diff --git a/backend-java/src/main/java/com/nanri/aiimage/config/RequestTraceFilter.java b/backend-java/src/main/java/com/nanri/aiimage/config/RequestTraceFilter.java index c9b3c1c7c..ce9f4bb9f 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/config/RequestTraceFilter.java +++ b/backend-java/src/main/java/com/nanri/aiimage/config/RequestTraceFilter.java @@ -7,6 +7,7 @@ import jakarta.servlet.http.HttpServletResponse; import java.util.Locale; import org.slf4j.Logger; import org.slf4j.LoggerFactory; +import org.slf4j.MDC; import org.springframework.beans.factory.annotation.Value; import org.springframework.stereotype.Component; import org.springframework.web.filter.OncePerRequestFilter; @@ -18,6 +19,13 @@ import java.util.Optional; @Component public class RequestTraceFilter extends OncePerRequestFilter { + /** + * AC-21(审计 §5.2):requestId 早在 2026-09 就被解析出来了,却只塞进这一行日志的参数里, + * 全仓没有一处 {@code MDC.put} ⇒ 一个请求里另外 570 处日志互相认不出彼此。 + * 入 MDC 之后,logback 的 {@code %X{requestId}} 把整条请求链缝起来(含异步 appender 落盘)。 + */ + static final String MDC_REQUEST_ID = "requestId"; + private static final Logger log = LoggerFactory.getLogger(RequestTraceFilter.class); private static final int MIN_REQUEST_BODY_CACHE_LIMIT_BYTES = 1024 * 1024; @@ -55,6 +63,9 @@ public class RequestTraceFilter extends OncePerRequestFilter { response.setHeader("X-AIIMAGE-Instance-Source", instanceMetadata.getSource()); response.setHeader("X-AIIMAGE-Instance-Stable", String.valueOf(instanceMetadata.isStable())); + // tomcat 线程是复用的:不清 MDC 的话,下一条请求(哪怕没有 requestId)会带上上一条的 id, + // 按请求关联反而变成按请求串案。缺失也写 "-",让 pattern 里这一位永远存在、格式不抖。 + MDC.put(MDC_REQUEST_ID, blankToDash(requestId)); try { filterChain.doFilter(requestToUse, response); } finally { @@ -96,6 +107,7 @@ public class RequestTraceFilter extends OncePerRequestFilter { costMs ); } + MDC.clear(); } } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskHeartbeatSpi.java b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskHeartbeatSpi.java index d2eec08fd..9aa82db26 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskHeartbeatSpi.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/AppearancePatentTaskHeartbeatSpi.java @@ -50,6 +50,12 @@ public class AppearancePatentTaskHeartbeatSpi implements TaskModuleHeartbeatSpi task.getId(), moduleType(), jobTouchIntervalMillis()); } + /** 只把 {@code task.getId()} 交给作业续期,不写模块缓存 ⇒ 心跳可以吃瘦读(AC-19/P1-B3)。 */ + @Override + public boolean heartbeatNeedsFullTaskEntity() { + return false; + } + private long jobTouchIntervalMillis() { return Math.max(1_000L, appearancePatentProperties.getDbJobTouchIntervalMillis()); } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/support/AppearancePatentDeliverySupport.java b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/support/AppearancePatentDeliverySupport.java index a88d54403..00911f9f3 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/support/AppearancePatentDeliverySupport.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/appearancepatent/service/support/AppearancePatentDeliverySupport.java @@ -17,6 +17,8 @@ import lombok.extern.slf4j.Slf4j; import java.util.ArrayList; import java.util.List; import java.util.Map; +import java.util.concurrent.ConcurrentHashMap; +import java.util.concurrent.atomic.AtomicLong; /** * 外观专利的交付闸门与补检(E 聚类之一),从 {@code AppearancePatentTaskService} 原样搬移 @@ -30,6 +32,20 @@ public class AppearancePatentDeliverySupport { private static final String STATUS_FAILED = TaskStatuses.FileTask.FAILED; + /** + * AC-21(审计 §5、PERF_REPORT §6):交付缺口补检 WARN 的按任务限流窗口。 + * + *

