update backend
This commit is contained in:
@ -12,6 +12,7 @@ import (
|
||||
"net/http"
|
||||
"path"
|
||||
"strconv"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
"fashionapi/internal/dto"
|
||||
@ -27,31 +28,44 @@ import (
|
||||
//
|
||||
// 队列复用一张 ingest_jobs 表,靠 kind 列分流两类任务:
|
||||
// - crawl:爬虫上报的走秀/街拍入库(下载图 → 写草稿)
|
||||
// - media_cleanup:删除图集时异步清理七牛孤儿图(引用计数归零才真删)
|
||||
// - media_cleanup:删除图集时异步清理S4孤儿图(引用计数归零才真删)
|
||||
type IngestService struct {
|
||||
repo repository.IngestRepository
|
||||
brandRepo repository.BrandRepository
|
||||
media repository.MediaRepository // 跨表图片引用计数(media_cleanup 删孤儿用)
|
||||
uploader storage.Uploader // 主上传器(七牛启用时为七牛,否则本地)
|
||||
del storage.Deleter // 删除器(七牛或本地,media_cleanup 真删用)
|
||||
local *storage.LocalUploader // 七牛失败时的兜底落地
|
||||
uploader storage.Uploader // 主上传器(S4启用时为S4,否则本地)
|
||||
del storage.Deleter // 删除器(S4或本地,media_cleanup 真删用)
|
||||
local *storage.LocalUploader // S4失败时的兜底落地
|
||||
httpClient *http.Client
|
||||
}
|
||||
|
||||
// NewIngestService 创建入库服务。
|
||||
//
|
||||
// uploader: 主上传器(七牛或本地)
|
||||
// del: 删除器(七牛或本地),用于 media_cleanup 真删七牛/本地孤儿文件
|
||||
// local: 本地兜底上传器(七牛上传失败时回退,避免图片完全丢失)
|
||||
// uploader: 主上传器(S4或本地)
|
||||
// del: 删除器(S4或本地),用于 media_cleanup 真删S4/本地孤儿文件
|
||||
// local: 本地兜底上传器(S4上传失败时回退,避免图片完全丢失)
|
||||
func NewIngestService(repo repository.IngestRepository, brandRepo repository.BrandRepository, media repository.MediaRepository, uploader storage.Uploader, del storage.Deleter, local *storage.LocalUploader) *IngestService {
|
||||
return &IngestService{
|
||||
repo: repo,
|
||||
brandRepo: brandRepo,
|
||||
media: media,
|
||||
uploader: uploader,
|
||||
del: del,
|
||||
local: local,
|
||||
httpClient: &http.Client{Timeout: 30 * time.Second},
|
||||
repo: repo,
|
||||
brandRepo: brandRepo,
|
||||
media: media,
|
||||
uploader: uploader,
|
||||
del: del,
|
||||
local: local,
|
||||
// 图片下载是并发执行的(见 imageFetchConcurrency),连接池须匹配并发度:
|
||||
// Go 默认 MaxIdleConnsPerHost=2,多余连接会在每轮下载时反复重建、白付 TLS 握手开销。
|
||||
httpClient: &http.Client{
|
||||
// 超时须显著大于单张图的正常下载耗时(实测 4~22s,且随源站波动剧烈)。
|
||||
// 原 30s 余量太薄:一次抖动就会让「单图失败=整任务失败」把整批打回重试。
|
||||
Timeout: 60 * time.Second,
|
||||
Transport: &http.Transport{
|
||||
Proxy: http.ProxyFromEnvironment,
|
||||
MaxIdleConns: imageFetchConcurrency * 4,
|
||||
MaxIdleConnsPerHost: imageFetchConcurrency * 2,
|
||||
IdleConnTimeout: 90 * time.Second,
|
||||
TLSHandshakeTimeout: 10 * time.Second,
|
||||
},
|
||||
},
|
||||
}
|
||||
}
|
||||
|
||||
@ -129,6 +143,7 @@ func (s *IngestService) drain(ctx context.Context, batch int) {
|
||||
// process 处理单条任务:先按 job.Kind(DB 列)分流到 crawl / media_cleanup 两条管线。
|
||||
// 旧爬虫任务(迁移前)kind 为空,回落 crawl 走秀/街拍入库管线。
|
||||
func (s *IngestService) process(ctx context.Context, job model.IngestJob) {
|
||||
start := time.Now()
|
||||
switch job.Kind {
|
||||
case model.IngestKindMediaCleanup:
|
||||
s.processMediaCleanup(ctx, job)
|
||||
@ -149,6 +164,7 @@ func (s *IngestService) process(ctx context.Context, job model.IngestJob) {
|
||||
if p.Kind == "" {
|
||||
p.Kind = dto.IngestKindRunway
|
||||
}
|
||||
log.Printf("[ingest] job=%d kind=%s 开始处理(图片=%d 张)", job.ID, p.Kind, len(p.Images))
|
||||
switch p.Kind {
|
||||
case dto.IngestKindRunway:
|
||||
s.processRunway(ctx, job, p)
|
||||
@ -158,15 +174,17 @@ func (s *IngestService) process(ctx context.Context, job model.IngestJob) {
|
||||
// 未知 kind 绝不静默当成走秀处理;所有爬虫数据都必须落到已注册的审核模块,
|
||||
// 否则标记任务失败,避免出现「未审核就入库」的脏数据。
|
