cl *storage.Client, photoID string) {
+	t.Helper()
+	rows, err := pool.Query(ctx,
+		`SELECT o.id, o.storage_key FROM photo_objects po JOIN objects o ON o.id = po.object_id
+		 WHERE po.photo_id = $1`, photoID)
+	if err != nil {
+		return
+	}
+	var ids, keys []string
+	for rows.Next() {
+		var id, key string
+		if rows.Scan(&id, &key) == nil {
+			ids = append(ids, id)
+			keys = append(keys, key)
+		}
+	}
+	rows.Close()
+	_, _ = pool.Exec(ctx, `DELETE FROM photos WHERE id = $1`, photoID)
+	for _, id := range ids {
+		_, _ = pool.Exec(ctx, `DELETE FROM objects WHERE id = $1`, id)
+	}
+	if cl != nil && len(keys) > 0 {
+		_ = cl.Delete(ctx, keys...)
+	}
+}

backend/internal/photos/keys.go

@@ -0,0 +1,17 @@
+package photos
+
+import (
+	"fmt"
+	"strings"
+)
+
+// objectKey builds an opaque, album-free storage key for a physical object.
+// The key never encodes album, photo or any other metadata entity; the mapping
+// between a key and its photo binding lives only in the metadata layer. The
+// role segment is cosmetic (for debugging) and is not authoritative.
+func objectKey(objectID, role, ext string) string {
+	if ext != "" && !strings.HasPrefix(ext, ".") {
+		ext = "." + ext
+	}
+	return fmt.Sprintf("objects/%s/%s%s", objectID, role, ext)
+}

backend/internal/photos/keys_test.go

@@ -0,0 +1,32 @@
+package photos
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestObjectKeyIsAlbumFree(t *testing.T) {
+	cases := []struct {
+		name string
+		id   string
+		role string
+		ext  string
+		want string
+	}{
+		{"original with dot", "obj1", RoleOriginal, ".jpg", "objects/obj1/original.jpg"},
+		{"thumb without dot", "obj1", RoleThumb, "webp", "objects/obj1/thumb.webp"},
+		{"preview", "obj2", RolePreview, ".webp", "objects/obj2/preview.webp"},
+		{"empty ext", "obj3", RoleOriginal, "", "objects/obj3/original"},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := objectKey(tc.id, tc.role, tc.ext)
+			if got != tc.want {
+				t.Fatalf("objectKey = %q, want %q", got, tc.want)
+			}
+			if strings.Contains(got, "album") {
+				t.Fatalf("key must not encode an album entity: %q", got)
+			}
+		})
+	}
+}