现网实录 task 32609/job 23345 三分钟 478 条(≈9,560 条/小时):结果文件作业每 15s + * 被 claim 一轮,同一批缺口行每轮原样再报一次,字段一字不差,日志里没有新信息,只有量。 + * + *

本类按调用现取(不是 Spring bean,{@code AppearancePatentTaskService#deliverySupport()} + * 每次 new 一只),所以限流态只能挂静态:一个 JVM 一份,两台实例各打各的,不互相吞计数。 + * 条目数由 {@link #WARN_TASKS_MAX} 兜住,超了就先把窗口已过期的清出去。 + */ + static final long RECHECK_WARN_WINDOW_MILLIS = 300_000L; + static final int WARN_TASKS_MAX = 1024; + private static final ConcurrentHashMap RECHECK_WARN_WINDOWS = new ConcurrentHashMap<>(); + private final String moduleType; private final AppearancePatentProperties properties; private final AppearancePatentTaskCacheService taskCacheService; @@ -120,8 +136,7 @@ public class AppearancePatentDeliverySupport { boolean requeued = taskFileJobService.requeue(job.getId(), "有 " + facts.missingRows() + " 行未完成检测,补检一轮后重新组装结果文件"); if (requeued) { - log.warn("[appearance-patent] 交付缺口触发补检轮 taskId={} jobId={} missingRows={} keys={} noDataRows={}", - task.getId(), job.getId(), facts.missingRows(), keys.size(), facts.noDataRows()); + throttleRecheckWarning(task, job, facts); return true; } log.warn("[appearance-patent] 交付缺口补检排不上队,转失败兜底出口按补检预算处置 taskId={} jobId={} missingRows={}", @@ -129,6 +144,47 @@ public class AppearancePatentDeliverySupport { throw new BusinessException("有 " + facts.missingRows() + " 行未完成检测,等待补检轮"); } + /** + * 补检轮 WARN 的限流出口(AC-21):同一任务在一个窗口内只留一条 WARN,窗口里被压掉的条数 + * 随下一条 WARN 一起汇总出去({@code suppressedInLastWindow}),窗口内的重复轮次降 DEBUG。 + * + *

刻意不去碰 {@code taskCacheService.recheckRounds/countRecheckRound} 那对轮次计数: + * 那是 {@link #scheduleFallbackRecheck} 的补检预算({@code recheckMaxRounds})。从这条路径 + * 再计一次,会让「日志限流」吃掉业务预算,任务更早被判「补检已耗尽」——那是行为变更,不是治理。 + * + *

并发口径:读-判-写三步之间有竞争,最坏情况是同窗口多打一条 WARN。限流是为了压日志量, + * 不是为了恰好一条,不值得为它加锁或把整张表串行化。 + */ + private void throttleRecheckWarning(FileTaskEntity task, TaskFileJobEntity job, + AppearancePatentRowOutcome.Facts facts) { + long now = System.currentTimeMillis(); + RecheckWarnWindow current = RECHECK_WARN_WINDOWS.get(task.getId()); + if (current != null && now - current.startedAtMillis < RECHECK_WARN_WINDOW_MILLIS) { + log.debug("[appearance-patent] 交付缺口触发补检轮 taskId={} jobId={} missingRows={} keys={}(限流:窗口内第 {} 条)", + task.getId(), job.getId(), facts.missingRows(), facts.missingKeys().size(), + current.suppressed.incrementAndGet()); + return; + } + log.warn("[appearance-patent] 交付缺口触发补检轮 taskId={} jobId={} missingRows={} keys={} noDataRows={} suppressedInLastWindow={}", + task.getId(), job.getId(), facts.missingRows(), facts.missingKeys().size(), + facts.noDataRows(), current == null ? 0L : current.suppressed.get()); + if (RECHECK_WARN_WINDOWS.size() >= WARN_TASKS_MAX) { + RECHECK_WARN_WINDOWS.entrySet().removeIf(entry -> + now - entry.getValue().startedAtMillis >= RECHECK_WARN_WINDOW_MILLIS); + } + RECHECK_WARN_WINDOWS.put(task.getId(), new RecheckWarnWindow(now)); + } + + /** 一个任务在一个限流窗口里的补检告警态:窗口起始时刻 + 窗口内被压到 DEBUG 的条数。 */ + private static final class RecheckWarnWindow { + private final long startedAtMillis; + private final AtomicLong suppressed = new AtomicLong(); + + private RecheckWarnWindow(long startedAtMillis) { + this.startedAtMillis = startedAtMillis; + } + } + /** * 只量行账、不出表。补检的决定必须做在 {@code assembleResultWorkbook} 之前,而那一跑完就把 * 成品传上 OSS 并挂好了 {@code resultFileUrl}。