||||
_ = s.repo.MarkFailed(ctx, job.ID, "unknown ingest kind: "+p.Kind)
|
||||
return
|
||||
}
|
||||
log.Printf("[ingest] job=%d kind=%s 处理结束 总耗时=%v", job.ID, p.Kind, time.Since(start).Round(time.Millisecond))
|
||||
}
|
||||
|
||||
// MediaCleanupPayload 是「清理七牛孤儿图」任务的 payload:待清理的图片 key 列表。
|
||||
// MediaCleanupPayload 是「清理S4孤儿图」任务的 payload:待清理的图片 key 列表。
|
||||
type MediaCleanupPayload struct {
|
||||
Keys []string `json:"keys"`
|
||||
}
|
||||
|
||||
// EnqueueMediaCleanup 把一批待清理的七牛 key 异步入队;真正删除由 worker 的
|
||||
// EnqueueMediaCleanup 把一批待清理的S4 key 异步入队;真正删除由 worker 的
|
||||
// processMediaCleanup 按引用计数判定,仅当 key 在所有图集/草稿表中引用归零才真删
|
||||
// (内容寻址共享 key 不会被误删)。
|
||||
func (s *IngestService) EnqueueMediaCleanup(ctx context.Context, keys []string) error {
|
||||
@ -180,7 +198,7 @@ func (s *IngestService) EnqueueMediaCleanup(ctx context.Context, keys []string)
|
||||
return s.repo.EnqueueMediaCleanup(ctx, string(raw))
|
||||
}
|
||||
|
||||
// processMediaCleanup 清理七牛孤儿图任务:解析 key 列表,按跨表引用计数删除真孤儿。
|
||||
// processMediaCleanup 清理S4孤儿图任务:解析 key 列表,按跨表引用计数删除真孤儿。
|
||||
// 删除/计数失败一律保守跳过(宁可留文件),因此本任务几乎总是成功置 done。
|
||||
func (s *IngestService) processMediaCleanup(ctx context.Context, job model.IngestJob) {
|
||||
if s.media == nil || s.del == nil {
|
||||
@ -211,7 +229,7 @@ func (s *IngestService) failOrRetry(ctx context.Context, id uint32, errMsg strin
|
||||
}
|
||||
}
|
||||
|
||||
// processRunway 走秀入库:品牌校验 → 去重 → 补季节码 → 下载图 → 写 brand_runway_draft。
|
||||
// processRunway 走秀入库:品牌校验 → 去重 → 补季节码 → 下载图 → 写 brand_runway_drafts。
|
||||
func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob, p dto.RunwayIngest) {
|
||||
// 1) 品牌必须存在(爬虫负责先建/复用品牌)
|
||||
brandID, err := hashid.Decode(p.BrandUID)
|
||||
@ -228,26 +246,31 @@ func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob,
|
||||
// 避免重爬重复下载。注意:只判正式表、不判 pending 草稿——否则多来源(Vogue + theImpression)
|
||||
// 爬同一场秀时第二个来源会被误判重复而丢弃,破坏晋升阶段的图片聚合。
|
||||
seasonCode := season.Derive(p.Year, p.CollectionType, p.Season)
|
||||
dedupStart := time.Now()
|
||||
if _, found, err := s.repo.RunwayIDByEntity(ctx, brandID, seasonCode, p.CollectionType); err == nil && found {
|
||||
log.Printf("[ingest] job=%d runway 实体去重命中(season=%s type=%s),跳过 耗时=%v",
|
||||
job.ID, seasonCode, p.CollectionType, time.Since(dedupStart).Round(time.Millisecond))
|
||||
_ = s.repo.MarkDone(ctx, job.ID)
|
||||
return
|
||||
}
|
||||
log.Printf("[ingest] job=%d runway 实体去重 耗时=%v(未命中,继续下载)", job.ID, time.Since(dedupStart).Round(time.Millisecond))
|
||||
|
||||
// 3) season_code 已在去重前补齐(vogue.go 历史漏填的 bug,统一在此兜底)
|
||||
|
||||
// 4) 下载图片并上传到存储(结构化 Looks 优先:主图+细节图分组;否则回退 Images 全部视为主图)。
|
||||
// 内容哈希(sha1)key 保证重爬不产生孤儿文件:失败回滚删本批 key 即可。
|
||||
var cover string
|
||||
var cover string
|
||||
var draftImages []model.BrandRunwayDraftImage
|
||||
var keys []string
|
||||
var imgFailed bool
|
||||
var imageCount uint16
|
||||
var timing *fetchTiming
|
||||
if len(p.Looks) > 0 {
|
||||
cover, draftImages, keys, imgFailed = s.fetchLookImages(ctx, p.Looks, "runway")
|
||||
cover, draftImages, keys, imgFailed, timing = s.fetchLookImages(ctx, p.Looks, "runway")
|
||||
imageCount = uint16(len(p.Looks))
|
||||
} else {
|
||||
var fimgs []fetchedImage
|
||||
cover, fimgs, keys, imgFailed = s.fetchImages(ctx, p.Images, "runway", dto.IngestKindRunway)
|
||||
cover, fimgs, keys, imgFailed, timing = s.fetchImages(ctx, p.Images, "runway", dto.IngestKindRunway)
|
||||
imageCount = uint16(len(fimgs))
|
||||