backend/internal/photos/objects.go

@@ -0,0 +1,370 @@
+package photos
+
+import (
+	"context"
+	"errors"
+	"time"
+
+	"github.com/jackc/pgx/v5"
+)
+
+// Object lifecycle states. The DB row is the ledger entry that governs the
+// lifecycle of the corresponding blob in the storage layer.
+const (
+	ObjectPending  = "pending"
+	ObjectReady    = "ready"
+	ObjectDeleting = "deleting"
+)
+
+// Object roles as bound to a photo. The role is a property of the binding.
+const (
+	RoleOriginal = "original"
+	RoleThumb    = "thumb"
+	RolePreview  = "preview"
+)
+
+var (
+	// ErrObjectNotFound is returned when an object row does not exist.
+	ErrObjectNotFound = errors.New("object not found")
+	// ErrObjectNotReady is returned when binding an object that is not ready.
+	ErrObjectNotReady = errors.New("object is not ready")
+)
+
+// Object is a physical file in the storage layer, described by metadata.
+type Object struct {
+	ID         string
+	StorageKey string
+	MimeType   string
+	Size       int64
+	Width      int
+	Height     int
+	Checksum   string
+	State      string
+	CreatedAt  time.Time
+}
+
+// CreateObject inserts a new object row in the pending state.
+func (s *Store) CreateObject(ctx context.Context, o *Object) error {
+	if o.State == "" {
+		o.State = ObjectPending
+	}
+	return s.pool.QueryRow(ctx,
+		`INSERT INTO objects (id, storage_key, mime_type, size, width, height, checksum, state)
+		 VALUES ($1, $2, $3, $4, $5, $6, $7, $8)
+		 RETURNING created_at`,
+		o.ID, o.StorageKey, o.MimeType, o.Size, o.Width, o.Height, o.Checksum, o.State).Scan(&o.CreatedAt)
+}
+
+// SetObjectState updates an object's lifecycle state.
+func (s *Store) SetObjectState(ctx context.Context, id, state string) error {
+	tag, err := s.pool.Exec(ctx, `UPDATE objects SET state = $2 WHERE id = $1`, id, state)
+	if err != nil {
+		return err
+	}
+	if tag.RowsAffected() == 0 {
+		return ErrObjectNotFound
+	}
+	return nil
+}
+
+// ActivateObject marks an object ready after its bytes have been stored.
+func (s *Store) ActivateObject(ctx context.Context, id string) error {
+	return s.SetObjectState(ctx, id, ObjectReady)
+}
+
+// MarkObjectDeleting marks an object as pending removal from storage.
+func (s *Store) MarkObjectDeleting(ctx context.Context, id string) error {
+	return s.SetObjectState(ctx, id, ObjectDeleting)
+}
+
+// DeleteObjectRow permanently removes an object row (after its blob is gone).
+func (s *Store) DeleteObjectRow(ctx context.Context, id string) error {
+	tag, err := s.pool.Exec(ctx, `DELETE FROM objects WHERE id = $1`, id)
+	if err != nil {
+		return err
+	}
+	if tag.RowsAffected() == 0 {
+		return ErrObjectNotFound
+	}
+	return nil
+}
+
+// BindObject attaches a ready object to a photo under the given role. It locks
+// the object row and refuses objects that are not ready, so an object being
+// garbage-collected can never be bound.
+func (s *Store) BindObject(ctx context.Context, photoID, objectID, role string) error {
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	var state string
+	if err := tx.QueryRow(ctx, `SELECT state FROM objects WHERE id = $1 FOR UPDATE`, objectID).Scan(&state); err != nil {
+		if errors.Is(err, pgx.ErrNoRows) {
+			return ErrObjectNotFound
+		}
+		return err
+	}
+	if state != ObjectReady {
+		return ErrObjectNotReady
+	}
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO photo_objects (photo_id, object_id, role) VALUES ($1, $2, $3)
+		 ON CONFLICT (photo_id, role) DO UPDATE SET object_id = EXCLUDED.object_id`,
+		photoID, objectID, role); err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
+}
+
+// PhotoObjectIDs returns the object ids currently bound to a photo.
+func (s *Store) PhotoObjectIDs(ctx context.Context, photoID string) ([]string, error) {
+	rows, err := s.pool.Query(ctx, `SELECT object_id FROM photo_objects WHERE photo_id = $1`, photoID)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var ids []string
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			return nil, err
+		}
+		ids = append(ids, id)
+	}
+	return ids, rows.Err()
+}
+
+// ObjectRefCount counts how many photo bindings reference an object.
+func (s *Store) ObjectRefCount(ctx context.Context, objectID string) (int, error) {
+	var n int
+	err := s.pool.QueryRow(ctx, `SELECT count(*) FROM photo_objects WHERE object_id = $1`, objectID).Scan(&n)
+	return n, err
+}
+
+// OriginalObject returns the original object bound to a photo.
+func (s *Store) OriginalObject(ctx context.Context, photoID string) (*Object, error) {
+	var o Object
+	err := s.pool.QueryRow(ctx,
+		`SELECT o.id, o.storage_key, o.mime_type, o.size, o.width, o.height, o.checksum, o.state, o.created_at
+		 FROM photo_objects po JOIN objects o ON o.id = po.object_id
+		 WHERE po.photo_id = $1 AND po.role = $2`, photoID, RoleOriginal).
+		Scan(&o.ID, &o.StorageKey, &o.MimeType, &o.Size, &o.Width, &o.Height, &o.Checksum, &o.State, &o.CreatedAt)
+	if err != nil {
+		if errors.Is(err, pgx.ErrNoRows) {
+			return nil, ErrObjectNotFound
+		}
+		return nil, err
+	}
+	return &o, nil
+}
+
+// PhotosByOriginalObject returns the ids of all photos bound to the given
+// original object, so generated variants can be shared between copies.
+func (s *Store) PhotosByOriginalObject(ctx context.Context, originalObjectID string) ([]string, error) {
+	rows, err := s.pool.Query(ctx,
+		`SELECT photo_id FROM photo_objects WHERE object_id = $1 AND role = $2`, originalObjectID, RoleOriginal)
+	if err != nil {
+		return nil, err
+	}
+	defer rows.Close()
+	var ids []string
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			return nil, err
+		}
+		ids = append(ids, id)
+	}
+	return ids, rows.Err()
+}
+
+// ClaimGarbage claims collectible objects: ready objects with no bindings and
+// objects already marked deleting. Rows are locked with SKIP LOCKED so
+// concurrent workers do not contend and a concurrent bind cannot slip in.
+func (s *Store) ClaimGarbage(ctx context.Context, limit int) ([]Object, error) {
+	if limit <= 0 {
+		limit = 10
+	}
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	rows, err := tx.Query(ctx,
+		`SELECT o.id, o.storage_key FROM objects o
+		 WHERE o.state = ANY($1)
+		   AND NOT EXISTS (SELECT 1 FROM photo_objects po WHERE po.object_id = o.id)
+		 ORDER BY o.created_at
+		 LIMIT $2
+		 FOR UPDATE SKIP LOCKED`, []string{ObjectReady, ObjectDeleting}, limit)
+	if err != nil {
+		return nil, err
+	}
+	var objs []Object
+	for rows.Next() {
+		var o Object
+		if err := rows.Scan(&o.ID, &o.StorageKey); err != nil {
+			rows.Close()
+			return nil, err
+		}
+		objs = append(objs, o)
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	if len(objs) == 0 {
+		return nil, tx.Commit(ctx)
+	}
+	ids := make([]string, len(objs))
+	for i, o := range objs {
+		ids[i] = o.ID
+	}
+	if _, err := tx.Exec(ctx, `UPDATE objects SET state = $2 WHERE id = ANY($1)`, ids, ObjectDeleting); err != nil {
+		return nil, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return nil, err
+	}
+	return objs, nil
+}
+
+// ClaimStalePending marks abandoned pending objects (whose upload never
+// completed) as deleting and returns them for cleanup.
+func (s *Store) ClaimStalePending(ctx context.Context, olderThan time.Duration, limit int) ([]Object, error) {
+	if limit <= 0 {
+		limit = 10
+	}
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	rows, err := tx.Query(ctx,
+		`SELECT id, storage_key FROM objects
+		 WHERE state = $1 AND created_at < $2
+		 ORDER BY created_at
+		 LIMIT $3
+		 FOR UPDATE SKIP LOCKED`, ObjectPending, time.Now().Add(-olderThan), limit)
+	if err != nil {
+		return nil, err
+	}
+	var objs []Object
+	for rows.Next() {
+		var o Object
+		if err := rows.Scan(&o.ID, &o.StorageKey); err != nil {
+			rows.Close()
+			return nil, err
+		}
+		objs = append(objs, o)
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	if len(objs) == 0 {
+		return nil, tx.Commit(ctx)
+	}
+	ids := make([]string, len(objs))
+	for i, o := range objs {
+		ids[i] = o.ID
+	}
+	if _, err := tx.Exec(ctx, `UPDATE objects SET state = $2 WHERE id = ANY($1)`, ids, ObjectDeleting); err != nil {
+		return nil, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return nil, err
+	}
+	return objs, nil
+}
+
+// ClaimVariantJobs leases due variant jobs and returns their photo ids. The
+// lease pushes next_attempt_at forward so the same job is not picked up twice
+// while it is being processed.
+func (s *Store) ClaimVariantJobs(ctx context.Context, limit int, lease time.Duration) ([]string, error) {
+	if limit <= 0 {
+		limit = 10
+	}
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return nil, err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	rows, err := tx.Query(ctx,
+		`SELECT photo_id FROM variant_jobs
+		 WHERE next_attempt_at <= now()
+		 ORDER BY next_attempt_at
+		 LIMIT $1
+		 FOR UPDATE SKIP LOCKED`, limit)
+	if err != nil {
+		return nil, err
+	}
+	var ids []string
+	for rows.Next() {
+		var id string
+		if err := rows.Scan(&id); err != nil {
+			rows.Close()
+			return nil, err
+		}
+		ids = append(ids, id)
+	}
+	rows.Close()
+	if err := rows.Err(); err != nil {
+		return nil, err
+	}
+	if len(ids) == 0 {
+		return nil, tx.Commit(ctx)
+	}
+	if _, err := tx.Exec(ctx,
+		`UPDATE variant_jobs SET next_attempt_at = $2 WHERE photo_id = ANY($1)`,
+		ids, time.Now().Add(lease)); err != nil {
+		return nil, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return nil, err
+	}
+	return ids, nil
+}
+
+// CompleteVariantJob removes a finished job.
+func (s *Store) CompleteVariantJob(ctx context.Context, photoID string) error {
+	_, err := s.pool.Exec(ctx, `DELETE FROM variant_jobs WHERE photo_id = $1`, photoID)
+	return err
+}
+
+// FailVariantJob records a failure and schedules a retry with exponential
+// backoff. After maxAttempts the job is abandoned so a permanently broken photo
+// cannot loop forever.
+func (s *Store) FailVariantJob(ctx context.Context, photoID, errMsg string, maxAttempts int) error {
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	var attempts int
+	err = tx.QueryRow(ctx,
+		`UPDATE variant_jobs
+		 SET attempts = attempts + 1,
+		     next_attempt_at = now() + make_interval(secs => least(300, 10 * power(2, attempts))),
+		     last_error = $2
+		 WHERE photo_id = $1
+		 RETURNING attempts`, photoID, errMsg).Scan(&attempts)
+	if err != nil {
+		if errors.Is(err, pgx.ErrNoRows) {
+			return tx.Commit(ctx)
+		}
+		return err
+	}
+	if maxAttempts > 0 && attempts >= maxAttempts {
+		if _, err := tx.Exec(ctx, `DELETE FROM variant_jobs WHERE photo_id = $1`, photoID); err != nil {
+			return err
+		}
+	}
+	return tx.Commit(ctx)
+}