判据与出表共用 diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskHeartbeatSpi.java b/backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskHeartbeatSpi.java index 03021d0d5..e0aaaf8bf 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskHeartbeatSpi.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/collectdata/service/CollectDataTaskHeartbeatSpi.java @@ -25,4 +25,13 @@ public class CollectDataTaskHeartbeatSpi implements TaskModuleHeartbeatSpi { public void touchHeartbeat(Long taskId, TaskHeartbeatRequest request) { collectDataService.updateProgress(taskId, request); } + + /** + * 进度写走 {@code updateProgress(taskId, request)} 自己按主键取数,心跳不需要整实体 + * (AC-19/P1-B3:采集任务的心跳最密,两列大 JSON 每 30s 捞一次纯属白付)。 + */ + @Override + public boolean heartbeatNeedsFullTaskEntity() { + return false; + } } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/keywordshare/service/KeywordShareService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/keywordshare/service/KeywordShareService.java index ff05572a9..65a8821cb 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/keywordshare/service/KeywordShareService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/keywordshare/service/KeywordShareService.java @@ -41,11 +41,13 @@ import org.springframework.transaction.annotation.Transactional; import java.time.LocalDate; import java.time.LocalDateTime; import java.util.ArrayList; +import java.util.Collections; import java.util.LinkedHashMap; import java.util.LinkedHashSet; import java.util.List; import java.util.Map; import java.util.Set; +import java.util.concurrent.ThreadLocalRandom; /** * 五国关键词共享(桌面 Tkinter 工具的 web 化)。 @@ -200,12 +202,23 @@ public class KeywordShareService { Long count) { KeywordShareFilterQuery filter = buildFilter(companyId, viewerUserName, country, category, subCategory, categoryL3, categoryL4, search, FILTER_UNUSED, null, false); - filter.setOffset(0); long safeCount = count == null ? DEFAULT_RANDOM_COUNT : Math.min(Math.max(count, 1), MAX_RANDOM_COUNT); filter.setLimit(safeCount); - return keywordShareMapper.selectRandomUnused(filter).stream() + // AC-19(审计 P1-B2):窗口起点先数命中集再掷,MySQL 就不必给全表打 RAND 后 filesort。 + // 掷的是起点而非每一行,所以同一批词的相邻关系仍在(一次抽到的是连续 id 的一段), + // 服务层洗牌补回"散着"的观感;命中集不足一批时起点恒 0,与旧口径同为"全给你"。 + filter.setOffset(randomWindowOffset(keywordShareMapper.countHit(filter), safeCount)); + List window = new ArrayList<>(keywordShareMapper.selectRandomUnused(filter).stream() .map(this::normalizeRow) - .toList(); + .toList()); + Collections.shuffle(window); + return window; + } + + /** 随机窗口起点:落在 {@code [0, total - count]},装得下整批才有得掷。 */ + static long randomWindowOffset(long total, long count) { + long span = total - count; + return span > 0 ? ThreadLocalRandom.current().nextLong(span + 1) : 0L; } // ---------------------------------------------------------------- 写操作 diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/permission/service/PermissionMenuService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/permission/service/PermissionMenuService.java index 0aa276fdc..40976fefc 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/permission/service/PermissionMenuService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/permission/service/PermissionMenuService.java @@ -261,6 +261,7 @@ public class PermissionMenuService { entity.setCompanyId(companyId); entity.setSortOrder(resolveSortOrder(companyId, menuType, request.getSortOrder(), null)); permissionMenuMapper.insert(entity); + invalidateMenuTreeCache(); return