From e71fa7fa83fb78579b2c90f8eb8f7a3fd87ce39f Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?=E9=BB=84=E8=87=AA=E8=BE=BE?= <980324341@qq.com> Date: Wed, 4 Nov 2026 02:55:11 +0800 Subject: [PATCH] =?UTF-8?q?fix(phase-6.1):=20CR-P1=20=E6=92=9E=E6=AC=BE?= =?UTF-8?q?=E5=88=86=E9=A1=B5=E6=BA=A2=E5=87=BA=E9=92=B3=E5=88=B6=EF=BC=8B?= =?UTF-8?q?B74-02=20=E5=9C=A8=E9=A3=9E=E7=99=BB=E8=AE=B0=E6=B3=84=E6=BC=8F?= =?UTF-8?q?=E6=94=B6=E5=8F=A3?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Phase 6 全量 Code Review 的后端四处(四面合计 0 个 P0 / 7 个 P1,这是其中两处,另带两处邻接纠偏): 1. `AdminShopDataDuplicateCheckController.parsePage` 补页码上限(CR-18/铁律 15 同形, 参照 `KeywordUsageAdminService.MAX_PAGE`)。原先只钳下限:`page=11000000 / page_size=200` 溢出成 `-2094967296`,`Math.min` 之后仍是负数,喂进 `subList(负,负)` 就是 HTTP 500—— 表里有至少一行重复数据就能打出来。这一处是 Phase 5 的 `c8cd75cf6`(B60 行窗)带进来的。 2. `ShopDataDuplicateCheckQueryService` 四处页窗统一走 `windowFrom/windowTo`:溢出发生在 **那次乘法本身**,只钳边界的话,下一个直接调本服务的调用方(或一支新用例)能把 500 原样请回来而没有任何尺子知道错在哪。判据 `ShopDataDuplicateCheckPageWindowTest`(5)。 其中 `page=Integer.MIN_VALUE` 那一格是先咬到修的人自己的:草稿只在 int 上做 `page - 1`, MIN_VALUE 会翻成"最后一窗之后"而不是第一窗,必须先把减数升到 long。 3. `TaskResultFileJobWorker`:`markInFlight` 从 try 外挪进 try 内(心跳 executor 在容器关闭时 `shutdownNow()`,`startJobHeartbeat` 抛 RejectedExecutionException 就跳过 finally,jobId 永久留在在飞登记里=卡死扫描永远跳过这一行,既不失败也不完成的静默停摆)。 4. `TaskResumeService` 的双读注释纠偏:它写的是**被反向验证打回的那份草稿口径** ("有列读列"),留着等于把下一个人引回去把代码改成真 bug。正本口径是 "文案判得出按文案、判不出才按列"(Phase 2 ✅ 行在册)。 尺子登记:`ShopDataDuplicateCheckPageWindowTest` 已入 ops `scripts/score_ratchet.py` 后端点名册 (不在册=闸门跑不到=等于没被执行,10-29 教训的形状)。 门禁实录:`mvn -o test-compile` rc=0;定向 `mvn -o test -Dtest=ShopDataDuplicateCheckPageWindowTest` Tests run 5 / Failures 0 / BUILD SUCCESS;后端面无 lint、无入库格式器 ⇒ 如实登记非漏执行。 归属:本笔只圈上述 5 支文件。`modules/usersecret/**` 六支+`V188__user_secret_usage_source.sql` +`MigratedSqlEquivalenceTest`/`migrated-statements.tsv`、以及 `useLiveTaskBoard.ts`+task-delete 三件,都是另一路会话的在飞件:未 add、未改、未评价。 --- ...AdminShopDataDuplicateCheckController.java | 12 +- .../ShopDataDuplicateCheckQueryService.java | 43 ++++++-- .../task/service/TaskResultFileJobWorker.java | 15 ++- .../task/service/TaskResumeService.java | 8 +- .../ShopDataDuplicateCheckPageWindowTest.java | 103 ++++++++++++++++++ 5 files changed, 167 insertions(+), 14 deletions(-) create mode 100644 backend-java/src/test/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckPageWindowTest.java diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/controller/AdminShopDataDuplicateCheckController.java b/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/controller/AdminShopDataDuplicateCheckController.java index b8c3e2c54..cc11934a0 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/controller/AdminShopDataDuplicateCheckController.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/controller/AdminShopDataDuplicateCheckController.java @@ -329,12 +329,22 @@ public class AdminShopDataDuplicateCheckController { + /** + * 页码上限(CR-18/铁律 15 同形,判据参照 {@code KeywordUsageAdminService.MAX_PAGE})。 + * 只钳下限不够:{@code ShopDataDuplicateCheckQueryService} 四处页窗算的都是 + * {@code (page - 1) * pageSize},{@code page=11000000 / page_size=200} 会溢出成负数, + * {@code Math.min(负数, total)} 仍是负数,喂给 {@code subList(负, 负)} 直接 + * IndexOutOfBounds = HTTP 500(有至少一行重复数据就能触发)。钳到 10000 页 × + * 本文件最大 page_size 200 = 200 万行,已远超实际可翻范围。 + */ + private static final int MAX_PAGE = 10_000; + private static int parsePage(String page) { if (page == null || page.isBlank()) { return 1; } try { - return Math.max(1, Integer.parseInt(page.trim())); + return clamp(Integer.parseInt(page.trim()), 1, MAX_PAGE); } catch (NumberFormatException ex) { throw new BusinessException(400, "page 参数不合法"); } diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckQueryService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckQueryService.java index 6b9698dfc..05bb76f6e 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckQueryService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckQueryService.java @@ -198,14 +198,41 @@ public class ShopDataDuplicateCheckQueryService { "total", 0, "page", page, "page_size", pageSize, "scanned_at", ""); } + /** + * 页窗下界(铁律 15 同形,判据参照 {@code KeywordUsageAdminService} 的 CR-18 注释)。 + * + *

四处页窗原先算的都是 {@code Math.min((page - 1) * pageSize, total)}:这个乘法先于 + * {@code min} 发生,{@code page=11000000 / page_size=200} 溢出成 {@code -2094967296}, + * {@code min} 之后仍是负数,喂给 {@code subList(负, 负)} 就是 IndexOutOfBounds = HTTP 500 + * ——只要表里有至少一行重复数据就能触发。2026-11-04 Phase 6 CR 在 {@code consoleData} + * 那处抓到(Phase 5 的 B60 行窗改的就是这一位),另外三处同形所以一起收进这里。 + * + *

先转 {@code long} 再乘:{@code page} 传得再离谱也只会拿到空窗,不会拿到负下标。 + * 控制器的 {@code parsePage} 那道钳制(上限 10000)是边界卫生,这一位是**算不坏**本身: + * 只钳边界的话,下一个直接调本服务的调用方(或一支新用例)能把 500 原样请回来, + * 而谁也看不出哪里错了。 + */ + private static int windowFrom(int page, int pageSize, int total) { + // `page - 1` 本身也要在 long 上算:page=Integer.MIN_VALUE 时在 int 上会翻成 + // +2147483647(="最后一窗之后"),升 long 才是"第一窗"。这一位是本支用例第 4 格 + // 当场抓出来的——判据不是装饰,它先咬到了修的人自己。 + long start = Math.max(0L, (long) page - 1) * Math.max(0, pageSize); + return (int) Math.min(start, total); + } + + /** 页窗上界:同样先升 long,`from` 已 ≤ total,加一档 pageSize 也不越 int。 */ + private static int windowTo(int from, int pageSize, int total) { + return (int) Math.min((long) from + Math.max(0, pageSize), total); + } + public Map itemsData(DuplicateScanView view, Set visibleKeys, int page, int pageSize, String viewMode, Filters filters) { boolean monitor = "monitor".equals(viewMode); List cropped = cropItems(view.payload().items(), visibleKeys); List matched = filterItems(cropped, monitor, filters); int total = matched.size(); - int from = Math.min((page - 1) * pageSize, total); - int to = Math.min(from + pageSize, total); + int from = windowFrom(page, pageSize, total); + int to = windowTo(from, pageSize, total); List slice = matched.isEmpty() ? List.of() : new ArrayList<>(matched.subList(from, to)); List shops = cropShops(view.payload().shops(), visibleKeys); @@ -219,8 +246,8 @@ public class ShopDataDuplicateCheckQueryService { List matched = filterItems(cropped, true, filters); matched.sort(detailItemComparator()); int total = matched.size(); - int from = Math.min((page - 1) * pageSize, total); - int to = Math.min(from + pageSize, total); + int from = windowFrom(page, pageSize, total); + int to = windowTo(from, pageSize, total); List> pageItems = new ArrayList<>(); for (int i = from; i < to; i++) { pageItems.add(detailItem(matched.get(i))); @@ -265,8 +292,8 @@ public class ShopDataDuplicateCheckQueryService { } rows.sort((a, b) -> ledgerCompare(a, b, key, asc)); int total = rows.size(); - int from = Math.min((page - 1) * pageSize, total); - int to = Math.min(from + pageSize, total); + int from = windowFrom(page, pageSize, total); + int to = windowTo(from, pageSize, total); List> slice = new ArrayList<>(); for (int i = from; i < to; i++) { slice.add(ledgerMap(rows.get(i))); @@ -310,8 +337,8 @@ public class ShopDataDuplicateCheckQueryService { rows.sort(Comparator.comparing((DuplicateItem item) -> safe(item.asin()))); List> groups = groupStats(shops, filterItems(cropped, false, filters)); int total = rows.size(); - int from = Math.min(Math.max(0, page - 1) * pageSize, total); - int to = Math.min(from + pageSize, total); + int from = windowFrom(page, pageSize, total); + int to = windowTo(from, pageSize, total); List window = rows.isEmpty() ? List.of() : new ArrayList<>(rows.subList(from, to)); return mapOf("pending", false, "scanned_at", safe(view.scannedAt()), "summary", summary, "shops", shops, "dup", window, "groups", groups, "total_dup", total, diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResultFileJobWorker.java b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResultFileJobWorker.java index 78ac742fa..cf286d2ba 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResultFileJobWorker.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResultFileJobWorker.java @@ -273,12 +273,21 @@ public class TaskResultFileJobWorker { // (stage=TERMINAL_STOP_RENEW),updated_at 从此无限期冻结,卡死扫描若只看时间戳就会 // 每 5 分钟给还在跑的执行线程记一次 stuck_requeue_count,20 次把还能出表的单判死 // (2026-11-03 现网 jobId=23815/taskId=33179 正是这个环)。注销必须在 finally。 - taskFileJobService.markInFlight(job.getId()); - ScheduledFuture heartbeat = startJobHeartbeat(job, parentTerminalObserved); + // (2026-11-04 CR-P2)登记与心跳都必须在这道 try **内**起:旧形状把 markInFlight + // 放在 try 外、注销放在 finally,而 `startJobHeartbeat` 在容器关闭时会抛 + // RejectedExecutionException(jobHeartbeatExecutor 已 shutdownNow)——那一跳直接绕过 + // finally,把这个 jobId 永久留在在飞登记里:卡死扫描从此永远跳过这一行, + // 既不失败也不完成=静默停摆,正是下面这段注释本身要防的形状。重启进程才清得掉, + // 而"重启即空"不能当兜底——那等于把这一单无限期挂在"看起来还在跑"上。 + ScheduledFuture heartbeat = null; try { + taskFileJobService.markInFlight(job.getId()); + heartbeat = startJobHeartbeat(job, parentTerminalObserved); processInternal(job); } finally { - heartbeat.cancel(false); + if (heartbeat != null) { + heartbeat.cancel(false); + } try { // CR-08:心跳只负责停续期;终态判定必须等执行线程退出后再做。 // 此时若 markSuccess 已落库,finishAfterExecution 会保留 SUCCESS; diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResumeService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResumeService.java index 81d9ba427..428cb1ce0 100644 --- a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResumeService.java +++ b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskResumeService.java @@ -227,8 +227,12 @@ public class TaskResumeService { continue; } TaskRetryEligibility.Category notRetryable = - // 双读(B51-R1):这行登记过分类就以列为准,没登记才回落到文案——自动续跑 - // 是每天真在跑的判死面,不该让"某句话改了措辞"在这种地方动摇历史行的判定。 + // 双读(B51-R1)定档是**文案判得出按文案、判不出才按列**,不是"有列读列" + // (2026-11-04 CR 纠正本注释:它写的正是被反向验证打回的那份草稿口径, + // 留着就是把下一个人引回去把代码改成真 bug)。实体写面 + // (setErrorMessage+updateById)改文案带不动列,无条件读列=让一列**陈旧** + // 的判决判死当前这一单,比它要修的 bug 更糟;背离时逐字等于加列前的今天。 + // 自动续跑是每天真在跑的判死面,更要按这条读。 TaskRetryEligibility.category(original); if (notRetryable != null) { stats.forceMajeureSkippedCount++; diff --git a/backend-java/src/test/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckPageWindowTest.java b/backend-java/src/test/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckPageWindowTest.java new file mode 100644 index 000000000..401efc8ab --- /dev/null +++ b/backend-java/src/test/java/com/nanri/aiimage/modules/shopduplicatecheck/service/ShopDataDuplicateCheckPageWindowTest.java @@ -0,0 +1,103 @@ +package com.nanri.aiimage.modules.shopduplicatecheck.service; + +import com.nanri.aiimage.modules.shopduplicatecheck.TestFixtures; +import com.nanri.aiimage.modules.shopduplicatecheck.mapper.ShopDuplicateCheckSourceMapper; +import org.junit.jupiter.api.BeforeEach; +import org.junit.jupiter.api.Test; + +import java.util.List; +import java.util.Map; + +import static org.assertj.core.api.Assertions.assertThat; +import static org.assertj.core.api.Assertions.assertThatCode; +import static org.mockito.Mockito.mock; + +/** + * 页窗算式的溢出档(铁律 15;2026-11-04 Phase 6 Code Review 的 CR-P1)。 + * + *