backend/internal/photos/objects_test.go

@@ -0,0 +1,187 @@
+package photos
+
+import (
+	"context"
+	"errors"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+
+	"shared-album/internal/auth"
+)
+
+func TestObjectStateTransitions(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	st := NewStore(pool)
+
+	objID := auth.NewID()
+	obj := &Object{ID: objID, StorageKey: "objects/" + objID + "/original.jpg", MimeType: "image/jpeg", Size: 42}
+	t.Cleanup(func() { _, _ = pool.Exec(ctx, `DELETE FROM objects WHERE id = $1`, objID) })
+
+	if err := st.CreateObject(ctx, obj); err != nil {
+		t.Fatalf("create object: %v", err)
+	}
+	if obj.State != ObjectPending {
+		t.Fatalf("new object state = %q, want %q", obj.State, ObjectPending)
+	}
+	if obj.CreatedAt.IsZero() {
+		t.Fatalf("created_at not populated")
+	}
+
+	if err := st.ActivateObject(ctx, objID); err != nil {
+		t.Fatalf("activate: %v", err)
+	}
+	if got := objectState(t, ctx, pool, objID); got != ObjectReady {
+		t.Fatalf("state after activate = %q", got)
+	}
+	if err := st.MarkObjectDeleting(ctx, objID); err != nil {
+		t.Fatalf("mark deleting: %v", err)
+	}
+	if got := objectState(t, ctx, pool, objID); got != ObjectDeleting {
+		t.Fatalf("state after mark = %q", got)
+	}
+	if err := st.DeleteObjectRow(ctx, objID); err != nil {
+		t.Fatalf("delete row: %v", err)
+	}
+	if err := st.DeleteObjectRow(ctx, objID); !errors.Is(err, ErrObjectNotFound) {
+		t.Fatalf("second delete = %v, want ErrObjectNotFound", err)
+	}
+}
+
+func TestClaimStalePending(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	st := NewStore(pool)
+
+	objID := auth.NewID()
+	obj := &Object{ID: objID, StorageKey: "objects/" + objID + "/original.jpg"}
+	t.Cleanup(func() { _, _ = pool.Exec(ctx, `DELETE FROM objects WHERE id = $1`, objID) })
+	if err := st.CreateObject(ctx, obj); err != nil {
+		t.Fatalf("create object: %v", err)
+	}
+	if _, err := pool.Exec(ctx, `UPDATE objects SET created_at = now() - interval '1 hour' WHERE id = $1`, objID); err != nil {
+		t.Fatalf("age object: %v", err)
+	}
+
+	objs, err := st.ClaimStalePending(ctx, 10*time.Minute, 100)
+	if err != nil {
+		t.Fatalf("claim stale: %v", err)
+	}
+	if !containsObject(objs, objID) {
+		t.Fatalf("stale pending object not claimed: %+v", objs)
+	}
+	if got := objectState(t, ctx, pool, objID); got != ObjectDeleting {
+		t.Fatalf("claimed stale object state = %q, want deleting", got)
+	}
+}
+
+func TestBindObjectRequiresReadyAndShares(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	st := NewStore(pool)
+
+	user := createUser(t, ctx, pool)
+	p1 := createPhoto(t, ctx, pool, user)
+	p2 := createPhoto(t, ctx, pool, user)
+
+	objID := auth.NewID()
+	obj := &Object{ID: objID, StorageKey: "objects/" + objID + "/original.jpg"}
+	t.Cleanup(func() {
+		_, _ = pool.Exec(ctx, `DELETE FROM photo_objects WHERE object_id = $1`, objID)
+		_, _ = pool.Exec(ctx, `DELETE FROM objects WHERE id = $1`, objID)
+	})
+	if err := st.CreateObject(ctx, obj); err != nil {
+		t.Fatalf("create object: %v", err)
+	}
+
+	if err := st.BindObject(ctx, p1, objID, RoleOriginal); !errors.Is(err, ErrObjectNotReady) {
+		t.Fatalf("bind pending object = %v, want ErrObjectNotReady", err)
+	}
+	if err := st.ActivateObject(ctx, objID); err != nil {
+		t.Fatalf("activate: %v", err)
+	}
+	if err := st.BindObject(ctx, p1, objID, RoleOriginal); err != nil {
+		t.Fatalf("bind p1: %v", err)
+	}
+	if err := st.BindObject(ctx, p2, objID, RoleOriginal); err != nil {
+		t.Fatalf("bind p2: %v", err)
+	}
+
+	if n, err := st.ObjectRefCount(ctx, objID); err != nil || n != 2 {
+		t.Fatalf("ref count = %d err=%v, want 2", n, err)
+	}
+	siblings, err := st.PhotosByOriginalObject(ctx, objID)
+	if err != nil || len(siblings) != 2 {
+		t.Fatalf("siblings = %v err=%v, want 2", siblings, err)
+	}
+}
+
+func TestClaimGarbageRespectsReferences(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	st := NewStore(pool)
+
+	user := createUser(t, ctx, pool)
+	photo := createPhoto(t, ctx, pool, user)
+
+	boundID := auth.NewID()
+	freeID := auth.NewID()
+	t.Cleanup(func() {
+		_, _ = pool.Exec(ctx, `DELETE FROM photo_objects WHERE object_id IN ($1, $2)`, boundID, freeID)
+		_, _ = pool.Exec(ctx, `DELETE FROM objects WHERE id IN ($1, $2)`, boundID, freeID)
+	})
+	for _, id := range []string{boundID, freeID} {
+		if err := st.CreateObject(ctx, &Object{ID: id, StorageKey: "objects/" + id + "/original.jpg"}); err != nil {
+			t.Fatalf("create object: %v", err)
+		}
+		if err := st.ActivateObject(ctx, id); err != nil {
+			t.Fatalf("activate: %v", err)
+		}
+	}
+	if err := st.BindObject(ctx, photo, boundID, RoleOriginal); err != nil {
+		t.Fatalf("bind: %v", err)
+	}
+
+	claimed, err := st.ClaimGarbage(ctx, 100)
+	if err != nil {
+		t.Fatalf("claim garbage: %v", err)
+	}
+	if !containsObject(claimed, freeID) {
+		t.Fatalf("unreferenced object not claimed: %+v", claimed)
+	}
+	if containsObject(claimed, boundID) {
+		t.Fatalf("referenced object was claimed for deletion")
+	}
+
+	// Removing the last reference makes both collectible.
+	if _, err := pool.Exec(ctx, `DELETE FROM photos WHERE id = $1`, photo); err != nil {
+		t.Fatalf("delete photo: %v", err)
+	}
+	claimed, err = st.ClaimGarbage(ctx, 100)
+	if err != nil {
+		t.Fatalf("claim garbage 2: %v", err)
+	}
+	if !containsObject(claimed, boundID) {
+		t.Fatalf("object with no remaining references not claimed: %+v", claimed)
+	}
+}
+
+func objectState(t *testing.T, ctx context.Context, pool *pgxpool.Pool, id string) string {
+	t.Helper()
+	var state string
+	if err := pool.QueryRow(ctx, `SELECT state FROM objects WHERE id = $1`, id).Scan(&state); err != nil {
+		t.Fatalf("read state: %v", err)
+	}
+	return state
+}
+
+func containsObject(objs []Object, id string) bool {
+	for _, o := range objs {
+		if o.ID == id {
+			return true
+		}
+	}
+	return false
+}

backend/internal/photos/pipeline.go

@@ -2,6 +2,8 @@ package photos
 
 import (
 	"bytes"
+	"crypto/sha256"
+	"encoding/hex"
 	"fmt"
 	"image"
 	"strconv"
@@ -138,3 +140,9 @@ func buildVariants(data []byte) (thumb, preview []byte, err error) {
 	}
 	return thumbBuf.Bytes(), previewBuf.Bytes(), nil
 }
+
+// checksum returns the hex-encoded SHA-256 of the given bytes.
+func checksum(data []byte) string {
+	sum := sha256.Sum256(data)
+	return hex.EncodeToString(sum[:])
+}

backend/internal/photos/service.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"errors"
-	"fmt"
 	"net/http"
 	"path/filepath"
 	"strings"
@@ -64,22 +63,20 @@ func (s *Service) Upload(ctx context.Context, in UploadInput) (*models.Photo, er
 	}
 	meta := extractEXIF(in.Data)
 
-	id := auth.NewID()
 	ext := strings.ToLower(filepath.Ext(in.Filename))
 	if ext == "" {
 		ext = ".bin"
 	}
-	storageKey := fmt.Sprintf("albums/%s/original/%s%s", in.AlbumID, id, ext)
 