toItemVo(getMenuById(entity.getId())); } @@ -300,6 +301,7 @@ public class PermissionMenuService { entity.setRoutePath(routePath); entity.setSortOrder(resolveSortOrder(companyId, menuType, request.getSortOrder(), id)); permissionMenuMapper.updateById(entity); + invalidateMenuTreeCache(); return toItemVo(getMenuById(id)); } @@ -321,6 +323,7 @@ public class PermissionMenuService { // installations without the FK are cleaned up as well. userColumnPermissionMapper.deleteByColumnId(entity.getId()); permissionMenuMapper.deleteById(entity.getId()); + invalidateMenuTreeCache(); } /** @@ -376,6 +379,7 @@ public class PermissionMenuService { } sort += 10; } + invalidateMenuTreeCache(); } /** @@ -457,6 +461,7 @@ public class PermissionMenuService { } log.info("[menu-tree-copy] 新公司菜单树已建 companyId={} template={} rows={}", companyId, templateCompanyId, inserted); + invalidateMenuTreeCache(); return inserted; } @@ -1031,6 +1036,29 @@ public class PermissionMenuService { } } + /** + * AC-19(审计 P1-B4):目录树的短 TTL 缓存。 + * + *

侧边栏、每个受管入口的 {@code hasAnyAdminMenu}、可授予全集都从 {@code loadMenus} 取同一棵 + * 树,此前每请求重查一遍再在内存里排序展开。目录结构是"配一次、几个月不动"的数据, + * 30 秒的窗口换来热路径少一次全表取数。 + * + *

只缓存目录本体,**不缓存用户授权行**:撤权必须下一次请求就生效,那是安全语义; + * 目录结构最多陈旧一个 TTL(写路径都同机 {@code invalidateMenuTreeCache},跨实例靠 TTL 收敛)。 + * 键是 (companyId, menuType),company 维度来自库里的用户行而非请求参数,条数有界。 + */ + private static final long MENU_CACHE_TTL_MILLIS = 30_000L; + private final Map menuTreeCache = new java.util.concurrent.ConcurrentHashMap<>(); + + /** 一棵已排序的目录树快照与其入缓存的时刻。 */ + private record CachedMenuTree(List menus, long cachedAtMillis) { + } + + /** 目录写路径共用的失效出口:结构变了就把整份快照丢掉,下一次现取。 */ + private void invalidateMenuTreeCache() { + menuTreeCache.clear(); + } + /** * 目录取数的唯一出口(§4):{@code companyId} 非空时只返回那一棵树的行。 * 授权树、有效菜单、可授予全集、越权判定都从这里取数,所以公司收窄改这一处就够一片。 @@ -1038,14 +1066,23 @@ public class PermissionMenuService { * @param companyId null 只来自「无操作者 / 无归属」的历史与测试路径,此时退回全量目录 */ private List loadMenus(Long companyId, String menuType) { - List rows = loadRawMenus(companyId, menuType); - if (rows == null || rows.isEmpty()) { - return List.of(); + String cacheKey = (companyId == null ? "*" : companyId.toString()) + + '|' + normalizeMenuType(menuType); + long now = System.currentTimeMillis(); + CachedMenuTree cached = menuTreeCache.get(cacheKey); + if (cached != null && now - cached.cachedAtMillis() < MENU_CACHE_TTL_MILLIS) { + return cached.menus(); } - return rows.stream() - .filter(row -> row != null && row.getId() != null) - .sorted(menuComparator()) - .toList(); + List rows = loadRawMenus(companyId, menuType); + // 不可变快照:缓存出去的东西被调用方就地改一笔,下一次请求会读到别人的写入 + List sorted = rows == null || rows.isEmpty() + ? List.of() + : rows.stream() + .filter(row -> row != null && row.getId() != null) + .sorted(menuComparator()) + .toList(); + menuTreeCache.put(cacheKey, new CachedMenuTree(sorted, now)); + return sorted; } /** diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/service/SimilarAsinTaskHeartbeatSpi.java b/backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/service/SimilarAsinTaskHeartbeatSpi.java index bdd845873..9a2a0d926 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/service/SimilarAsinTaskHeartbeatSpi.