fix(交付): 已交付任务不再在 Python 超时恢复里空转(现网 33228 既不失败也不完成)
机制:恢复链路每 2 分钟走到 enqueueAssembleResult,它按 TaskFileJobService 的 .ne SUCCESS 刻意原样返回那行 SUCCESS(没人要求重做时不许替换用户可能已下载的表), 调用方据此打一条 "skipped llm submission because result job unavailable" 就跳过发批。 于是作业没排上、批没发出,交付闸门永远不被评估,补检/recheckMaxRounds 预算/终态 那条有界链一步都不走;而紧随其后的 touchJavaSideTaskActivity 又刷新 updated_at, 让陈旧扫描永远看不见这一行——正是铁律 28(c) 点名的「不计任何计数还刷 updated_at」。 33228 实测:表 11-03 10:50 就出好了(job SUCCESS+biz_file_result.success=1 带直链), activeAssembleJobs 恒 0,10 行永远"等待补检轮",而界面「重试」的判据(前后端一致的 FAILED/CANCELLED)永远拒绝它——用户被关在死胡同里。 三处改动: - TaskFileJobService 新增 hasDeliveredAssembleJob,与既有 hasExhaustedAssembleJob 成对 (那一格看"重试耗尽",这一格看"成品已交付",两种"恢复已经试过了"都该停止空转); - finalizer 在 uploadComplete ∧ activeAssembleJobs=0 ∧ 已交付 时返回 false 交回收尾, 不再 touch、不再空发批。收尾走既有 finalizeTask:它本就把部分数据缺口判成非致命 (partialGapSalvageable),缺口由行级文案标在表里(未采集到数据/未完成检测/ 未取得检测结果是三句不同的话,处置动作不同),是否成功仍由 statusResolver 按覆盖率判; - onTaskResumed 补第二半 rearmDeliveredAssembleJobForResume。只加上一条会引入新 bug: 刚点过「继续」的任务同样呈"已交付+无在飞作业",会被直接收尾,剩余行永远轮不到补检。 界面确认框本就承诺「当前已生成的部分结果文件会被替换」,这里复位那行 SUCCESS 是兑现 承诺,不是绕过 retry() 的保护——那是给没人要求重做时的。 判据 AppearancePatentStaleTaskFinalizerDeliveredGuardTest 4 支,四个方向都钉住:已交付→ 收尾且不 touch/未交付→照常恢复(保「继续」)/有在飞作业→不抢它正在产出的表/ 未封口→仍走封口那条路。后端整量 5,317 例 0 失败(新增前 5,313)。
This commit is contained in:
+124
@@ -0,0 +1,124 @@
|
||||
package com.nanri.aiimage.modules.appearancepatent.service.support;
|
||||
|
||||
import com.nanri.aiimage.modules.appearancepatent.service.support.AppearancePatentChunkSubmissionSupport.SubmitContext;
|
||||
import com.nanri.aiimage.modules.task.mapper.FileTaskMapper;
|
||||
import com.nanri.aiimage.modules.task.model.entity.FileTaskEntity;
|
||||
import com.nanri.aiimage.modules.task.model.entity.TaskChunkEntity;
|
||||
import com.nanri.aiimage.modules.task.service.TaskFileJobService;
|
||||
import org.junit.jupiter.api.BeforeEach;
|
||||
import org.junit.jupiter.api.Test;
|
||||
import org.junit.jupiter.api.extension.ExtendWith;
|
||||
import org.mockito.Mock;
|
||||
import org.mockito.junit.jupiter.MockitoExtension;
|
||||
|
||||
import java.util.List;
|
||||
import java.util.concurrent.atomic.AtomicBoolean;
|
||||
import java.util.concurrent.atomic.AtomicInteger;
|
||||
import java.util.function.BiConsumer;
|
||||
import java.util.function.Consumer;
|
||||
import java.util.function.Function;
|
||||
import java.util.function.Predicate;
|
||||
|
||||
import static org.junit.jupiter.api.Assertions.assertFalse;
|
||||
import static org.junit.jupiter.api.Assertions.assertTrue;
|
||||
import static org.mockito.Mockito.lenient;
|
||||
import static org.mockito.Mockito.when;
|
||||
|
||||
/**
|
||||
* 已交付任务不许在 Python 超时恢复里空转(2026-11-04 现网任务 33228)。
|
||||
*
|
||||
* <p>形状:表早在 11-03 10:50 就出好了({@code biz_task_file_job} ASSEMBLE_RESULT=SUCCESS、
|
||||
* {@code biz_file_result.success=1} 带直链),任务却仍 RUNNING、每 2 分钟被探针刷一次
|
||||
* {@code updated_at}。成因是恢复链路里两处"各自以为对方在管":
|
||||
* {@code enqueueAssembleResult} 原样返回那行 SUCCESS({@code retry} 的 {@code .ne SUCCESS} 是刻意的),
|
||||
* 调用方据此打一条 "skipped llm submission because result job unavailable" 就**跳过发批**;
|
||||
* 于是作业没排上、批没发出,交付闸门永远不被评估,补检/预算/终态那条有界链一步都不走,
|
||||
* 而紧随其后的 touch 又让陈旧扫描永远看不见这一行——既不失败也不完成。
|
||||
*
|
||||
* <p>反向同样要钉死:**刚点过「继续」的任务不能被这一格当成已交付直接收尾**,否则剩余行
|
||||
* 永远轮不到补检。这条靠 resume 复位那行 SUCCESS({@code rearmDeliveredAssembleJobForResume})
|
||||
* 保证,所以这里断言的是"没有已交付作业时必须照常恢复"。
|
||||
*/
|
||||
@ExtendWith(MockitoExtension.class)
|
||||