for i, fi := range fimgs {
|
||||
draftImages = append(draftImages, model.BrandRunwayDraftImage{
|
||||
@ -256,13 +279,13 @@ func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob,
|
||||
SortOrder: uint32(i + 1),
|
||||
LookIndex: uint32(i + 1),
|
||||
IsDetail: 0,
|
||||
ContentSha1: fi.sha1,
|
||||
Phash: sqlNull(fi.phash),
|
||||
IsDuplicate: fi.isDup,
|
||||
DupOf: strconv.FormatUint(uint64(fi.dupID), 10),
|
||||
})
|
||||
}
|
||||
}
|
||||
log.Printf("[ingest] job=%d runway 拉图 总耗时=%v %s", job.ID, timing.total.Round(time.Millisecond), timing)
|
||||
if imgFailed {
|
||||
// 单图失败=整任务失败:先回滚本批已上传的图(S4 + 本地兜底),避免孤儿文件永远堆在存储里,
|
||||
// 然后按指数退避自动重试,达上限才置 failed 等后台手动重试。
|
||||
@ -271,8 +294,8 @@ func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob,
|
||||
return
|
||||
}
|
||||
|
||||
|
||||
// 5) 写草稿表(status=pending),等待后台审核通过后再晋升正式表
|
||||
writeStart := time.Now()
|
||||
draft := &model.BrandRunwayDraft{
|
||||
JobID: job.ID,
|
||||
BrandID: brandID,
|
||||
@ -300,6 +323,7 @@ func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob,
|
||||
s.failOrRetry(ctx, job.ID, "create draft images: "+err.Error())
|
||||
return
|
||||
}
|
||||
log.Printf("[ingest] job=%d runway 写草稿 耗时=%v 图片行=%d", job.ID, time.Since(writeStart).Round(time.Millisecond), len(draftImages))
|
||||
_ = s.repo.MarkDone(ctx, job.ID)
|
||||
}
|
||||
|
||||
@ -307,102 +331,109 @@ func (s *IngestService) processRunway(ctx context.Context, job model.IngestJob,
|
||||
// BrandRunwayDraftImage 行(带 look_index / is_detail 分组)。任意一张下载/上传失败即把
|
||||
// failed 置 true,调用方据此把整条任务判失败并回滚本批已上传的 key,符合「单图失败=整任务失败」策略。
|
||||
// cover 取首个成功下载的主图;image_count(主图数)由调用方按 len(Looks) 计,不在此返回。
|
||||
// 每张图入库前做去重:content_sha1 已存在则整行跳过(精确重复);命中 dHash 近重复则仍入库但标记留痕。
|
||||
func (s *IngestService) fetchLookImages(ctx context.Context, looks []dto.RunwayLook, prefix string) (string, []model.BrandRunwayDraftImage, []string, bool) {
|
||||
// 每张图入库前做去重:命中 dHash 近重复则仍入库但标记留痕。
|
||||
//
|
||||
// 实现:先把全部「主图 + 细节图」按原始顺序摊平成任务列表,用 concurrentFetch 并发完成
|
||||
// 「下载 → phash → 上传」这段 IO 密集操作;随后再按下标顺序串行去重与组装。
|
||||
// 这样既拿到并发收益,又保证 cover / sort_order / 批次内去重与原串行实现一致。
|
||||
func (s *IngestService) fetchLookImages(ctx context.Context, looks []dto.RunwayLook, prefix string) (string, []model.BrandRunwayDraftImage, []string, bool, *fetchTiming) {
|
||||
start := time.Now()
|
||||
timing := &fetchTiming{}
|
||||
cover := ""
|
||||
rows := make([]model.BrandRunwayDraftImage, 0)
|
||||
keys := make([]string, 0)
|
||||
seen := make(map[string]bool)
|
||||
failed := false
|
||||
order := 0
|
||||
|
||||
// 摊平:顺序 = 逐个 look 先主图、再其细节图,与原串行遍历顺序完全一致。
|
||||
tasks := make([]lookImageTask, 0, len(looks))
|
||||
for li, look := range looks {
|
||||
lookIdx := li + 1
|
||||
if look.Main != "" {
|
||||
url, key, sha1h, ph, err := s.downloadOne(ctx, look.Main, order, prefix)
|
||||
if err != nil {
|
||||
failed = true
|
||||
} else {
|
||||
keys = append(keys, key)
|
||||
if seen[sha1h] {
|
||||
continue
|
||||
}
|
||||
seen[sha1h] = true
|
||||
skip, dupID, isDup := s.dedupImage(ctx, dto.IngestKindRunway, sha1h, ph)
|
||||
if skip {
|
||||
continue
|
||||
}
|
||||
order++
|
||||
if cover == "" {
|
||||
cover = url
|
||||
}
|
||||
rows = append(rows, model.BrandRunwayDraftImage{
|
||||
Image: url,
|
||||
Name: fmt.Sprintf("Look %d", lookIdx),
|
||||
SortOrder: uint32(order),
|
||||
LookIndex: uint32(lookIdx),
|
||||
IsDetail: 0,
|
||||
ContentSha1: sha1h,
|
||||
Phash: sqlNull(ph),
|
||||
IsDuplicate: isDup,
|