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/similarasin/service/SimilarAsinTaskHeartbeatSpi.java @@ -47,6 +47,12 @@ public class SimilarAsinTaskHeartbeatSpi implements TaskModuleHeartbeatSpi { return Duration.ofMillis(checkpointIntervalMillis()); } + /** 只把 {@code task.getId()} 交给作业续期,不写模块缓存 ⇒ 心跳可以吃瘦读(AC-19/P1-B3)。 */ + @Override + public boolean heartbeatNeedsFullTaskEntity() { + return false; + } + private long checkpointIntervalMillis() { return Math.max(1_000L, similarAsinProperties.getDbTaskTouchIntervalMillis()); } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatService.java index 7341a1dd7..e9bfda122 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatService.java @@ -79,7 +79,7 @@ public class TaskHeartbeatService { boolean brandOnly = MODULE_BRAND.equalsIgnoreCase(requestedModuleType); boolean skipBrandLookup = MODULE_SHOP_DATA_CRAWL.equalsIgnoreCase(requestedModuleType); - TaskHeartbeatVo fileResult = brandOnly ? null : touchFileTaskIfRunning(selectFileTask(taskId), request); + TaskHeartbeatVo fileResult = brandOnly ? null : touchFileTaskIfRunning(taskId, request); TaskHeartbeatVo brandResult = skipBrandLookup ? null : brandTaskHeartbeatSpi.heartbeat(taskId, request); // 心跳到达打点:陈旧判死线用它把「心跳时钟」改成从首次心跳起算(见 StaleTaskHeartbeatPolicy)。 @@ -112,7 +112,26 @@ public class TaskHeartbeatService { return TaskHeartbeatVo.notAlive(null, null, "task not found"); } + /** + * AC-19(审计 P1-B3):心跳热路径的任务读只取活性判定要的五列。 + * + *

{@code biz_file_task.request_json}/{@code result_json} 能到 MB 级,而这条读是 + * 「每单每 30s 一次」的热路径,原实现整行捞回来只为读 id/moduleType/status/userId/updatedAt。 + * 需要把实体交给模块 SPI 的({@code saveTaskCache} 会把整个实体序列化进模块缓存), + * 由 {@link TaskModuleHeartbeatSpi#heartbeatNeedsFullTaskEntity()} 声明后走 + * {@link #selectFullFileTask}——半行进缓存比多读两列难查得多,所以宁可多一次主键点查。 + */ private FileTaskEntity selectFileTask(Long taskId) { + LambdaQueryWrapper fileQuery = new LambdaQueryWrapper() + .select(FileTaskEntity::getId, FileTaskEntity::getModuleType, FileTaskEntity::getStatus, + FileTaskEntity::getUserId, FileTaskEntity::getUpdatedAt) + .eq(FileTaskEntity::getId, taskId) + .last("limit 1"); + return fileTaskMapper.selectOne(fileQuery); + } + + /** 整实体读:客户端中断恢复(要 errorMessage,并把实体同步进模块缓存与快照)与声明需要整行的模块心跳。 */ + private FileTaskEntity selectFullFileTask(Long taskId) { LambdaQueryWrapper fileQuery = new LambdaQueryWrapper() .eq(FileTaskEntity::getId, taskId) .last("limit 1"); @@ -132,7 +151,7 @@ public class TaskHeartbeatService { String safeReason = (reason == null || reason.isBlank()) ? "客户端异常中断,任务已自动失败" : "客户端异常中断: " + reason.trim(); - FileTaskEntity fileTask = selectFileTask(taskId); + FileTaskEntity fileTask = selectFullFileTask(taskId); if (fileTask != null) { String status = fileTask.getStatus(); int updated = fileTaskMapper.update(null, new LambdaUpdateWrapper() @@ -167,12 +186,23 @@ public class TaskHeartbeatService { return TaskHeartbeatVo.notAlive(null, null, "task not found"); } - private TaskHeartbeatVo touchFileTaskIfRunning(FileTaskEntity task, TaskHeartbeatRequest request) { + private TaskHeartbeatVo touchFileTaskIfRunning(Long taskId, TaskHeartbeatRequest request) { + FileTaskEntity task = selectFileTask(taskId); if (task == null) { return null; } String moduleType = task.getModuleType() == null ? "" : task.getModuleType(); TaskModuleHeartbeatSpi handler = moduleHandlers.get(moduleType); + if (handler != null && handler.heartbeatNeedsFullTaskEntity()) { + // 归属校验与模块缓存回写都吃整实体(publish/shop-data-crawl 读实例路由列, + // 七个模块的 saveTaskCache 把实体序列化落盘),瘦行给它们会写出半行 ⇒ 补一次主键读。 + task = selectFullFileTask(taskId); + if (task == null) { + return null; + } + moduleType = task.getModuleType() == null ? "" : task.getModuleType(); + handler = moduleHandlers.get(moduleType); + } if (handler != null) { handler.ensureOwnership(task, moduleType + " task heartbeat"); } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/task/spi/TaskModuleHeartbeatSpi.java b/backend-java/src/main/java/com/nanri/aiimage/modules/task/spi/TaskModuleHeartbeatSpi.java index 4e5aaacab..246008dfe 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/task/spi/TaskModuleHeartbeatSpi.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/task/spi/TaskModuleHeartbeatSpi.java @@ -44,6 +44,21 @@ public interface TaskModuleHeartbeatSpi { default void onHeartbeat(FileTaskEntity task) { } + /** + * 心跳时是否需要整行 {@code FileTaskEntity}。 + * + *

AC-19(审计 P1-B3):心跳是「每单每 30s 一次」的热路径,而 {@code biz_file_task} 的 + * {@code request_json}/{@code result_json} 能到 MB 级。task 侧默认只捞活性判定要的五列, + * 只有真把实体交出去用的模块才付整行读的钱——即 {@link #saveTaskCache(FileTaskEntity)} + * 会把实体序列化进模块缓存的那些(半行进缓存比多读一次列难查得多)。 + * + *

默认 {@code true}(保守):新模块不改这儿就维持今天的取整行行为。只用得到 + * {@code task.getId()} 的实现才显式改成 false。 + */ + default boolean heartbeatNeedsFullTaskEntity() { + return true; + } + /** * biz_file_task.updated_at 检查点节流间隔。 * diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/mapper/TaskCenterQueryMapper.java b/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/mapper/TaskCenterQueryMapper.java index 9c2767e16..84d66d047 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/mapper/TaskCenterQueryMapper.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/mapper/TaskCenterQueryMapper.java @@ -217,6 +217,27 @@ public interface TaskCenterQueryMapper { + USER_SCOPE_B + ")"; + /** + * 页首游标锚点(AC-19/P1-B1):第 N 页那一行的前一行——按同一套筛选与同一个排序键 + * {@code (createdAt DESC, id DESC)} 走到 {@code offset} 处,只取两列。 + * + *

深翻页的 skip 只在这条窄语句里发生,页查询本身改用游标;两条必须共用同一份 UNION + * (mapper XML 的 {@code taskUnionRows}),否则锚点算的是另一批行,第 N 页会安静地给错。 + */ + TaskCenterTaskRow selectTaskPageAnchor(@Param("userIds") List userIds, + @Param("moduleType") String moduleType, + @Param("statuses") List statuses, + @Param("shop") String shop, + @Param("startTime") LocalDateTime startTime, + @Param("endTime") LocalDateTime endTime, + @Param("offset") long offset); + + /** + * 任务列表一页:keyset 游标取数,不再吃 OFFSET。 + * + *

{@code afterCreatedAt}/{@code afterId} 同时给值才是游标(第一页两个都为 null = 从头取); + * 游标条件能被我方的 MySQL 8.4 压进 UNION 分支,从而用上 created_at 索引。 + */ List selectTaskPage(@Param("userIds") List userIds, @Param("moduleType") String moduleType, @Param("statuses") List statuses, @@ -224,7 +245,8 @@ public interface TaskCenterQueryMapper { @Param("startTime") LocalDateTime startTime, @Param("endTime") LocalDateTime endTime, @Param("limit") long limit, - @Param("offset") long offset); + @Param("afterCreatedAt") LocalDateTime afterCreatedAt, + @Param("afterId") Long afterId); long countTaskPage(@Param("userIds") List userIds, @Param("moduleType") String moduleType, diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/service/TaskCenterQueryService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/service/TaskCenterQueryService.java index 6cd939e05..1c5850f43 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/service/TaskCenterQueryService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/taskcenter/service/TaskCenterQueryService.java @@ -51,6 +51,11 @@ public class TaskCenterQueryService { private static final long ACTIVE_TASK_SCAN_CAP = 500; private static final long DEFAULT_PAGE_SIZE = 20; private static final long MAX_PAGE_SIZE = 100; + /** + * AC-19(审计 P1-B1):任务中心默认时间窗(天)。