-	if err := s.storage.Put(ctx, storageKey, bytes.NewReader(in.Data), int64(len(in.Data)), mime); err != nil {
-		return nil, err
-	}
+	photoID := auth.NewID()
+	objectID := auth.NewID()
+	key := objectKey(objectID, RoleOriginal, ext)
 
 	p := &models.Photo{
-		ID:               id,
+		ID:               photoID,
 		AlbumID:          in.AlbumID,
 		UploaderID:       in.UserID,
-		StorageKey:       storageKey,
+		StorageKey:       key,
 		OriginalFilename: in.Filename,
 		Size:             int64(len(in.Data)),
 		Width:            width,
@@ -90,12 +87,28 @@ func (s *Service) Upload(ctx context.Context, in UploadInput) (*models.Photo, er
 		Longitude:        meta.Longitude,
 		EXIF:             meta.Attributes,
 	}
-	if err := s.store.Create(ctx, p); err != nil {
-		return nil, err
+	obj := &Object{
+		ID:         objectID,
+		StorageKey: key,
+		MimeType:   mime,
+		Size:       int64(len(in.Data)),
+		Width:      width,
+		Height:     height,
+		Checksum:   checksum(in.Data),
 	}
 
-	data := in.Data
-	go s.processVariants(p.ID, in.AlbumID, data)
+	// The metadata layer is the ledger of record: the row is written first,
+	// then the bytes are stored, then the object is activated.
+	if err := s.store.CreateWithOriginal(ctx, p, obj); err != nil {
+		return nil, err
+	}
+	if err := s.storage.Put(ctx, key, bytes.NewReader(in.Data), int64(len(in.Data)), mime); err != nil {
+		_ = s.store.DeleteUpload(ctx, photoID, objectID)
+		return nil, err
+	}
+	if err := s.store.ActivateObject(ctx, objectID); err != nil {
+		return nil, err
+	}
 
 	if err := s.attachURLs(ctx, p); err != nil {
 		return nil, err
@@ -103,33 +116,6 @@ func (s *Service) Upload(ctx context.Context, in UploadInput) (*models.Photo, er
 	return p, nil
 }
 
-func (s *Service) processVariants(photoID, albumID string, data []byte) {
-	ctx, cancel := context.WithTimeout(context.Background(), 60*time.Second)
-	defer cancel()
-
-	thumb, preview, err := buildVariants(data)
-	if err != nil {
-		return
-	}
-	thumbKey := fmt.Sprintf("albums/%s/thumb/%s.webp", albumID, photoID)
-	previewKey := fmt.Sprintf("albums/%s/preview/%s.webp", albumID, photoID)
-	if err := s.storage.Put(ctx, thumbKey, bytes.NewReader(thumb), int64(len(thumb)), "image/webp"); err != nil {
-		return
-	}
-	if err := s.storage.Put(ctx, previewKey, bytes.NewReader(preview), int64(len(preview)), "image/webp"); err != nil {
-		return
-	}
-	if err := s.store.SetVariants(ctx, photoID, thumbKey, previewKey); err != nil {
-		return
-	}
-	s.store.SetCoverIfEmpty(ctx, albumID, photoID)
-	if p, err := s.store.GetByID(ctx, photoID, ""); err == nil {
-		if err := s.attachURLs(ctx, p); err == nil {
-			s.hub.Publish(albumID, "photo.created", p)
-		}
-	}
-}
-
 func (s *Service) attachURLs(ctx context.Context, p *models.Photo) error {
 	if p.StorageKey != "" {
 		url, err := s.storage.PresignGet(ctx, p.StorageKey)
@@ -416,21 +402,22 @@ func (s *Service) PermanentDelete(ctx context.Context, photoID, userID string) e
 	if !canDeletePhoto(role, p.UploaderID == userID) {
 		return httpx.Forbidden("forbidden", "you cannot delete this photo")
 	}
+	objectIDs, err := s.store.PhotoObjectIDs(ctx, photoID)
+	if err != nil {
+		return err
+	}
 	if err := s.store.HardDelete(ctx, photoID); err != nil {
 		return err
 	}
 	if p.AlbumID != "" {
 		s.publishDeleted(p)
 	}
-	if n, err := s.store.StorageKeyRefCount(ctx, p.StorageKey); err == nil && n == 0 {
-		keys := []string{p.StorageKey}
-		if p.ThumbKey != "" {
-			keys = append(keys, p.ThumbKey)
+	// Mark objects that no longer have any binding for garbage collection.
+	// The background worker removes the blobs from storage.
+	for _, oid := range objectIDs {
+		if n, err := s.store.ObjectRefCount(ctx, oid); err == nil && n == 0 {
+			_ = s.store.MarkObjectDeleting(ctx, oid)
 		}
-		if p.PreviewKey != "" {
-			keys = append(keys, p.PreviewKey)
-		}
-		_ = s.storage.Delete(ctx, keys...)
 	}
 	return nil
 }

backend/internal/photos/store.go

@@ -23,9 +23,14 @@ type Store struct {
 
 func NewStore(pool *pgxpool.Pool) *Store { return &Store{pool: pool} }
 
-const photoColumns = `p.id, COALESCE(p.album_id, ''), p.uploader_id, p.storage_key,
-	COALESCE(p.thumb_key, ''), COALESCE(p.preview_key, ''),
-	p.original_filename, p.size, p.width, p.height, p.mime_type,
+// photoColumns selects photo rows together with the keys and physical facts of
+// their bound objects (original, thumb, preview) and the original's size,
+// dimensions and mime type.
+const photoColumns = `p.id, COALESCE(p.album_id, ''), p.uploader_id,
+	COALESCE(o_o.storage_key, ''),
+	COALESCE(o_t.storage_key, ''),
+	COALESCE(o_p.storage_key, ''),
+	p.original_filename, COALESCE(o_o.size, 0), COALESCE(o_o.width, 0), COALESCE(o_o.height, 0), COALESCE(o_o.mime_type, ''),
 	p.taken_at, p.latitude, p.longitude, COALESCE(p.caption, ''), p.created_at,
 	u.id, u.name, COALESCE(u.avatar_url, ''),
 	(SELECT count(*) FROM likes l WHERE l.photo_id = p.id),
@@ -36,6 +41,14 @@ const photoColumns = `p.id, COALESCE(p.album_id, ''), p.uploader_id, p.storage_k
 		FROM photo_tags pt JOIN tags t ON t.id = pt.tag_id
 		WHERE pt.photo_id = p.id AND t.user_id = $2), '[]')::text`
 
+const photoJoins = `
+	LEFT JOIN photo_objects po_o ON po_o.photo_id = p.id AND po_o.role = 'original'
+	LEFT JOIN objects o_o ON o_o.id = po_o.object_id
+	LEFT JOIN photo_objects po_t ON po_t.photo_id = p.id AND po_t.role = 'thumb'
+	LEFT JOIN objects o_t ON o_t.id = po_t.object_id
+	LEFT JOIN photo_objects po_p ON po_p.photo_id = p.id AND po_p.role = 'preview'
+	LEFT JOIN objects o_p ON o_p.id = po_p.object_id`
+
 func scanPhoto(row pgx.Row) (*models.Photo, error) {
 	var p models.Photo
 	var exifText, tagsText string
@@ -76,31 +89,68 @@ func setExif(p *models.Photo, raw string) {
 	}
 }
 
