diff --git a/.agents/notes/implemented/bug-fix/2026-11-21-recall-owner-lock-deadlock-and-read-batching.md b/.agents/notes/implemented/bug-fix/2026-11-21-recall-owner-lock-deadlock-and-read-batching.md index 90952d0..1b843d2 100644 --- a/.agents/notes/implemented/bug-fix/2026-11-21-recall-owner-lock-deadlock-and-read-batching.md +++ b/.agents/notes/implemented/bug-fix/2026-11-21-recall-owner-lock-deadlock-and-read-batching.md @@ -51,7 +51,11 @@ Status: implemented - `PUT /markers/rename`:逐条标记一次 `exists` + 一次 `update` → 一次预取目标标签命中的对象、一条批量 `UPDATE`、一条批量 `DELETE`。 - 隐私清算按范围类型过滤:标签改名/标签变化只解析 `scope_type='TAG'` 的回忆,删任务只解析该任务的 `TASK` 范围回忆, 不再整表解析 `document_json`。 -- `GET /review/report/{id}/structured`:按归属分组时逐个查任务名 → 先收集 ownerTaskNum,一次 `IN` 取齐。 +- `GET /review/report/{id}/structured`:按归属分组时逐个查任务名 → 先收集 ownerTaskNum,一次 `IN` 取齐(会话任务名并入同一次)。 +- 回忆删除与隐私抹除:对每条 AI 请求各查各改(每条请求 2–3 条语句)→ 整份回忆一次批量失效 + (`invalidateTasks`:查远程任务号 + 改请求 + 改 AI 任务,共 3 条),`deleteUnreadable` 的逐条删除也改成一条。 +- `GET /review/feed` 的随机/智能取样:`selectPage` 白算了一条 COUNT → `Page(current, size, false)`。 +- 标签范围校验加载标记时只取 `object_num` / `marker_name` 两列,不再把审计列一起搬进内存。 ## Alternatives considered @@ -64,6 +68,16 @@ Status: implemented 真正需要跨表的是详情里的两批装饰,`IN` 已经够用,JOIN 只会把 JSON 文档拆成多行再在内存里重组,反而更慢。 - **把 `reconcile` 挪到定时任务、读路径不再校验。** 事务最短,但来源被删除后列表会继续显示 `VALID`/可用正文,直到下一轮扫描,不符合隐私口径。 +## 索引补齐 + +`V20261121_1__add_query_indexes.sql` 补两条被漏掉的索引(2026-11-21 复查发现,不是本次批量化引入的问题): + +- `study_expectations (created_by, session_num)`:这张表此前只有主键,而报告分页、报告详情、会话详情都按会话号查它,等于每次全表扫。 +- `review_recalls (created_by, deleted, last_modified_time, id)`:原有索引以 `scope_type` / `scope_key` 打头,未指定范围时(首页的「最近回忆」、 + `status=ACTIVE` 取草稿)无法利用排序,只能取完 `created_by` 前缀再 filesort;`status` 由索引外过滤(活跃草稿通常就是最近修改的那条)。 + +两条索引都按「租户拦截器统一追加 `created_by`」的约定以 `created_by` 打头;键长远小于 InnoDB 的 3072 字节上限,`MigrationIndexKeyLengthTest` 已复算。 + ## Consequences 列表仍是「每个用户串行」的读路径:并发两个列表请求时第二个会等第一个提交,只是等待从几十秒降到几百毫秒。 @@ -75,6 +89,10 @@ Status: implemented 优先级批量重算不再经过 `updateById`,因此自动填充不再参与:`last_modified_time` 由调用方传入,`last_modified_by` / `device_info` 只在取到身份时写入。 新增「任务优先级」批量 SQL 时必须保留 `deleted = 0`(自定义 SQL 不会自动追加逻辑删除条件),租户条件仍由拦截器追加。 +已知的下一层瓶颈(语句数已经是常数,但数据量随用户增长):回忆读路径仍会整表取「可达碎片 id + 用户标签」来做范围校验与可达性判断 +(`RecallReferenceLoader`),隐私清算 `sourceDeleted` 仍需扫用户全部回忆并解析 `document_json`。要再降一档需要一张「来源 → 回忆」的索引表, +或把范围校验下推成聚合 SQL,属于独立改动。 + ## Verification - 真实 MySQL 并发复现:脚本对同一用户并发发两个列表请求,旧锁形态 6/6 轮死锁,新锁形态 0/12 次失败;`performance_schema.data_locks` 中不再出现同事务的 S+X 并存。 @@ -83,3 +101,6 @@ Status: implemented 资料清除、AI 结果装饰等断言,以及新增的 `PriorityWeightsIntegrationTest`(真实事务里校验批量 CASE 更新写回的值与审计时间)。 - 批量 `CASE id` 更新语句在真实 MySQL 上单独执行验证(同一事务内 affected=2 后回滚),确认语法与索引条件可用;H2 集成测试同时覆盖了 租户拦截器对这条 `UPDATE` 的改写。 +- 真实 MySQL 上跑过一遍首页会话的读接口并数语句:`GET /review/feed` 4 条、`GET /review/summary` 3 条、`GET /review/tasks` 6 条、 + `GET /tasks` 2 条、`GET /tasks/options` 1 条、`GET /report-fragments/pending-ownership` 1 条、`GET /study-sessions/active` 2 条, + 均不随行数增长;迁移在真实库执行成功(`flyway_schema_history` 20261121.1,两条索引已建)。 