||||
DupOf: strconv.FormatUint(uint64(dupID), 10),
|
||||
})
|
||||
}
|
||||
tasks = append(tasks, lookImageTask{url: look.Main, lookIdx: lookIdx})
|
||||
}
|
||||
for di, d := range look.Details {
|
||||
if d == "" {
|
||||
continue
|
||||
}
|
||||
url, key, sha1h, ph, err := s.downloadOne(ctx, d, order, prefix)
|
||||
if err != nil {
|
||||
failed = true
|
||||
continue
|
||||
}
|
||||
keys = append(keys, key)
|
||||
if seen[sha1h] {
|
||||
continue
|
||||
}
|
||||
seen[sha1h] = true
|
||||
skip, dupID, isDup := s.dedupImage(ctx, dto.IngestKindRunway, sha1h, ph)
|
||||
if skip {
|
||||
continue
|
||||
}
|
||||
order++
|
||||
rows = append(rows, model.BrandRunwayDraftImage{
|
||||
Image: url,
|
||||
Name: fmt.Sprintf("Look %d — Detail %d", lookIdx, di+1),
|
||||
SortOrder: uint32(order),
|
||||
LookIndex: uint32(lookIdx),
|
||||
IsDetail: 1,
|
||||
ContentSha1: sha1h,
|
||||
Phash: sqlNull(ph),
|
||||
IsDuplicate: isDup,
|
||||
DupOf: strconv.FormatUint(uint64(dupID), 10),
|
||||
})
|
||||
tasks = append(tasks, lookImageTask{url: d, lookIdx: lookIdx, isDetail: true, detailNo: di + 1})
|
||||
}
|
||||
}
|
||||
return cover, rows, keys, failed
|
||||
|
||||
// 并发段:下载 + phash + 上传(耗时大头)。
|
||||
results := concurrentFetch(ctx, len(tasks), func(c context.Context, i int) downloadResult {
|
||||
return s.downloadAndUpload(c, tasks[i].url, prefix)
|
||||
})
|
||||
|
||||
// 串行段:按原始顺序累计耗时、做批次内去重、组装行。
|
||||
order := 0
|
||||
for i, r := range results {
|
||||
task := tasks[i]
|
||||
timing.download += r.download
|
||||
timing.phash += r.phash
|
||||
timing.upload += r.upload
|
||||
if r.err != nil {
|
||||
failed = true
|
||||
continue
|
||||
}
|
||||
keys = append(keys, r.key)
|
||||
if seen[r.sha1] {
|
||||
continue
|
||||
}
|
||||
seen[r.sha1] = true
|
||||
dupID, isDup := s.dedupImage(ctx, dto.IngestKindRunway, r.phashBits, timing)
|
||||
order++
|
||||
if cover == "" && !task.isDetail {
|
||||
cover = r.url
|
||||
}
|
||||
name := fmt.Sprintf("Look %d", task.lookIdx)
|
||||
isDetail := uint8(0)
|
||||
if task.isDetail {
|
||||
name = fmt.Sprintf("Look %d — Detail %d", task.lookIdx, task.detailNo)
|
||||
isDetail = 1
|
||||
}
|
||||
rows = append(rows, model.BrandRunwayDraftImage{
|
||||
Image: r.url,
|
||||
Name: name,
|
||||
SortOrder: uint32(order),
|
||||
LookIndex: uint32(task.lookIdx),
|
||||
IsDetail: isDetail,
|
||||
Phash: sqlNull(r.phashBits),
|
||||
IsDuplicate: isDup,
|
||||
DupOf: strconv.FormatUint(uint64(dupID), 10),
|
||||
})
|
||||
}
|
||||
timing.images = len(rows)
|
||||
timing.total = time.Since(start)
|
||||
return cover, rows, keys, failed, timing
|
||||
}
|
||||
|
||||
// processStreet 街拍入库:去重 → 下载图 → 写 street_snap_draft(无品牌)。
|
||||
func (s *IngestService) processStreet(ctx context.Context, job model.IngestJob, p dto.RunwayIngest) {
|
||||
// 1) 按实体键(city + year)去重:正式表已存在该街拍则跳过,避免重爬重复下载。
|
||||
// 只判正式表、不判 pending 草稿,保留多来源街拍图片在晋升阶段聚合。
|
||||
dedupStart := time.Now()
|
||||
if _, found, err := s.repo.StreetSnapIDByEntity(ctx, p.City, p.Year); err == nil && found {
|
||||
log.Printf("[ingest] job=%d street 实体去重命中(city=%s year=%d),跳过 耗时=%v",
|
||||
job.ID, p.City, p.Year, time.Since(dedupStart).Round(time.Millisecond))
|
||||
_ = s.repo.MarkDone(ctx, job.ID)
|
||||
return
|
||||
}
|
||||
log.Printf("[ingest] job=%d street 实体去重 耗时=%v(未命中,继续下载)", job.ID, time.Since(dedupStart).Round(time.Millisecond))
|
||||
|
||||
// 2) 下载图片并上传到存储(七牛优先,失败兜底本地)。
|
||||
cover, fimgs, keys, imgFailed := s.fetchImages(ctx, p.Images, "street", dto.IngestKindStreet)