-func (s *Store) Create(ctx context.Context, p *models.Photo) error {
+// CreateWithOriginal records a new photo together with its pending original
+// object, the original binding and a variant job — all in one transaction. The
+// caller stores the bytes and then activates the object. The metadata layer is
+// the ledger of record; the storage layer follows.
+func (s *Store) CreateWithOriginal(ctx context.Context, p *models.Photo, obj *Object) error {
 	exifJSON, err := json.Marshal(p.EXIF)
 	if err != nil || len(exifJSON) == 0 {
 		exifJSON = []byte("{}")
 	}
-	return s.pool.QueryRow(ctx,
-		`INSERT INTO photos (id, album_id, uploader_id, storage_key, original_filename,
-			size, width, height, mime_type, taken_at, latitude, longitude, caption, exif)
-		 VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14::jsonb)
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO objects (id, storage_key, mime_type, size, width, height, checksum, state)
+		 VALUES ($1, $2, $3, $4, $5, $6, $7, $8)`,
+		obj.ID, obj.StorageKey, obj.MimeType, obj.Size, obj.Width, obj.Height, obj.Checksum, ObjectPending); err != nil {
+		return err
+	}
+	if err := tx.QueryRow(ctx,
+		`INSERT INTO photos (id, album_id, uploader_id, original_filename, taken_at, latitude, longitude, caption, exif)
+		 VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9::jsonb)
 		 RETURNING created_at`,
-		p.ID, p.AlbumID, p.UploaderID, p.StorageKey, p.OriginalFilename,
-		p.Size, p.Width, p.Height, p.MimeType, p.TakenAt, p.Latitude, p.Longitude, p.Caption, string(exifJSON),
-	).Scan(&p.CreatedAt)
+		p.ID, p.AlbumID, p.UploaderID, p.OriginalFilename, p.TakenAt, p.Latitude, p.Longitude, p.Caption, string(exifJSON),
+	).Scan(&p.CreatedAt); err != nil {
+		return err
+	}
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO photo_objects (photo_id, object_id, role) VALUES ($1, $2, $3)`,
+		p.ID, obj.ID, RoleOriginal); err != nil {
+		return err
+	}
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO variant_jobs (photo_id) VALUES ($1) ON CONFLICT (photo_id) DO NOTHING`, p.ID); err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
 }
 
-func (s *Store) SetVariants(ctx context.Context, id, thumbKey, previewKey string) error {
-	_, err := s.pool.Exec(ctx,
-		`UPDATE photos SET thumb_key = $2, preview_key = $3 WHERE id = $1`, id, thumbKey, previewKey)
-	return err
+// DeleteUpload rolls back a failed upload: the photo (and its bindings and job)
+// plus the pending object row are removed.
+func (s *Store) DeleteUpload(ctx context.Context, photoID, objectID string) error {
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+	if _, err := tx.Exec(ctx, `DELETE FROM photos WHERE id = $1`, photoID); err != nil {
+		return err
+	}
+	if _, err := tx.Exec(ctx, `DELETE FROM objects WHERE id = $1`, objectID); err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
 }
 
 func (s *Store) GetByID(ctx context.Context, id, viewerID string) (*models.Photo, error) {
 	return scanPhoto(s.pool.QueryRow(ctx,
 		`SELECT `+photoColumns+`
-		 FROM photos p JOIN users u ON u.id = p.uploader_id
+		 FROM photos p JOIN users u ON u.id = p.uploader_id`+photoJoins+`
 		 WHERE p.id = $1`, id, viewerID))
 }
 
@@ -118,7 +168,7 @@ func (s *Store) List(ctx context.Context, in ListParams) ([]models.Photo, error)
 	}
 	rows, err := s.pool.Query(ctx,
 		`SELECT `+photoColumns+`
-		 FROM photos p JOIN users u ON u.id = p.uploader_id
+		 FROM photos p JOIN users u ON u.id = p.uploader_id`+photoJoins+`
 		 WHERE p.album_id = $1
 		   AND p.deleted_at IS NULL
 		   AND ($3 = '' OR (p.created_at, p.id) < (SELECT created_at, id FROM photos WHERE id = $3))
@@ -181,30 +231,36 @@ func (s *Store) UpdateAlbum(ctx context.Context, id, albumID string) error {
 	return nil
 }
 
-// Clone creates a new photo row that references the same stored objects as src.
+// Clone creates a new photo row in albumID that references the same objects as
+// src. No storage operation is performed: the clone shares the physical files.
 func (s *Store) Clone(ctx context.Context, src *models.Photo, albumID, uploaderID, id string) error {
 	exifJSON, err := json.Marshal(src.EXIF)
 	if err != nil || len(exifJSON) == 0 {
 		exifJSON = []byte("{}")
 	}
-	_, err = s.pool.Exec(ctx,
-		`INSERT INTO photos (id, album_id, uploader_id, storage_key, thumb_key, preview_key,
-			original_filename, size, width, height, mime_type, taken_at, latitude, longitude, caption, exif)
-		 VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16::jsonb)`,
-		id, albumID, uploaderID, src.StorageKey, nullIfEmpty(src.ThumbKey), nullIfEmpty(src.PreviewKey),
-		src.OriginalFilename, src.Size, src.Width, src.Height, src.MimeType,
-		src.TakenAt, src.Latitude, src.Longitude, src.Caption, string(exifJSON))
-	return err
+	tx, err := s.pool.Begin(ctx)
+	if err != nil {
+		return err
+	}
+	defer func() { _ = tx.Rollback(ctx) }()
+
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO photos (id, album_id, uploader_id, original_filename, taken_at, latitude, longitude, caption, exif)
+		 VALUES ($1, $2, $3, $4, $5, $6, $7, $8, $9::jsonb)`,
+		id, albumID, uploaderID, src.OriginalFilename, src.TakenAt, src.Latitude, src.Longitude, src.Caption, string(exifJSON)); err != nil {
+		return err
+	}
+	if _, err := tx.Exec(ctx,
+		`INSERT INTO photo_objects (photo_id, object_id, role)
+		 SELECT $1, object_id, role FROM photo_objects WHERE photo_id = $2`,
+		id, src.ID); err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
 }
 
-// StorageKeyRefCount counts photo rows (any state) referencing the stored object.
-func (s *Store) StorageKeyRefCount(ctx context.Context, storageKey string) (int, error) {
-	var n int
-	err := s.pool.QueryRow(ctx, `SELECT count(*) FROM photos WHERE storage_key = $1`, storageKey).Scan(&n)
-	return n, err
-}
-
-// HardDelete permanently removes the photo row.
+// HardDelete permanently removes the photo row. Bound objects are not touched
+// here; the caller marks orphaned objects for collection.
 func (s *Store) HardDelete(ctx context.Context, id string) error {
 	tag, err := s.pool.Exec(ctx, `DELETE FROM photos WHERE id = $1`, id)
 	if err != nil {
@@ -223,7 +279,7 @@ func (s *Store) ListTrash(ctx context.Context, userID string) ([]models.Photo, e
 		`SELECT `+photoColumns+`, COALESCE(a.title, '')
 		 FROM photos p
 		 JOIN users u ON u.id = p.uploader_id
-		 LEFT JOIN albums a ON a.id = p.album_id
+		 LEFT JOIN albums a ON a.id = p.album_id`+photoJoins+`
 		 WHERE p.deleted_at IS NOT NULL
 		   AND (
 		   	(a.id IS NOT NULL AND (a.owner_id = $1 OR EXISTS(
@@ -362,13 +418,6 @@ func (s *Store) TagOwnedBy(ctx context.Context, tagID, userID string) (bool, err
 	return ok, err
 }
 
-func nullIfEmpty(s string) *string {
-	if s == "" {
-		return nil
-	}
-	return &s
-}
-
 func (s *Store) SetCoverIfEmpty(ctx context.Context, albumID, photoID string) {
 	_, _ = s.pool.Exec(ctx,
 		`UPDATE albums SET cover_photo_id = $2 WHERE id = $1 AND cover_photo_id IS NULL`,
@@ -391,11 +440,16 @@ type backfillRow struct {
 	StorageKey string
 }
 