diff --git a/src/main/java/com/guo/learningprogresstracker/service/recall/RecallReferenceLoader.java b/src/main/java/com/guo/learningprogresstracker/service/recall/RecallReferenceLoader.java index a94021a..4fe4bd9 100644 --- a/src/main/java/com/guo/learningprogresstracker/service/recall/RecallReferenceLoader.java +++ b/src/main/java/com/guo/learningprogresstracker/service/recall/RecallReferenceLoader.java @@ -180,9 +180,11 @@ public class RecallReferenceLoader { } private List actualMarkers() { + // 只取判断范围要用到的两列:标签表是整用户量级,不裁列会把审计列一起搬进内存 return markers.selectList(Wrappers.lambdaQuery() .eq(MarkerEntity::getObjectType, "FRAGMENT") - .eq(MarkerEntity::getKind, "USER")); + .eq(MarkerEntity::getKind, "USER") + .select(MarkerEntity::getObjectNum, MarkerEntity::getMarkerName)); } public List selectedFragments(ScopeSnapshot scope) { diff --git a/src/main/java/com/guo/learningprogresstracker/service/recall/ReviewRecallStore.java b/src/main/java/com/guo/learningprogresstracker/service/recall/ReviewRecallStore.java index d716e58..a91aa8f 100644 --- a/src/main/java/com/guo/learningprogresstracker/service/recall/ReviewRecallStore.java +++ b/src/main/java/com/guo/learningprogresstracker/service/recall/ReviewRecallStore.java @@ -23,6 +23,7 @@ import com.guo.learningprogresstracker.service.ai.RecallJson; import java.time.Instant; import java.time.LocalDateTime; import java.time.ZoneOffset; +import java.util.Collection; import java.util.HashSet; import java.util.List; import java.util.Objects; @@ -92,12 +93,12 @@ public class ReviewRecallStore { public void deleteUnreadable(ReviewRecallEntity row) { List linked = requests.selectList(Wrappers.lambdaQuery() .eq(ReviewRecallRequestEntity::getRecallNum, row.getRecallNum())); - for (ReviewRecallRequestEntity request : linked) { - if (request.getTaskNum() != null) { - invalidateTask(request.getTaskNum()); - } - requests.deleteById(request.getId()); - } + invalidateTasks(linked.stream() + .map(ReviewRecallRequestEntity::getTaskNum) + .filter(Objects::nonNull) + .toList()); + requests.delete(Wrappers.lambdaQuery() + .eq(ReviewRecallRequestEntity::getRecallNum, row.getRecallNum())); receipts.delete(Wrappers.lambdaQuery() .eq(ReviewRecallReceiptEntity::getRecallNum, row.getRecallNum())); recalls.update( @@ -255,9 +256,7 @@ public class ReviewRecallStore { } public void redactRequests(RecallDocument doc) { - for (RecallRequest request : doc.getDetail().getRequests()) { - invalidateTask(request.taskNum()); - } + invalidateTasks(taskNums(doc)); doc.getDetail() .setRequests(doc.getDetail().getRequests().stream() .map(r -> new RecallRequest( @@ -276,22 +275,29 @@ public class ReviewRecallStore { } public void invalidateTask(String taskNum) { - ReviewRecallRequestEntity req = request(taskNum); - if (req != null) { + invalidateTasks(List.of(taskNum)); + } + + /** 一批 AI 任务一起失效:整份回忆只发 3 条语句,不再每条请求各查各改。 */ + public void invalidateTasks(Collection taskNums) { + if (taskNums.isEmpty()) { + return; + } + List live = requests.selectList(Wrappers.lambdaQuery() + .in(ReviewRecallRequestEntity::getTaskNum, taskNums) + .select(ReviewRecallRequestEntity::getId, ReviewRecallRequestEntity::getRemoteTaskId)); + if (!live.isEmpty()) { requests.update( null, Wrappers.lambdaUpdate() - .eq(ReviewRecallRequestEntity::getId, req.getId()) + .in(ReviewRecallRequestEntity::getTaskNum, taskNums) .set(ReviewRecallRequestEntity::getRedacted, 1) .set(ReviewRecallRequestEntity::getParamsJson, null)); - if (req.getRemoteTaskId() != null) { - purgeAfterCommit(req.getRemoteTaskId()); - } } aiTasks.update( null, Wrappers.lambdaUpdate() - .eq(AiTaskEntity::getTaskNum, taskNum) + .in(AiTaskEntity::getTaskNum, taskNums) .in(AiTaskEntity::getAiType, "RECALL_HINT", "RECALL_CHECK") .set(AiTaskEntity::getStatus, "FAILED") .set(AiTaskEntity::getRequestPayload, null) @@ -299,6 +305,19 @@ public class ReviewRecallStore { .set(AiTaskEntity::getErrorCode, "SOURCE_REDACTED") .set(AiTaskEntity::getErrorMessage, "资料已清除,这次 AI 结果不再展示") .set(AiTaskEntity::getFinishedTime, LocalDateTime.now())); + for (ReviewRecallRequestEntity req : live) { + if (req.getRemoteTaskId() != null) { + purgeAfterCommit(req.getRemoteTaskId()); + } + } + } + + /** 文档里记录过的 AI 任务号(可能为空),用于批量失效。 */ + private static List taskNums(RecallDocument doc) { + return doc.getDetail().getRequests().stream() + .map(RecallRequest::taskNum) + .filter(Objects::nonNull) + .toList(); } public void purgeAfterCommit(String remoteId) { @@ -315,9 +334,7 @@ public class ReviewRecallStore { } public void delete(RecallDocument doc) { - for (RecallRequest r : doc.getDetail().getRequests()) { - invalidateTask(r.taskNum()); - } + invalidateTasks(taskNums(doc)); recalls.update( null, Wrappers.lambdaUpdate() diff --git a/src/main/java/com/guo/learningprogresstracker/service/review/ReviewFeedService.java b/src/main/java/com/guo/learningprogresstracker/service/review/ReviewFeedService.java index 2199e2f..ca57386 100644 --- a/src/main/java/com/guo/learningprogresstracker/service/review/ReviewFeedService.java +++ b/src/main/java/com/guo/learningprogresstracker/service/review/ReviewFeedService.java @@ -77,7 +77,8 @@ public class ReviewFeedService { ? randomFragments(safeLimit) : studyReportFragmentsMapper .selectPage( - new Page<>(1, safeLimit), + // 只要前 N 条:关掉分页的 COUNT,别为一次取样多查一条 + new Page<>(1, safeLimit, false), Wrappers.lambdaQuery() .orderByDesc(StudyReportFragmentsEntity::getCreatedTime)) .getRecords(); @@ -121,7 +122,7 @@ public class ReviewFeedService { } List window = studyReportFragmentsMapper .selectPage( - new Page<>(1, count), + new Page<>(1, count, false), Wrappers.lambdaQuery() .ge(StudyReportFragmentsEntity::getId, startId) .orderByAsc(StudyReportFragmentsEntity::getId)) diff --git a/src/main/resources/db/migration/V20261121_1__add_query_indexes.sql b/src/main/resources/db/migration/V20261121_1__add_query_indexes.sql new file mode 100644 index 0000000..f25b923 --- /dev/null +++ b/src/main/resources/db/migration/V20261121_1__add_query_indexes.sql @@ -0,0 +1,9 @@ +-- 查询索引补齐(2026-11-21 的查询性能排查)。 +-- 说明:所有 SELECT/UPDATE/DELETE 都会由租户拦截器追加 created_by = ?,因此复合索引统一以 created_by 打头。 + +-- 学习预期:按会话取一条(报告分页逐页取、报告详情、会话详情)。此前只有主键,按 session_num 查等于全表扫。 +CREATE INDEX idx_expectations_session ON study_expectations (created_by, session_num); + +-- 回忆列表:按最近修改翻页(含 status 筛选)。原有索引以 scope_type/scope_key 打头, +-- 未指定范围时用不上排序,只能靠 created_by 前缀取完再 filesort。 +CREATE INDEX idx_recall_list ON review_recalls (created_by, deleted, last_modified_time, id);