|
||||
// 2) 下载图片并上传到存储(S4优先,失败兜底本地)。
|
||||
cover, fimgs, keys, imgFailed, timing := s.fetchImages(ctx, p.Images, "street", dto.IngestKindStreet)
|
||||
log.Printf("[ingest] job=%d street 拉图 总耗时=%v %s", job.ID, timing.total.Round(time.Millisecond), timing)
|
||||
if len(p.Images) > 0 && imgFailed {
|
||||
// 单图失败=整任务失败:先回滚本批已上传的图(七牛 + 本地兜底),避免孤儿文件永远堆在存储里,
|
||||
// 单图失败=整任务失败:先回滚本批已上传的图(S4 + 本地兜底),避免孤儿文件永远堆在存储里,
|
||||
// 然后按指数退避自动重试,达上限才置 failed 等后台手动重试。
|
||||
s.cleanupUploads(ctx, keys)
|
||||
s.failOrRetry(ctx, job.ID, "image download failed")
|
||||
return
|
||||
}
|
||||
|
||||
// 3) 写草稿表(status=pending),等待后台审核通过后再晋升 street_snap 正式表
|
||||
// 3) 写草稿表(status=pending),等待后台审核通过后再晋升 street_snaps 正式表
|
||||
writeStart := time.Now()
|
||||
draft := &model.StreetSnapDraft{
|
||||
JobID: job.ID,
|
||||
Title: p.TitleEn, // 街拍单标题,爬虫优先填 title_en
|
||||
@ -424,66 +455,73 @@ func (s *IngestService) processStreet(ctx context.Context, job model.IngestJob,
|
||||
Image: fi.url,
|
||||
Name: fmt.Sprintf("Look %d", i+1),
|
||||
SortOrder: uint32(i + 1),
|
||||
ContentSha1: fi.sha1,
|
||||
Phash: sqlNull(fi.phash),
|
||||
IsDuplicate: fi.isDup,
|
||||
DupOf: strconv.FormatUint(uint64(fi.dupID), 10),
|
||||
})
|
||||
}
|
||||
|
||||
|
||||
if err := s.repo.CreateStreetSnapDraftImages(ctx, rows); err != nil {
|
||||
s.failOrRetry(ctx, job.ID, "create street draft images: "+err.Error())
|
||||
return
|
||||
}
|
||||
log.Printf("[ingest] job=%d street 写草稿 耗时=%v 图片行=%d", job.ID, time.Since(writeStart).Round(time.Millisecond), len(rows))
|
||||
_ = s.repo.MarkDone(ctx, job.ID)
|
||||
}
|
||||
|
||||
// fetchImages 下载图片并上传到存储,返回 (cover 地址, 全部图片地址, 已成功上传对象的 key 列表, 是否有任意一张失败)。
|
||||
// prefix 为七牛 key 前缀(runway/ 或 street/)。只要任意一张下载/上传失败,failed 即置 true,
|
||||
// 调用方据此把整条任务判为失败(不再写草稿),并拿 keys 回滚本批已上传的对象,符合「单图失败=整任务失败」策略。
|
||||
|
||||
// fetchImages 下载图片并上传到存储,返回 (cover 地址, 已下载图结构, 已成功上传对象的 key 列表, 是否有任意一张失败)。
|
||||
// prefix 为七牛 key 前缀(runway/ 或 street/)。kind 用于选择去重比对表。
|
||||
// prefix 为S4 key 前缀(runway/ 或 street/)。kind 用于选择去重比对表。
|
||||
// 只要任意一张下载/上传失败,failed 即置 true,调用方据此把整条任务判为失败并回滚本批已上传的对象,
|
||||
// 符合「单图失败=整任务失败」策略。每张图入库前做去重(精确跳过 + 近重复标记)。
|
||||
func (s *IngestService) fetchImages(ctx context.Context, urls []string, prefix, kind string) (string, []fetchedImage, []string, bool) {
|
||||
// 符合「单图失败=整任务失败」策略。每张图入库前做去重(近重复标记)。
|
||||
//
|
||||
// 实现:先用 concurrentFetch 并发完成全部图片的「下载 → phash → 上传」,再按下标顺序串行
|
||||
// 做去重与组装,保证 cover / 顺序 / 批次内去重与原串行实现一致。
|
||||
func (s *IngestService) fetchImages(ctx context.Context, urls []string, prefix, kind string) (string, []fetchedImage, []string, bool, *fetchTiming) {
|
||||
start := time.Now()
|
||||
timing := &fetchTiming{}
|
||||
cover := ""
|
||||
out := make([]fetchedImage, 0, len(urls))
|
||||
keys := make([]string, 0, len(urls))
|
||||
seen := make(map[string]bool)
|
||||
failed := false
|
||||
for _, u := range urls {
|
||||
url, key, sha1h, ph, err := s.downloadOne(ctx, u, 0, prefix)
|
||||
if err != nil {
|
||||
|
||||
// 并发段:下载 + phash + 上传(耗时大头)。
|
||||
results := concurrentFetch(ctx, len(urls), func(c context.Context, i int) downloadResult {
|
||||
return s.downloadAndUpload(c, urls[i], prefix)
|
||||
})
|
||||
|
||||
// 串行段:按原始顺序累计耗时、做批次内去重、组装结果。
|
||||
for _, r := range results {
|
||||
timing.download += r.download
|
||||
timing.phash += r.phash
|
||||
timing.upload += r.upload
|
||||
if r.err != nil {
|
||||
failed = true
|
||||