+// BackfillEXIF re-reads the original object of every photo with empty exif and
+// fills the exif/geo fields on the photo.
 func (s *Store) BackfillEXIF(ctx context.Context, fetch func(context.Context, string) ([]byte, error)) (int, error) {
 	rows, err := s.pool.Query(ctx,
-		`SELECT id, storage_key FROM photos
-		 WHERE exif IS NULL OR exif = '{}'::jsonb
-		 ORDER BY created_at`)
+		`SELECT p.id, o.storage_key
+		 FROM photos p
+		 JOIN photo_objects po ON po.photo_id = p.id AND po.role = 'original'
+		 JOIN objects o ON o.id = po.object_id
+		 WHERE p.exif IS NULL OR p.exif = '{}'::jsonb
+		 ORDER BY p.created_at`)
 	if err != nil {
 		return 0, err
 	}

backend/internal/photos/upload_test.go

@@ -0,0 +1,70 @@
+package photos
+
+import (
+	"context"
+	"strings"
+	"testing"
+)
+
+func TestUploadCreatesLedgerAndAlbumFreeKey(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	svc, cl := newTestService(t, pool)
+	user := createUser(t, ctx, pool)
+	album := createAlbum(t, ctx, pool, user)
+
+	p, err := svc.Upload(ctx, UploadInput{
+		AlbumID:  album,
+		UserID:   user,
+		Filename: "photo.png",
+		MimeType: "image/png",
+		Data:     makePNG(t),
+	})
+	if err != nil {
+		t.Fatalf("upload: %v", err)
+	}
+	t.Cleanup(func() { cleanupPhoto(t, ctx, pool, cl, p.ID) })
+
+	if !strings.HasPrefix(p.StorageKey, "objects/") {
+		t.Fatalf("storage key %q is not an object key", p.StorageKey)
+	}
+	if strings.Contains(p.StorageKey, album) {
+		t.Fatalf("storage key leaks album id: %q", p.StorageKey)
+	}
+	if p.OriginalURL == "" {
+		t.Fatalf("original url not set")
+	}
+	if p.ThumbURL != "" || p.PreviewURL != "" {
+		t.Fatalf("variants should be pending, got thumb=%q preview=%q", p.ThumbURL, p.PreviewURL)
+	}
+
+	orig, err := svc.store.OriginalObject(ctx, p.ID)
+	if err != nil {
+		t.Fatalf("original object: %v", err)
+	}
+	if orig.State != ObjectReady {
+		t.Fatalf("original object state = %q, want ready", orig.State)
+	}
+	if n, err := svc.store.ObjectRefCount(ctx, orig.ID); err != nil || n != 1 {
+		t.Fatalf("original ref count = %d err=%v, want 1", n, err)
+	}
+
+	var jobs int
+	if err := pool.QueryRow(ctx, `SELECT count(*) FROM variant_jobs WHERE photo_id = $1`, p.ID).Scan(&jobs); err != nil {
+		t.Fatalf("count jobs: %v", err)
+	}
+	if jobs != 1 {
+		t.Fatalf("variant jobs = %d, want 1", jobs)
+	}
+
+	list, _, err := svc.List(ctx, album, user, "", nil, 50)
+	if err != nil || len(list) != 1 {
+		t.Fatalf("list: %v len=%d", err, len(list))
+	}
+	if list[0].MimeType != "image/png" || list[0].Width != 4 || list[0].Height != 4 {
+		t.Fatalf("derived fields wrong: mime=%q %dx%d", list[0].MimeType, list[0].Width, list[0].Height)
+	}
+	if list[0].OriginalURL == "" {
+		t.Fatalf("list original url not set")
+	}
+}

backend/internal/photos/worker.go

@@ -0,0 +1,210 @@
+package photos
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"log/slog"
+	"time"
+
+	"shared-album/internal/auth"
+)
+
+const maxVariantAttempts = 5
+
+// Worker is the background reconciler. It drives the variant generation queue
+// and the object garbage collector entirely from the state recorded in the
+// metadata layer, so it resumes correctly after a restart.
+type Worker struct {
+	svc        *Service
+	interval   time.Duration
+	batch      int
+	pendingTTL time.Duration
+	lease      time.Duration
+}
+
+// NewWorker builds a worker. Parameters are clamped to sane defaults.
+func NewWorker(svc *Service, interval time.Duration, batch int, pendingTTL time.Duration) *Worker {
+	if interval <= 0 {
+		interval = 5 * time.Second
+	}
+	if batch <= 0 {
+		batch = 10
+	}
+	if pendingTTL <= 0 {
+		pendingTTL = 10 * time.Minute
+	}
+	return &Worker{
+		svc:        svc,
+		interval:   interval,
+		batch:      batch,
+		pendingTTL: pendingTTL,
+		lease:      interval * 4,
+	}
+}
+
+// Run loops until the context is cancelled.
+func (w *Worker) Run(ctx context.Context) {
+	slog.Info("background worker started", "interval", w.interval)
+	defer slog.Info("background worker stopped")
+	ticker := time.NewTicker(w.interval)
+	defer ticker.Stop()
+	for {
+		w.tick(ctx)
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+		}
+	}
+}
+
+func (w *Worker) tick(ctx context.Context) {
+	w.processVariantJobs(ctx)
+	w.collectGarbage(ctx)
+	w.reapStalePending(ctx)
+}
+
+func (w *Worker) processVariantJobs(ctx context.Context) {
+	ids, err := w.svc.store.ClaimVariantJobs(ctx, w.batch, w.lease)
+	if err != nil {
+		slog.Error("claim variant jobs", "error", err)
+		return
+	}
+	for _, id := range ids {
+		w.generateVariants(ctx, id)
+	}
+}
+
+func (w *Worker) generateVariants(ctx context.Context, photoID string) {
+	orig, err := w.svc.store.OriginalObject(ctx, photoID)
+	if err != nil {
+		// The photo or its original binding is gone; nothing to do.
+		_ = w.svc.store.CompleteVariantJob(ctx, photoID)
+		return
+	}
+	if orig.State != ObjectReady {
+		w.fail(ctx, photoID, "original object is not ready")
+		return
+	}
+	data, err := w.svc.storage.Get(ctx, orig.StorageKey)
+	if err != nil {
+		w.fail(ctx, photoID, "fetch original: "+err.Error())
+		return
+	}
+	thumb, preview, err := buildVariants(data)
+	if err != nil {
+		w.fail(ctx, photoID, "build variants: "+err.Error())
+		return
+	}
+	siblings, err := w.svc.store.PhotosByOriginalObject(ctx, orig.ID)
+	if err != nil || len(siblings) == 0 {
+		siblings = []string{photoID}
+	}
+	if err := w.createVariant(ctx, siblings, RoleThumb, thumb); err != nil {
+		w.fail(ctx, photoID, "store thumb: "+err.Error())
+		return
+	}
+	if err := w.createVariant(ctx, siblings, RolePreview, preview); err != nil {
+		w.fail(ctx, photoID, "store preview: "+err.Error())
+		return
+	}
+	_ = w.svc.store.CompleteVariantJob(ctx, photoID)
+	for _, pid := range siblings {
+		w.notify(ctx, pid)
+	}
+}
+
+// createVariant stores one derived object and binds it to every photo sharing
+// the original, so copies reuse the same physical variant.
+func (w *Worker) createVariant(ctx context.Context, photoIDs []string, role string, data []byte) error {
+	objID := auth.NewID()
+	key := objectKey(objID, role, ".webp")
+	obj := &Object{
+		ID:         objID,
+		StorageKey: key,
+		MimeType:   "image/webp",
+		Size:       int64(len(data)),
+		Checksum:   checksum(data),
+	}
+	if err := w.svc.store.CreateObject(ctx, obj); err != nil {
+		return err
+	}
+	if err := w.svc.storage.Put(ctx, key, bytes.NewReader(data), int64(len(data)), "image/webp"); err != nil {
+		_ = w.svc.store.DeleteObjectRow(ctx, objID)
+		return err
+	}
+	if err := w.svc.store.ActivateObject(ctx, objID); err != nil {
+		return err
+	}
+	bound := false
+	for _, pid := range photoIDs {
+		if err := w.svc.store.BindObject(ctx, pid, objID, role); err != nil {
+			if errors.Is(err, ErrObjectNotReady) || errors.Is(err, ErrObjectNotFound) {
+				continue
+			}
+			return err
+		}
+		bound = true
+	}
+	if !bound {
+		// No photo could take the variant; let the collector reclaim it.
+		_ = w.svc.store.MarkObjectDeleting(ctx, objID)
+	}
+	return nil
+}
+
+func (w *Worker) notify(ctx context.Context, photoID string) {
+	p, err := w.svc.store.GetByID(ctx, photoID, "")
+	if err != nil {
+		return
+	}
+	if err := w.svc.attachURLs(ctx, p); err != nil {
+		return
+	}
+	w.svc.hub.Publish(p.AlbumID, "photo.created", p)
+}
+
+func (w *Worker) fail(ctx context.Context, photoID, msg string) {
+	slog.Warn("variant job failed", "photo", photoID, "error", msg)
+	_ = w.svc.store.FailVariantJob(ctx, photoID, msg, maxVariantAttempts)
+}
+
+func (w *Worker) collectGarbage(ctx context.Context) {
+	objs, err := w.svc.store.ClaimGarbage(ctx, w.batch)
+	if err != nil {
+		slog.Error("claim garbage", "error", err)
+		return
+	}
+	for _, o := range objs {
+		w.deleteObject(ctx, o)
+	}
+}
+
+func (w *Worker) reapStalePending(ctx context.Context) {
+	objs, err := w.svc.store.ClaimStalePending(ctx, w.pendingTTL, w.batch)
+	if err != nil {
+		slog.Error("claim stale pending", "error", err)
+		return
+	}
+	for _, o := range objs {
+		w.deleteObject(ctx, o)
+	}
+}
+
+// deleteObject removes an object's blob then its row. If a binding appeared
+// meanwhile the object is returned to ready; if the storage delete fails the
+// row stays deleting and is retried on the next pass.
+func (w *Worker) deleteObject(ctx context.Context, o Object) {
+	if n, err := w.svc.store.ObjectRefCount(ctx, o.ID); err == nil && n > 0 {
+		_ = w.svc.store.SetObjectState(ctx, o.ID, ObjectReady)
+		return
+	}
+	if err := w.svc.storage.Delete(ctx, o.StorageKey); err != nil {
+		slog.Warn("delete object blob", "object", o.ID, "error", err)
+		return
+	}
+	if err := w.svc.store.DeleteObjectRow(ctx, o.ID); err != nil && !errors.Is(err, ErrObjectNotFound) {
+		slog.Warn("delete object row", "object", o.ID, "error", err)
+	}
+}