class AppearancePatentStaleTaskFinalizerDeliveredGuardTest {
|
||||
|
||||
private static final String MODULE_TYPE = "APPEARANCE_PATENT";
|
||||
private static final Long TASK_ID = 33228L;
|
||||
|
||||
@Mock
|
||||
private FileTaskMapper fileTaskMapper;
|
||||
@Mock
|
||||
private TaskFileJobService taskFileJobService;
|
||||
|
||||
private final AtomicBoolean touched = new AtomicBoolean(false);
|
||||
private final AtomicInteger llmSubmissions = new AtomicInteger(0);
|
||||
private final AtomicBoolean finalized = new AtomicBoolean(false);
|
||||
|
||||
private FileTaskEntity runningTask;
|
||||
|
||||
@BeforeEach
|
||||
void setUp() {
|
||||
runningTask = new FileTaskEntity();
|
||||
runningTask.setId(TASK_ID);
|
||||
runningTask.setModuleType(MODULE_TYPE);
|
||||
runningTask.setStatus("RUNNING");
|
||||
}
|
||||
|
||||
private void runFinalizer(boolean uploadComplete,
|
||||
long activeAssembleJobs,
|
||||
boolean delivered) {
|
||||
when(fileTaskMapper.selectById(TASK_ID)).thenReturn(runningTask);
|
||||
when(taskFileJobService.countActiveAssembleJobs(TASK_ID, MODULE_TYPE)).thenReturn(activeAssembleJobs);
|
||||
when(taskFileJobService.hasExhaustedAssembleJob(TASK_ID, MODULE_TYPE)).thenReturn(false);
|
||||
// 这一格是短路求值(uploadComplete=false 或有在飞作业时根本不会问),严格 stub 会判多余
|
||||
lenient().when(taskFileJobService.hasDeliveredAssembleJob(TASK_ID, MODULE_TYPE)).thenReturn(delivered);
|
||||
|
||||
Predicate<Long> submissionComplete = taskId -> uploadComplete;
|
||||
Predicate<Long> persistedRowsPresent = taskId -> true;
|
||||
Function<Long, Integer> markSubmissionComplete = taskId -> 1;
|
||||
Function<Long, List<TaskChunkEntity>> chunksLoader = taskId -> List.of();
|
||||
Function<List<TaskChunkEntity>, Integer> uninspectedRowCounter = chunks -> 10;
|
||||
Consumer<SubmitContext> llmSubmitter = context -> llmSubmissions.incrementAndGet();
|
||||
Consumer<Long> toucher = taskId -> touched.set(true);
|
||||
BiConsumer<FileTaskEntity, String> finalizer = (task, error) -> finalized.set(true);
|
||||
|
||||
new AppearancePatentStaleTaskFinalizer(MODULE_TYPE, fileTaskMapper, taskFileJobService,
|
||||
task -> 154, submissionComplete, persistedRowsPresent, markSubmissionComplete,
|
||||
chunksLoader, uninspectedRowCounter, llmSubmitter, toucher, finalizer)
|
||||
.finalizeStaleTask(TASK_ID, "有 10 行未完成检测,等待补检轮");
|
||||
}
|
||||
|
||||
@Test
|
||||
void 已交付且无在飞作业时按已有行账收尾不再刷新活动() {
|
||||
runFinalizer(true, 0L, true);
|
||||
|
||||
assertTrue(finalized.get(), "已交付的任务必须交回 finalizeStaleTask 收尾,而不是继续空转");
|
||||
assertFalse(touched.get(), "收尾这一格不许再刷 updated_at——刷了就永远不被陈旧扫描看见");
|
||||
assertFalse(llmSubmissions.get() > 0, "已交付时不该再发一批 LLM(发了也只会拿回同一行 SUCCESS)");
|
||||
}
|
||||
|
||||
@Test
|
||||
void 没有已交付作业时必须照常恢复否则继续功能被这一格掐死() {
|
||||
runFinalizer(true, 0L, false);
|
||||
|
||||
assertFalse(finalized.get(), "未交付的任务不能被误判成已交付而直接收尾");
|
||||
assertTrue(touched.get(), "正常恢复路径仍要刷新活动");
|
||||
assertTrue(llmSubmissions.get() > 0, "正常恢复路径仍要发出 LLM 批次");
|
||||
}
|
||||
|
||||
@Test
|
||||
void 仍有在飞出表作业时不触发已交付收尾() {
|
||||
runFinalizer(true, 1L, true);
|
||||
|
||||
assertFalse(finalized.get(), "出表作业还在飞时谈不上已交付,收尾会抢掉它正在产出的表");
|
||||
assertTrue(llmSubmissions.get() > 0, "有在飞作业时恢复照常推进");
|
||||
}
|
||||
|
||||
@Test
|
||||
void 上传未封口时仍走封口那条路() {
|
||||
runFinalizer(false, 0L, true);
|
||||
|
||||
assertFalse(finalized.get(), "uploadComplete=false 是封口场景,与本格判据无关");
|
||||
assertTrue(touched.get(), "封口路径照常刷新活动");
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user