continue
|
||||
}
|
||||
keys = append(keys, key)
|
||||
if seen[sha1h] {
|
||||
continue
|
||||
}
|
||||
seen[sha1h] = true
|
||||
skip, dupID, isDup := s.dedupImage(ctx, kind, sha1h, ph)
|
||||
if skip {
|
||||
keys = append(keys, r.key)
|
||||
if seen[r.sha1] {
|
||||
continue
|
||||
}
|
||||
seen[r.sha1] = true
|
||||
dupID, isDup := s.dedupImage(ctx, kind, r.phashBits, timing)
|
||||
if cover == "" {
|
||||
cover = url
|
||||
cover = r.url
|
||||
}
|
||||
out = append(out, fetchedImage{
|
||||
url: url,
|
||||
key: key,
|
||||
sha1: sha1h,
|
||||
phash: ph,
|
||||
url: r.url,
|
||||
key: r.key,
|
||||
phash: r.phashBits,
|
||||
dupID: dupID,
|
||||
isDup: isDup,
|
||||
})
|
||||
}
|
||||
return cover, out, keys, failed
|
||||
timing.images = len(out)
|
||||
timing.total = time.Since(start)
|
||||
return cover, out, keys, failed, timing
|
||||
}
|
||||
|
||||
// cleanupUploads 删除一批本批次成功上传的对象(七牛 + 本地兜底),用于任务失败回滚:
|
||||
// cleanupUploads 删除一批本批次成功上传的对象(S4 + 本地兜底),用于任务失败回滚:
|
||||
// 「单图失败=整任务失败」时,前面已成功上传的图若放任不管就成了孤儿,永远堆在存储里。
|
||||
// 对两种存储都尝试删除(任一不存在即按幂等成功处理);删除失败仅告警,不阻断任务置失败。
|
||||
// 注意:key 为 sha1 内容寻址,若恰与其他已晋升图集共享同一内容哈希会被一并移除,重试会重新上传补齐。
|
||||
@ -491,7 +529,7 @@ func (s *IngestService) cleanupUploads(ctx context.Context, keys []string) {
|
||||
for _, k := range keys {
|
||||
if d, ok := s.uploader.(storage.Deleter); ok {
|
||||
if err := d.Delete(ctx, k); err != nil {
|
||||
log.Printf("[warn] ingest cleanup: 七牛删除失败 key=%s err=%v", k, err)
|
||||
log.Printf("[warn] ingest cleanup: S4删除失败 key=%s err=%v", k, err)
|
||||
}
|
||||
}
|
||||
if s.local != nil {
|
||||
@ -506,25 +544,64 @@ func (s *IngestService) cleanupUploads(ctx context.Context, keys []string) {
|
||||
type fetchedImage struct {
|
||||
url string
|
||||
key string
|
||||
sha1 string // 内容 sha1(与存储 key 同源)
|
||||
phash string // dHash 的 pgvector 二进制向量串,空串表示无法解码
|
||||
dupID uint32 // 命中近重复时的参考图 id
|
||||
isDup uint8 // 是否标记为近重复(供审核留痕)
|
||||
}
|
||||
|
||||
// slowImageThreshold 单张图各阶段合计超过该阈值时单独打一条告警,便于从大量图里定位异常慢图。
|
||||
const slowImageThreshold = 2 * time.Second
|
||||
|
||||
// imageFetchConcurrency 单条入库任务内「下载 + 上传」的并发度。
|
||||
//
|
||||
// 实测结论(theimpression.com,59 张图):
|
||||
// - 并发 1:单张下载 ~3.7s,整批 59 张约 3m54s,任务成功。
|
||||
// - 并发 5:单张下载涨到 ~23s(约 6 倍),5m24s 仅完成 12/59 张便因 30s 超时失败。
|
||||
//
|
||||
// 即源站对同 IP 的并发连接有强限流:并发不仅不提速,反而把吞吐从 ~0.27 张/秒
|
||||
// 打到 ~0.04 张/秒并直接拖垮任务。因此默认取 1(等效串行)。
|
||||
// 换到不限流的源站时,可把这个值调大(机制已就绪),但务必先做小批量实测。
|
||||
const imageFetchConcurrency = 1
|
||||
|
||||
// fetchTiming 汇总一次图集拉取各阶段的累计耗时,用于定位「慢在哪一步」。
|
||||
// 下载 / 上传是网络 IO,phash 是 CPU(解码 + 哈希),近重复是 DB(HNSW 检索)——
|
||||
// 三类瓶颈成因不同,分开计时才能对症下药。
|
||||
type fetchTiming struct {
|
||||
images int // 成功入行的图片数
|
||||
total time.Duration // 整个拉取过程的墙钟耗时
|
||||
download time.Duration // 各张 HTTP 下载耗时之和
|
||||
phash time.Duration // 各张 dHash 指纹计算耗时之和
|
||||
upload time.Duration // 各张上传存储耗时之和
|
||||
dedup time.Duration // 近重复查询累计
|
||||
}
|
||||
|
||||
// String 输出各阶段耗时(毫秒精度),一条日志看清瓶颈。
|
||||
// 注意:下载 / phash / 上传是「各张图耗时之和」,并发执行下其总和会大于墙钟总耗时,
|
||||
// 因此三者与「总耗时」的比值不再等于时间占比——它们用于横向对比哪个阶段最重。
|
||||
func (t *fetchTiming) String() string {
|
||||
if t == nil {
|
||||
return ""
|
||||
}
|
||||
return fmt.Sprintf("图片=%d 并发=%d 下载=%v phash=%v 上传=%v 近重复=%v(下载/phash/上传为各张累计值)",
|
||||
t.images,
|
||||
imageFetchConcurrency,
|
||||
t.download.Round(time.Millisecond),
|
||||
t.phash.Round(time.Millisecond),
|
||||
t.upload.Round(time.Millisecond),
|
||||
t.dedup.Round(time.Millisecond),