backend/internal/photos/worker_test.go

@@ -0,0 +1,79 @@
+package photos
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"github.com/jackc/pgx/v5/pgxpool"
+)
+
+func TestWorkerGeneratesVariantsAndCollects(t *testing.T) {
+	pool := testPool(t)
+	ctx := context.Background()
+	svc, cl := newTestService(t, pool)
+	user := createUser(t, ctx, pool)
+	album := createAlbum(t, ctx, pool, user)
+
+	p, err := svc.Upload(ctx, UploadInput{
+		AlbumID:  album,
+		UserID:   user,
+		Filename: "photo.png",
+		MimeType: "image/png",
+		Data:     makePNG(t),
+	})
+	if err != nil {
+		t.Fatalf("upload: %v", err)
+	}
+	t.Cleanup(func() { cleanupPhoto(t, ctx, pool, cl, p.ID) })
+
+	w := NewWorker(svc, time.Second, 10, time.Minute)
+	w.processVariantJobs(ctx)
+
+	refreshed, err := svc.store.GetByID(ctx, p.ID, "")
+	if err != nil {
+		t.Fatalf("get refreshed: %v", err)
+	}
+	if refreshed.ThumbKey == "" || refreshed.PreviewKey == "" {
+		t.Fatalf("variants not bound: thumb=%q preview=%q", refreshed.ThumbKey, refreshed.PreviewKey)
+	}
+	if err := svc.attachURLs(ctx, refreshed); err != nil {
+		t.Fatalf("attach urls: %v", err)
+	}
+	if refreshed.ThumbURL == "" || refreshed.PreviewURL == "" {
+		t.Fatalf("variant urls not set")
+	}
+
+	var jobs int
+	if err := pool.QueryRow(ctx, `SELECT count(*) FROM variant_jobs WHERE photo_id = $1`, p.ID).Scan(&jobs); err != nil {
+		t.Fatalf("count jobs: %v", err)
+	}
+	if jobs != 0 {
+		t.Fatalf("variant job not completed: %d left", jobs)
+	}
+
+	objectIDs, err := svc.store.PhotoObjectIDs(ctx, p.ID)
+	if err != nil || len(objectIDs) != 3 {
+		t.Fatalf("expected 3 bound objects, got %v err=%v", objectIDs, err)
+	}
+
+	if err := svc.PermanentDelete(ctx, p.ID, user); err != nil {
+		t.Fatalf("permanent delete: %v", err)
+	}
+	// Run the collector until the objects are gone (bounded to avoid hangs).
+	for attempt := 0; attempt < 100 && countObjects(t, ctx, pool, objectIDs) > 0; attempt++ {
+		w.collectGarbage(ctx)
+	}
+	if n := countObjects(t, ctx, pool, objectIDs); n != 0 {
+		t.Fatalf("%d objects were not collected", n)
+	}
+}
+
+func countObjects(t *testing.T, ctx context.Context, pool *pgxpool.Pool, ids []string) int {
+	t.Helper()
+	var n int
+	if err := pool.QueryRow(ctx, `SELECT count(*) FROM objects WHERE id = ANY($1)`, ids).Scan(&n); err != nil {
+		t.Fatalf("count objects: %v", err)
+	}
+	return n
+}

backend/internal/repo/repo_test.go

@@ -29,6 +29,7 @@ func testPool(t *testing.T) *pgxpool.Pool {
 	if err := pool.Ping(context.Background()); err != nil {
 		t.Skipf("cannot ping: %v", err)
 	}
+	t.Cleanup(pool.Close)
 	return pool
 }
 
@@ -48,7 +49,6 @@ func newUser(t *testing.T, ctx context.Context, pool *pgxpool.Pool, name string)
 
 func TestUserRepository(t *testing.T) {
 	pool := testPool(t)
-	defer pool.Close()
 	ctx := context.Background()
 	store := auth.NewStore(pool)
 
@@ -82,7 +82,6 @@ func TestUserRepository(t *testing.T) {
 
 func TestAlbumRepository(t *testing.T) {
 	pool := testPool(t)
-	defer pool.Close()
 	ctx := context.Background()
 	store := albums.NewStore(pool)
 
@@ -126,7 +125,6 @@ func TestAlbumRepository(t *testing.T) {
 
 func TestAlbumCoverRepository(t *testing.T) {
 	pool := testPool(t)
-	defer pool.Close()
 	ctx := context.Background()
 	store := albums.NewStore(pool)
 
@@ -138,18 +136,33 @@ func TestAlbumCoverRepository(t *testing.T) {
 		t.Fatalf("create album: %v", err)
 	}
 