病灶是 {@code Math.min((page - 1) * pageSize, total)}:乘法先于 {@code min} 发生, + * {@code page=11000000 / page_size=200} 在 int 上溢出成 {@code -2094967296}, + * {@code Math.min} 之后还是负数,喂进 {@code subList(负, 负)} 就是 IndexOutOfBounds = HTTP 500。 + * 表里只要有至少一行重复数据就能打出来,且四个界面入口(撞款明细/行级详情/台账/console 矩阵) + * 全是同一个形状。这条不是"深分页慢"那种性能问题,是**正确性**问题:管理员翻到最后一页 + * 或者被脚本扫一遍,接口就红。 + * + *

为什么判据打在服务而不是控制器:控制器的 {@code parsePage} 也加了上限钳制(边界卫生), + * 但溢出发生在服务里的那次乘法上——只钳边界的话,下一个直接调本服务的调用方能把 500 原样请回来, + * 而没有任何一把尺子知道错在哪。所以这里打真入口,把「算不坏」钉住。 + */ +class ShopDataDuplicateCheckPageWindowTest { + + /** 审计给的现网形状:这一对相乘正好跨过 Integer.MAX_VALUE。 */ + private static final int OVERFLOW_PAGE = 11_000_000; + private static final int OVERFLOW_SIZE = 200; + + private ShopDataDuplicateCheckQueryService service; + private ShopDataDuplicateCheckScanService.DuplicateScanView view; + private ShopDataDuplicateCheckQueryService.Filters noFilter; + + @BeforeEach + void setUp() { + service = new ShopDataDuplicateCheckQueryService(mock(ShopDuplicateCheckSourceMapper.class)); + view = TestFixtures.view(); + noFilter = ShopDataDuplicateCheckQueryService.Filters.clean("", "", "", "", "", ""); + } + + @Test + void consoleMatrixWithOverflowShapedPageYieldsEmptyWindowNotHttp500() { + // 本 Phase 的 B60 行窗改的就是这一处(c8cd75cf),CR 也是在这里抓到的。 + int fullTotal = (int) service.consoleData(view, null, noFilter, 1, 100, "").get("total_dup"); + assertThat(fullTotal).isGreaterThan(0); + + Map data = service.consoleData( + view, null, noFilter, OVERFLOW_PAGE, OVERFLOW_SIZE, ""); + + assertThat((List) data.get("dup")).isEmpty(); + assertThat(data.get("total_dup")).isEqualTo(fullTotal); + assertThat(data.get("page")).isEqualTo(OVERFLOW_PAGE); + } + + @Test + void otherThreeWindowSitesSurviveIntegerMaxValuePage() { + for (int page : new int[]{Integer.MAX_VALUE, OVERFLOW_PAGE, 1_073_741_824}) { + assertThatCode(() -> { + assertThat((List) service.itemsData(view, null, page, 200, "", noFilter).get("items")) + .isEmpty(); + assertThat((List) service.detailData(view, null, page, 200, noFilter).get("items")) + .isEmpty(); + assertThat((List) service.ledgerData(view, null, page, 200, null, null, noFilter) + .get("items")).isEmpty(); + }).as("page=%s 不该把负下标喂给 subList", page).doesNotThrowAnyException(); + } + } + + @Test + void totalStaysTheUnpaginatedCountBeyondTheLastPage() { + // 「越界页=空窗」与「total 仍是全量」必须同时成立:把 total 跟着窗走,界面会显示"没有重复"。 + int fullTotal = (int) service.itemsData(view, null, 1, 1, "", noFilter).get("total"); + assertThat(service.itemsData(view, null, 999_999, 100, "", noFilter).get("total")) + .isEqualTo(fullTotal); + assertThat(service.consoleData(view, null, noFilter, 999_999, 100, "").get("total_dup")) + .isEqualTo((int) service.consoleData(view, null, noFilter, 1, 1, "").get("total_dup")); + } + + @Test + void pageBelowOneFallsBackToFirstWindowInsteadOfNegativeIndex() { + for (int page : new int[]{0, -1, Integer.MIN_VALUE}) { + List items = (List) service.itemsData(view, null, page, 2, "", noFilter).get("items"); + assertThat(items).as("page=%s 应当按第一窗读", page).isNotEmpty(); + assertThat((List) service.consoleData(view, null, noFilter, page, 100, "").get("dup")) + .as("page=%s 的 console 矩阵同理", page).isNotEmpty(); + } + } + + @Test + void middleWindowStillReturnsExactlyPageSizeRowsInOrder() { + // 反向自证:把溢出修成"整窗返回空"这种过头样子,这条会红。 + List first = (List) service.itemsData(view, null, 1, 1, "dup", noFilter).get("items"); + List second = (List) service.itemsData(view, null, 2, 1, "dup", noFilter).get("items"); + assertThat(first).hasSize(1); + assertThat(second).hasSizeLessThanOrEqualTo(1); + assertThat(first).isNotEqualTo(second); + } +}