|
||||
)
|
||||
}
|
||||
|
||||
// sqlNull 把 phash 字符串转成可空向量字段:空串 → NULL(不参与近邻检索)。
|
||||
func sqlNull(ph string) sql.NullString {
|
||||
return sql.NullString{String: ph, Valid: ph != ""}
|
||||
}
|
||||
|
||||
// dedupImage 判断单张图是否重复:
|
||||
// - 精确重复(content_sha1 已在库)→ 返回 skip=true(不入库该行);
|
||||
// - 近重复(dHash 汉明距离 ≤ 阈值)→ 仍入库,但标记 is_duplicate + dup_of;
|
||||
// - 否则正常入库。
|
||||
func (s *IngestService) dedupImage(ctx context.Context, kind, sha1hash, phashBits string) (skip bool, dupID uint32, isDup uint8) {
|
||||
// dedupImage 判断单张图是否近重复(dHash 汉明距离 ≤ 阈值):命中则仍入库,但标记 is_duplicate + dup_of。
|
||||
func (s *IngestService) dedupImage(ctx context.Context, kind, phashBits string, timing *fetchTiming) (dupID uint32, isDup uint8) {
|
||||
// repo 未注入(如离线单测)时跳过去重,不阻断主流程。生产环境 repo 必不为空。
|
||||
if s.repo == nil {
|
||||
return false, 0, 0
|
||||
return 0, 0
|
||||
}
|
||||
var tables []string
|
||||
switch kind {
|
||||
@ -533,61 +610,135 @@ func (s *IngestService) dedupImage(ctx context.Context, kind, sha1hash, phashBit
|
||||
case dto.IngestKindStreet:
|
||||
tables = []string{"street_snap_draft_images", "street_snap_images"}
|
||||
default:
|
||||
return false, 0, 0
|
||||
}
|
||||
if exists, err := s.repo.ImageExistsBySha1(ctx, tables, sha1hash); err == nil && exists {
|
||||
return true, 0, 0
|
||||
return 0, 0
|
||||
}
|
||||
if phashBits != "" {
|
||||
if id, found, err := s.repo.FindNearDuplicateImage(ctx, tables, phashBits, phash.DefaultThreshold); err == nil && found {
|
||||
return false, id, 1
|
||||
start := time.Now()
|
||||
id, found, err := s.repo.FindNearDuplicateImage(ctx, tables, phashBits, phash.DefaultThreshold)
|
||||
if timing != nil {
|
||||
timing.dedup += time.Since(start)
|
||||
}
|
||||
if err == nil && found {
|
||||
return id, 1
|
||||
}
|
||||
}
|
||||
return false, 0, 0
|
||||
return 0, 0
|
||||
}
|
||||
|
||||
// downloadOne 把单张远程图下载后上传到存储(七牛优先,失败兜底本地),
|
||||
// 返回 (访问地址, 对象 key, 内容 sha1, dHash 向量串, error)。访问地址可直接写入数据库
|
||||
// (七牛为完整 https URL,本地为相对 /uploads 路径)。
|
||||
func (s *IngestService) downloadOne(ctx context.Context, u string, idx int, prefix string) (string, string, string, string, error) {
|
||||
// downloadResult 是单张图「下载 → phash → 上传」的结果,供并发执行后按下标串行组装。
|
||||
// 各阶段耗时随结果一并带回、由调用方在串行段汇总(而非直接累加到共享 timing),
|
||||
// 这样并发路径上无需加锁即可保证计时准确。
|
||||
type downloadResult struct {
|
||||
url string // 存储访问地址(失败为空)
|
||||
key string // 对象 key(内容寻址)
|
||||
sha1 string // 内容 sha1,用于批次内去重
|
||||
phashBits string // dHash 向量串,空串表示无法解码
|
||||
download time.Duration // 本张下载耗时
|
||||
phash time.Duration // 本张指纹计算耗时
|
||||
upload time.Duration // 本张上传耗时
|
||||
err error // 下载或上传失败原因(非空即该张失败)
|
||||
}
|
||||
|
||||
// concurrentFetch 以 imageFetchConcurrency 路并发执行 fn(ctx, i)(i ∈ [0,n)),返回长度 n、
|
||||
// 下标与入参一一对齐的结果切片。
|
||||
//
|
||||
// 下标对齐是刻意的:调用方随后按原始顺序串行组装,使 cover / sort_order / 批次内去重结果
|
||||
// 与串行实现完全一致,只把「下载+上传」这段 IO 并行化。
|
||||
// 每个下标只由一个 goroutine 写入,故无需加锁。
|
||||
func concurrentFetch(ctx context.Context, n int, fn func(context.Context, int) downloadResult) []downloadResult {
|
||||
results := make([]downloadResult, n)
|
||||
if n == 0 {
|
||||
return results
|
||||
}
|
||||
sem := make(chan struct{}, imageFetchConcurrency)
|
||||
var wg sync.WaitGroup
|
||||
for i := 0; i < n; i++ {
|
||||
wg.Add(1)
|
||||
go func(i int) {
|
||||
defer wg.Done()
|
||||
sem <- struct{}{} // 占坑:超过并发度即在此排队
|
||||
defer func() { <-sem }() // 释放坑位
|
||||
results[i] = fn(ctx, i)
|
||||
}(i)
|
||||