-	insertPhoto := func(id, thumbKey string) {
+	insertPhoto := func(id string) string {
 		t.Helper()
+		origObj := auth.NewID()
+		thumbObj := auth.NewID()
+		thumbKey := "objects/" + thumbObj + "/thumb.webp"
 		if _, err := pool.Exec(ctx,
-			`INSERT INTO photos (id, album_id, uploader_id, storage_key, thumb_key)
-			 VALUES ($1, $2, $3, $4, $5)`,
-			id, a.ID, owner.ID, "albums/"+a.ID+"/original/"+id+".jpg", thumbKey); err != nil {
+			`INSERT INTO objects (id, storage_key, mime_type, state) VALUES
+			 ($1, $2, 'image/jpeg', 'ready'),
+			 ($3, $4, 'image/webp', 'ready')`,
+			origObj, "objects/"+origObj+"/original.jpg", thumbObj, thumbKey); err != nil {
+			t.Fatalf("insert objects: %v", err)
+		}
+		if _, err := pool.Exec(ctx,
+			`INSERT INTO photos (id, album_id, uploader_id) VALUES ($1, $2, $3)`,
+			id, a.ID, owner.ID); err != nil {
 			t.Fatalf("insert photo: %v", err)
 		}
+		if _, err := pool.Exec(ctx,
+			`INSERT INTO photo_objects (photo_id, object_id, role) VALUES ($1, $2, 'original'), ($1, $3, 'thumb')`,
+			id, origObj, thumbObj); err != nil {
+			t.Fatalf("insert bindings: %v", err)
+		}
+		return thumbKey
 	}
 
 	photo1 := auth.NewID()
-	insertPhoto(photo1, "albums/"+a.ID+"/thumb/"+photo1+".webp")
+	thumb1 := insertPhoto(photo1)
 
 	got, err := store.GetByID(ctx, a.ID)
 	if err != nil || got.CoverThumbKey != "" {
@@ -162,12 +175,12 @@ func TestAlbumCoverRepository(t *testing.T) {
 	if got.CoverPhotoID == nil || *got.CoverPhotoID != photo1 {
 		t.Fatalf("cover not set: %+v", got.CoverPhotoID)
 	}
-	if want := "albums/" + a.ID + "/thumb/" + photo1 + ".webp"; got.CoverThumbKey != want {
-		t.Fatalf("cover thumb key = %q want %q", got.CoverThumbKey, want)
+	if got.CoverThumbKey != thumb1 {
+		t.Fatalf("cover thumb key = %q want %q", got.CoverThumbKey, thumb1)
 	}
 
 	photo2 := auth.NewID()
-	insertPhoto(photo2, "thumb-2")
+	thumb2 := insertPhoto(photo2)
 	ps.SetCoverIfEmpty(ctx, a.ID, photo2)
 	got, _ = store.GetByID(ctx, a.ID)
 	if got.CoverPhotoID == nil || *got.CoverPhotoID != photo1 {
@@ -184,14 +197,13 @@ func TestAlbumCoverRepository(t *testing.T) {
 
 	ps.SetCoverIfEmpty(ctx, a.ID, photo2)
 	list, err := store.ListForUser(ctx, owner.ID)
-	if err != nil || len(list) != 1 || list[0].CoverThumbKey != "thumb-2" {
+	if err != nil || len(list) != 1 || list[0].CoverThumbKey != thumb2 {
 		t.Fatalf("list cover key: %v %+v", err, list)
 	}
 }
 
 func TestMemberRepository(t *testing.T) {
 	pool := testPool(t)
-	defer pool.Close()
 	ctx := context.Background()
 
 	owner := newUser(t, ctx, pool, "Owner")
@@ -255,7 +267,6 @@ func TestMemberRepository(t *testing.T) {
 
 func TestShareLinkRepository(t *testing.T) {
 	pool := testPool(t)
-	defer pool.Close()
 	ctx := context.Background()
 
 	owner := newUser(t, ctx, pool, "Owner")

backend/internal/storage/storage.go

@@ -104,24 +104,6 @@ func (c *Client) Delete(ctx context.Context, keys ...string) error {
 	return nil
 }
 
-func (c *Client) DeletePrefix(ctx context.Context, prefix string) error {
-	var keys []string
-	paginator := s3.NewListObjectsV2Paginator(c.s3, &s3.ListObjectsV2Input{
-		Bucket: aws.String(c.bucket),
-		Prefix: aws.String(prefix),
-	})
-	for paginator.HasMorePages() {
-		page, err := paginator.NextPage(ctx)
-		if err != nil {
-			return fmt.Errorf("list objects: %w", err)
-		}
-		for _, o := range page.Contents {
-			keys = append(keys, aws.ToString(o.Key))
-		}
-	}
-	return c.Delete(ctx, keys...)
-}
-
 func (c *Client) Get(ctx context.Context, key string) ([]byte, error) {
 	out, err := c.s3.GetObject(ctx, &s3.GetObjectInput{
 		Bucket: aws.String(c.bucket),

backend/migrations/000008_media_objects.down.sql

@@ -0,0 +1,3 @@
+DROP TABLE IF EXISTS variant_jobs;
+DROP TABLE IF EXISTS photo_objects;
+DROP TABLE IF EXISTS objects;

backend/migrations/000008_media_objects.up.sql

@@ -0,0 +1,39 @@
+-- Physical objects stored in S3. Keys are opaque and never encode album, photo
+-- or any other metadata entity. The row is the ledger entry that governs the
+-- lifecycle of the corresponding blob.
+CREATE TABLE IF NOT EXISTS objects (
+    id          text PRIMARY KEY,
+    storage_key text NOT NULL UNIQUE,
+    mime_type   text NOT NULL DEFAULT '',
+    size        bigint NOT NULL DEFAULT 0,
+    width       integer NOT NULL DEFAULT 0,
+    height      integer NOT NULL DEFAULT 0,
+    checksum    text NOT NULL DEFAULT '',
+    state       text NOT NULL DEFAULT 'pending'
+        CHECK (state IN ('pending', 'ready', 'deleting')),
+    created_at  timestamptz NOT NULL DEFAULT now()
+);
+
+CREATE INDEX IF NOT EXISTS objects_state_created_idx ON objects (state, created_at);
+
+-- Binding of a photo to its objects by role. The role is a property of the
+-- binding, not of the object itself.
+CREATE TABLE IF NOT EXISTS photo_objects (
+    photo_id  text NOT NULL REFERENCES photos(id) ON DELETE CASCADE,
+    object_id text NOT NULL REFERENCES objects(id) ON DELETE RESTRICT,
+    role      text NOT NULL CHECK (role IN ('original', 'thumb', 'preview')),
+    PRIMARY KEY (photo_id, role)
+);
+
+CREATE INDEX IF NOT EXISTS photo_objects_object_idx ON photo_objects (object_id);
+
+-- Durable queue of pending variant generation work, one job per photo.
+CREATE TABLE IF NOT EXISTS variant_jobs (
+    photo_id        text PRIMARY KEY REFERENCES photos(id) ON DELETE CASCADE,
+    attempts        integer NOT NULL DEFAULT 0,
+    next_attempt_at timestamptz NOT NULL DEFAULT now(),
+    last_error      text NOT NULL DEFAULT '',
+    created_at      timestamptz NOT NULL DEFAULT now()
+);
+
+CREATE INDEX IF NOT EXISTS variant_jobs_due_idx ON variant_jobs (next_attempt_at);

backend/migrations/000009_drop_photo_keys.down.sql

@@ -0,0 +1,8 @@
+ALTER TABLE photos
+    ADD COLUMN IF NOT EXISTS storage_key text NOT NULL DEFAULT '',
+    ADD COLUMN IF NOT EXISTS thumb_key text,
+    ADD COLUMN IF NOT EXISTS preview_key text,
+    ADD COLUMN IF NOT EXISTS size bigint NOT NULL DEFAULT 0,
+    ADD COLUMN IF NOT EXISTS width integer NOT NULL DEFAULT 0,
+    ADD COLUMN IF NOT EXISTS height integer NOT NULL DEFAULT 0,
+    ADD COLUMN IF NOT EXISTS mime_type text NOT NULL DEFAULT '';

backend/migrations/000009_drop_photo_keys.up.sql

@@ -0,0 +1,10 @@
+-- Physical facts move to the objects table; photos keeps only membership and
+-- presentation data. size/width/height/mime_type are derived from the original
+-- object via photo_objects + objects.
+ALTER TABLE photos DROP COLUMN IF EXISTS storage_key;
+ALTER TABLE photos DROP COLUMN IF EXISTS thumb_key;
+ALTER TABLE photos DROP COLUMN IF EXISTS preview_key;
+ALTER TABLE photos DROP COLUMN IF EXISTS size;
+ALTER TABLE photos DROP COLUMN IF EXISTS width;
+ALTER TABLE photos DROP COLUMN IF EXISTS height;
+ALTER TABLE photos DROP COLUMN IF EXISTS mime_type;