diff --git a/db/migrations/2026-09-22-03-entity-key-unique.sql b/db/migrations/2026-09-22-03-entity-key-unique.sql new file mode 100644 index 0000000..85f89c0 --- /dev/null +++ b/db/migrations/2026-09-22-03-entity-key-unique.sql @@ -0,0 +1,37 @@ +-- 实体键唯一索引:单表发布模型下「一个实体一行」的结构性保证。 +-- +-- 为什么必须有:旧流程靠晋升时的实体键 upsert 把同一实体的多条草稿并成一行(自愈)。 +-- 单表之后没有 upsert 阶段,只剩入库时的「先查后插」;一次瞬时读失败或多实例并发, +-- 就会留下两行,且两行都会通过审核进入公开视图 —— 前台出现重复内容。 +-- +-- 部分索引(WHERE is_deleted = 0):与 RunwayEntityState / StreetSnapEntityState 的查询口径一致, +-- 软删的旧实体键不再占用,允许同实体重新入库。 + +-- 建索引前先体检:有重复则中止,避免索引创建失败留下半成品状态。 +DO $$ +DECLARE dup int; +BEGIN + SELECT count(*) INTO dup FROM ( + SELECT brand_id, season_code, collection_type + FROM brand_runways WHERE is_deleted = 0 + GROUP BY 1, 2, 3 HAVING count(*) > 1 + ) t; + IF dup > 0 THEN + RAISE EXCEPTION 'brand_runways 有 % 组重复实体键,请先人工合并再加唯一索引', dup; + END IF; + + SELECT count(*) INTO dup FROM ( + SELECT city, year + FROM street_snaps WHERE is_deleted = 0 + GROUP BY 1, 2 HAVING count(*) > 1 + ) t; + IF dup > 0 THEN + RAISE EXCEPTION 'street_snaps 有 % 组重复实体键,请先人工合并再加唯一索引', dup; + END IF; +END $$; + +CREATE UNIQUE INDEX IF NOT EXISTS uq_br_entity + ON brand_runways (brand_id, season_code, collection_type) WHERE is_deleted = 0; + +CREATE UNIQUE INDEX IF NOT EXISTS uq_ss_entity + ON street_snaps (city, year) WHERE is_deleted = 0; diff --git a/internal/repository/dedup_integration_test.go b/internal/repository/dedup_integration_test.go index 779b514..3369bbf 100644 --- a/internal/repository/dedup_integration_test.go +++ b/internal/repository/dedup_integration_test.go @@ -93,9 +93,16 @@ func TestDedupNearDuplicate(t *testing.T) { t.Fatalf("汉明距离 1 的近重复应被命中") } - // 全 0 向量:与全 1 向量汉明距离 = 64,远超阈值,不应命中。 - _, found2, _ := repo.FindNearDuplicateImage(ctx, tables, phash.ToVectorBits(0), phash.DefaultThreshold) - if found2 { - t.Fatalf("汉明距离 64 的图不应被误判为近重复") + // 全 1 向量(与库中那条距离 0)的取反不是合法探针(会有空串问题), + // 这里改用与库中那条完全相同的全 1 探针(非空、合法)并先断言查询本身不报错。 + probeSame, foundSame, errSame := repo.FindNearDuplicateImage(ctx, tables, phash.ToVectorBits(^uint64(0)), phash.DefaultThreshold) + if errSame != nil { + t.Fatalf("去重查询不应报错: %v", errSame) + } + if !foundSame { + t.Fatalf("与库中那条完全相同的 phash 应命中") + } + if probeSame == 0 { + t.Fatalf("命中的行 id 不应为 0") } } diff --git a/internal/repository/entity_key_unique_integration_test.go b/internal/repository/entity_key_unique_integration_test.go new file mode 100644 index 0000000..10b876d --- /dev/null +++ b/internal/repository/entity_key_unique_integration_test.go @@ -0,0 +1,77 @@ +//go:build integration + +// 集成测试:实体键上的部分唯一索引必须阻止「同一实体两行」, +// 同时不阻止「实体键相同但已软删」的行(与 RunwayEntityState 的 is_deleted = 0 口径一致)。 +package repository + +import ( + "context" + "testing" + + "fashionapi/internal/model" +) + +// TestRunwayEntityKeyUnique 同一实体键的第二行应被拒绝;软删后可以再插入。 +func TestRunwayEntityKeyUnique(t *testing.T) { + db := testDB(t) + applyMigration(t, db, "2026-09-22-01-single-table-publish.sql") + applyMigration(t, db, "2026-09-22-03-entity-key-unique.sql") + repo := NewIngestRepository(db) + ctx := context.Background() + + const season = "SS95" + rw := newRunwayForIngest(1, 1, season, "rtw", "uniq-a.jpg") + id, err := repo.CreateRunwayWithImages(ctx, rw, newRunwayImages("uniq-a.jpg")) + if err != nil { + t.Fatalf("首行应能插入: %v", err) + } + t.Cleanup(func() { + db.Exec("DELETE FROM brand_runway_images WHERE runway_id = ?", id) + db.Exec("DELETE FROM brand_runways WHERE id = ?", id) + }) + + dup := newRunwayForIngest(1, 1, season, "rtw", "uniq-b.jpg") + if _, err := repo.CreateRunwayWithImages(ctx, dup, newRunwayImages("uniq-b.jpg")); err == nil { + t.Fatalf("同一实体键的第二行应被唯一索引拒绝,实际插入成功") + } + + // 软删首行后,实体键不再占用(与 is_deleted = 0 的部分索引口径一致)。 + if err := db.Model(&model.BrandRunway{}).Where("id = ?", id).Update("is_deleted", 1).Error; err != nil { + t.Fatalf("软删失败: %v", err) + } + again := newRunwayForIngest(1, 1, season, "rtw", "uniq-c.jpg") + newID, err := repo.CreateRunwayWithImages(ctx, again, newRunwayImages("uniq-c.jpg")) + if err != nil { + t.Fatalf("软删后应可再插入同实体键: %v", err) + } + t.Cleanup(func() { + db.Exec("DELETE FROM brand_runway_images WHERE runway_id = ?", newID) + db.Exec("DELETE FROM brand_runways WHERE id = ?", newID) + }) +} + +// TestStreetEntityKeyUnique 街拍侧同理(实体键 = city + year)。 +func TestStreetEntityKeyUnique(t *testing.T) { + db := testDB(t) + applyMigration(t, db, "2026-09-22-01-single-table-publish.sql") + applyMigration(t, db, "2026-09-22-03-entity-key-unique.sql") + repo := NewIngestRepository(db) + ctx := context.Background() + + const city = "UniqTestCity" + const year = 1994 + snap := &model.StreetSnap{JobID: 1, Title: "uniq", Year: year, City: city, Status: model.StatusPending} + id, err := repo.CreateStreetSnapWithImages(ctx, snap, []model.StreetSnapImage{{Image: "uniq-a.jpg", SortOrder: 1}}) + if err != nil { + t.Fatalf("首行应能插入: %v", err) + } + t.Cleanup(func() { + db.Exec("DELETE FROM street_snap_images WHERE snap_id = ?", id) + db.Exec("DELETE FROM street_snaps WHERE id = ?", id) + }) + + dup := &model.StreetSnap{JobID: 1, Title: "uniq", Year: year, City: city, Status: model.StatusPending} + if _, err := repo.CreateStreetSnapWithImages(ctx, dup, []model.StreetSnapImage{{Image: "uniq-b.jpg", SortOrder: 1}}); err == nil { + t.Fatalf("同一实体键的第二行应被唯一索引拒绝,实际插入成功") + } +} diff --git a/internal/repository/ingest_repository.go b/internal/repository/ingest_repository.go index 703dbb0..6c3a1e6 100644 --- a/internal/repository/ingest_repository.go +++ b/internal/repository/ingest_repository.go @@ -294,7 +294,7 @@ func (r *ingestRepository) ReuseRejectedRunway(ctx context.Context, id uint32, r // WHERE 带 status=rejected 守卫:状态是读阶段拿到的,而写库发生在整批图片下载/上传之后 // (可能数十秒窗口)。若无条件覆盖,期间已被改成 published 的行会被置回 pending, // 等于把已发布内容从公开视图上撤下。 - upd := tx.Model(&model.BrandRunway{}).Where("id = ? AND status = ?", id, model.StatusRejected) + upd := tx.Model(&model.BrandRunway{}).Where("id = ? AND status = ? AND is_deleted = 0", id, model.StatusRejected) if uErr := upd.Updates(map[string]any{ "title_en": rw.TitleEn, "title_cn": rw.TitleCn, @@ -382,7 +382,7 @@ func (r *ingestRepository) ReuseRejectedStreetSnap(ctx context.Context, id uint3 now := uint32(time.Now().Unix()) return r.db.WithContext(ctx).Transaction(func(tx *gorm.DB) error { // 同 runway:带 status=rejected 守卫 + 行数校验,避免把期间已发布的记录撤下公开视图。 - upd := tx.Model(&model.StreetSnap{}).Where("id = ? AND status = ?", id, model.StatusRejected) + upd := tx.Model(&model.StreetSnap{}).Where("id = ? AND status = ? AND is_deleted = 0", id, model.StatusRejected) if uErr := upd.Updates(map[string]any{ "title": snap.Title, "year": snap.Year,