}
|
||||
wg.Wait()
|
||||
return results
|
||||
}
|
||||
|
||||
// lookImageTask 是 fetchLookImages 摊平后的一张待下载任务:记录它属于哪个 look、
|
||||
// 是主图还是第几张细节图,供并发下载完成后重建 Name / LookIndex / IsDetail。
|
||||
type lookImageTask struct {
|
||||
url string
|
||||
lookIdx int // 1-based look 序号
|
||||
isDetail bool // true=细节图
|
||||
detailNo int // 细节图序号(1-based),主图为 0
|
||||
}
|
||||
|
||||
// downloadAndUpload 把单张远程图下载后上传到存储(S4优先,失败兜底本地)。
|
||||
// 内容 sha1 用于生成内容寻址的存储 key(重爬幂等)与批次内去重;访问地址可直接写入数据库
|
||||
// (S4为完整 https URL,本地为相对 /uploads 路径)。
|
||||
//
|
||||
// 本函数是并发调用点:每次调用各自返回一份独立的 downloadResult,不共享可变状态;
|
||||
// 耗时汇总与去重留痕由调用方在串行段统一完成。
|
||||
func (s *IngestService) downloadAndUpload(ctx context.Context, u, prefix string) downloadResult {
|
||||
res := downloadResult{}
|
||||
|
||||
// ① 下载
|
||||
dlStart := time.Now()
|
||||
resp, err := s.httpClient.Get(u)
|
||||
if err != nil {
|
||||
return "", "", "", "", err
|
||||
res.err = err
|
||||
return res
|
||||
}
|
||||
defer resp.Body.Close()
|
||||
if resp.StatusCode != http.StatusOK {
|
||||
return "", "", "", "", fmt.Errorf("status %d", resp.StatusCode)
|
||||
res.err = fmt.Errorf("status %d", resp.StatusCode)
|
||||
return res
|
||||
}
|
||||
data, err := io.ReadAll(resp.Body)
|
||||
if err != nil {
|
||||
return "", "", "", "", err
|
||||
res.err = err
|
||||
return res
|
||||
}
|
||||
res.download = time.Since(dlStart)
|
||||
|
||||
ext := path.Ext(u)
|
||||
if ext == "" || len(ext) > 5 {
|
||||
ext = ".jpg"
|
||||
}
|
||||
// 内容哈希作 key:同一张图(无论来自哪篇文章/job)永远得到相同 key,
|
||||
// 七牛覆盖写即天然幂等,不会因重复采集、崩溃重试、reject 重爬而累积孤儿文件。
|
||||
// S4覆盖写即天然幂等,不会因重复采集、崩溃重试、reject 重爬而累积孤儿文件。
|
||||
h := sha1.Sum(data)
|
||||
hash := hex.EncodeToString(h[:])
|
||||
key := fmt.Sprintf("%s/%s%s", prefix, hash, ext)
|
||||
res.sha1 = hex.EncodeToString(h[:])
|
||||
res.key = fmt.Sprintf("%s/%s%s", prefix, res.sha1, ext)
|
||||
ct := resp.Header.Get("Content-Type")
|
||||
|
||||
// 计算 dHash 指纹(仅用于近重复检索;解码失败返回空串 → NULL,不参与检索)。
|
||||
ph := phash.Of(data)
|
||||
phashBits := phash.ToVectorBits(ph)
|
||||
// ② dHash 指纹(仅用于近重复检索;解码失败返回空串 → NULL,不参与检索)。
|
||||
phStart := time.Now()
|
||||
res.phashBits = phash.ToVectorBits(phash.Of(data))
|
||||
res.phash = time.Since(phStart)
|
||||
|
||||
// 主上传器(七牛)
|
||||
if url, err := s.uploader.Upload(ctx, key, data, ct); err == nil {
|
||||
return url, key, hash, phashBits, nil
|
||||
// ③ 上传(主存储 S4,失败兜底本地)
|
||||
upStart := time.Now()
|
||||
if url, uerr := s.uploader.Upload(ctx, res.key, data, ct); uerr == nil {
|
||||
res.url = url
|
||||
} else if s.local != nil {
|
||||
// 兜底本地,避免图片完全丢失
|
||||
if lurl, lerr := s.local.Upload(ctx, key, data, ct); lerr == nil {
|
||||
return lurl, key, hash, phashBits, nil
|
||||
if lurl, lerr := s.local.Upload(ctx, res.key, data, ct); lerr == nil {
|
||||
res.url = lurl
|
||||
} else {
|
||||
return "", "", "", "", err
|
||||
res.err = uerr
|
||||
}
|
||||
} else {
|
||||
res.err = uerr
|
||||
}
|
||||
return "", "", "", "", err
|
||||
}
|
||||
res.upload = time.Since(upStart)
|
||||
|
||||
// 单图异常慢(合计超阈值)单独告警:便于从几十上百张图里一眼锁定是哪张、卡在哪一段。
|
||||
if total := res.download + res.phash + res.upload; total > slowImageThreshold {
|
||||
log.Printf("[ingest] 慢图 url=%s 合计=%v 下载=%v phash=%v 上传=%v",
|
||||
u, total.Round(time.Millisecond),
|
||||
res.download.Round(time.Millisecond), res.phash.Round(time.Millisecond), res.upload.Round(time.Millisecond))
|
||||
}
|
||||
return res
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user