请求不带 {@code startTime} 时由服务层补上, + * 页码跳转与总数都只在这个窗口里发生;带 {@code startTime} 就完全按请求走。 + */ + private static final int DEFAULT_WINDOW_DAYS = 90; /** * 心跳位置读取封顶,与页大小同源:看板一页最多 100 行、30 秒一轮。 * 沿用 light 侧那个 20 的话,第 21 行之后的在跑任务永远读不到进度。 @@ -117,15 +122,34 @@ public class TaskCenterQueryService { } List statuses = statusFilter(query.getStatus()); + // AC-19(审计 P1-B1):时间窗先补再传,count 与页查询共用同一个窗口口径。 + LocalDateTime startTime = startOfDefaultWindow(query.getStartTime()); + LocalDateTime endTime = query.getEndTime(); long total = queryMapper.countTaskPage(userIds, query.getModuleType(), statuses, - query.getShop(), query.getStartTime(), query.getEndTime()); + query.getShop(), startTime, endTime); if (total <= 0) { return TaskCenterTaskPageVo.empty(page, pageSize); } + // 深翻页的 skip 交给窄锚点语句(只投影 createdAt,id 两列),页查询本身走 keyset 游标: + // 游标条件能被 MySQL 压进 UNION 分支用上 created_at 索引,OFFSET 不能,只能物化整个 UNION。 + long offset = (page - 1) * pageSize; + LocalDateTime afterCreatedAt = null; + Long afterId = null; + if (offset > 0) { + TaskCenterTaskRow anchor = queryMapper.selectTaskPageAnchor(userIds, query.getModuleType(), + statuses, query.getShop(), startTime, endTime, offset); + if (anchor == null || anchor.getCreatedAt() == null || anchor.getId() == null) { + // 页码落在窗口之外:与旧实现取到空行同样返回空页,不拿 null 游标去要第一页 + return TaskCenterTaskPageVo.empty(page, pageSize); + } + afterCreatedAt = anchor.getCreatedAt(); + afterId = anchor.getId(); + } + List rows = queryMapper.selectTaskPage(userIds, query.getModuleType(), - statuses, query.getShop(), query.getStartTime(), query.getEndTime(), - pageSize, (page - 1) * pageSize); + statuses, query.getShop(), startTime, endTime, + pageSize, afterCreatedAt, afterId); if (rows == null || rows.isEmpty()) { return TaskCenterTaskPageVo.empty(page, pageSize); } @@ -138,6 +162,17 @@ public class TaskCenterQueryService { return vo; } + /** + * 默认时间窗起点(AC-19/P1-B1):没带 {@code startTime} 才补 {@code now - 90 天}。 + * + *

这一条给 UNION 的两个分支各加一条 {@code created_at} 下界,是把「全历史物化再排序」 + * 换成「按索引取窗口」的那一刀。代价是对管理员可见的行为收紧:90 天前的任务不再默认出现在 + * 列表与总数里,必须显式传 startTime 才能翻旧账——Phase 6 真机总验收要按这条复现一次翻页路径。 + */ + static LocalDateTime startOfDefaultWindow(LocalDateTime requested) { + return requested != null ? requested : LocalDateTime.now().minusDays(DEFAULT_WINDOW_DAYS); + } + /** * 状态筛选展开:空白 = 不裁剪;{@code ACTIVE} = 前端默认的「执行中」,等于三个进行中状态的合并 * (与判卡住用的 {@link #ACTIVE_STATUSES} 同一组,不另定口径);{@code SUCCESS}/{@code FAILED} diff --git a/backend-java/src/main/resources/application.yml b/backend-java/src/main/resources/application.yml index 3816fd357..128c154df 100644 --- a/backend-java/src/main/resources/application.yml +++ b/backend-java/src/main/resources/application.yml @@ -80,8 +80,10 @@ logging: root: INFO com.nanri.aiimage: INFO pattern: - file: "%d{yyyy-MM-dd HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n" - console: "%d{yyyy-MM-dd HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n" + # AC-21:与 logback-spring.xml 的默认值同口径(本文件显式设了这两条,默认值会被盖掉), + # %X{requestId} 由 RequestTraceFilter 入 MDC。两处必须一起改,漏一处线上就看不见请求链。 + file: "%d{yyyy-MM-dd HH:mm:ss.SSS} [%thread] %-5level [%X{requestId}] %logger{36} - %msg%n" + console: "%d{yyyy-MM-dd HH:mm:ss.SSS} [%thread] %-5level [%X{requestId}] %logger{36} - %msg%n" mybatis-plus: mapper-locations: classpath*:mapper/**/*.xml diff --git a/backend-java/src/main/resources/logback-spring.xml b/backend-java/src/main/resources/logback-spring.xml index 973add4f8..3c57e258c 100644 --- a/backend-java/src/main/resources/logback-spring.xml +++ b/backend-java/src/main/resources/logback-spring.xml @@ -4,8 +4,11 @@ - - + + + diff --git a/backend-java/src/main/resources/mapper/keywordshare/KeywordShareMapper.xml b/backend-java/src/main/resources/mapper/keywordshare/KeywordShareMapper.xml index f3eee8ffb..658703784 100644 --- a/backend-java/src/main/resources/mapper/keywordshare/KeywordShareMapper.xml +++ b/backend-java/src/main/resources/mapper/keywordshare/KeywordShareMapper.xml @@ -45,13 +45,17 @@ LIMIT #{f.offset}, #{f.limit} + SELECT t.id AS taskId, COALESCE(NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.filters.countryCode')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.countryCode')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.country')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.publish_country')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.countryCodes')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.country_codes')), ''), 'null'),NULLIF(NULLIF(JSON_UNQUOTE(JSON_EXTRACT(t.request_json, '$.filters.countryCodes')), ''), 'null')) AS country FROM biz_file_task t WHERE t.id IN #{tid} + + t.module_type IN ('PUBLISH','DELETE_BRAND','PRODUCT_RISK_RESOLVE','PRICE_TRACK','SHOP_MATCH','PATROL_DELETE','QUERY_ASIN','WITHDRAW','APPEARANCE_PATENT','SIMILAR_ASIN','COLLECT_DATA','SHOP_DATA_CRAWL') AND t.user_id IN #{uid} AND UPPER(t.module_type) = #{moduleType} AND CASE WHEN UPPER(t.status) = 'CANCELED' THEN 'CANCELLED' ELSE UPPER(t.status) END IN #{st} AND t.created_at >= #{startTime} AND t.created_at <= #{endTime} AND (EXISTS (SELECT 1 FROM biz_publish_file p WHERE p.task_id = t.id AND LOWER(TRIM(p.shop_name)) = LOWER(TRIM(#{shop}))) OR (t.module_type IN ('PATROL_DELETE','PRICE_TRACK','PRODUCT_RISK_RESOLVE','QUERY_ASIN','SHOP_DATA_CRAWL','SHOP_MATCH') AND EXISTS (SELECT 1 FROM biz_file_result f WHERE f.task_id = t.id AND LOWER(TRIM(f.source_filename)) = LOWER(TRIM(#{shop}))))) + 1 = 1 AND b.user_id IN #{uid} AND 'BRAND' = #{moduleType} AND CASE WHEN UPPER(b.status) = 'CANCELED' THEN 'CANCELLED' ELSE UPPER(b.status) END IN #{st} AND b.created_at >= #{startTime} AND b.created_at <= #{endTime} AND 1 = 0 + + (SELECT t.id AS id, t.task_no AS taskNo, UPPER(t.module_type) AS moduleType, CASE WHEN UPPER(t.status) = 'CANCELED' THEN 'CANCELLED' ELSE UPPER(t.status) END AS status, t.user_id AS userId, t.device_id AS deviceId, t.created_at AS createdAt, t.finished_at AS finishedAt, t.updated_at AS updatedAt, t.source_file_count AS totalCount, t.success_file_count AS successCount, t.failed_file_count AS failedCount, 0 AS brand FROM biz_file_task t WHERE UNION ALL SELECT b.id, CONCAT('brand-', b.id), 'BRAND', CASE WHEN UPPER(b.status) = 'CANCELED' THEN 'CANCELLED' ELSE UPPER(b.status) END, b.user_id, b.device_id, b.created_at, CASE WHEN b.status IN ('cancelled','failed','success') THEN b.updated_at END, b.updated_at, b.progress_total, b.progress_current, 0, 1 FROM brand_crawl_tasks b WHERE ) + + + (SELECT t.id AS id, t.created_at AS createdAt FROM biz_file_task t WHERE UNION ALL SELECT b.id, b.created_at FROM brand_crawl_tasks b WHERE ) + +