From eb6f082674159daff3c6bde8053c92b4f5ea66a0 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E9=BB=84=E8=87=AA=E8=BE=BE?= <980324341@qq.com>
Date: Tue, 20 Oct 2026 12:30:09 +0800
Subject: [PATCH] =?UTF-8?q?feat(=E5=BF=83=E8=B7=B3):=20=E5=BD=B1=E5=AD=90?=
=?UTF-8?q?=E5=88=A4=E5=AE=98=E5=85=88=E8=B7=91=E8=AF=BB=E6=95=B0=E2=80=94?=
=?UTF-8?q?=E2=80=94=E5=BF=83=E8=B7=B3=E6=96=AD/=E8=BF=9B=E5=BA=A6?=
=?UTF-8?q?=E5=81=9C=E6=BB=9E/=E6=8E=92=E9=98=9F/=E6=9C=8D=E5=8A=A1?=
=?UTF-8?q?=E7=AB=AF=E5=A4=84=E7=90=86=E5=88=86=E5=BC=80=E5=88=A4?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit
现有判死线各看各的证据,假活(心跳线程活着、主线程卡死)与假死(长任务/排队被
判死)同时存在。新增只读影子扫描:每 5 分钟扫 RUNNING 文件任务,按「心跳中断 /
进度停滞 / 排队宽限 / 服务端结果任务 / 24h 硬顶」输出 would-finalize 日志,
不改任何任务状态。进度用「计数+文本+阶段」指纹,写影子独立快照键,不影响正式
判死线的未变化起点。规则纯函数单测 9 例 + 只读边界守卫 3 例。
---
.../service/TaskHeartbeatPositionService.java | 64 ++++-
.../task/service/TaskLivenessShadowRules.java | 132 ++++++++++
.../service/TaskLivenessShadowService.java | 236 ++++++++++++++++++
.../src/main/resources/application.yml | 12 +
.../TaskHeartbeatPositionServiceTest.java | 32 +++
.../service/TaskLivenessShadowGuardTest.java | 76 ++++++
.../service/TaskLivenessShadowRulesTest.java | 125 ++++++++++
7 files changed, 671 insertions(+), 6 deletions(-)
create mode 100644 backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskLivenessShadowRules.java
create mode 100644 backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskLivenessShadowService.java
create mode 100644 backend-java/src/test/java/com/nanri/aiimage/modules/task/service/TaskLivenessShadowGuardTest.java
create mode 100644 backend-java/src/test/java/com/nanri/aiimage/modules/task/service/TaskLivenessShadowRulesTest.java
diff --git a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatPositionService.java b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatPositionService.java
index 37ab015d8..cccf33c79 100644
--- a/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatPositionService.java
+++ b/backend-java/src/main/java/com/nanri/aiimage/modules/task/service/TaskHeartbeatPositionService.java
@@ -28,6 +28,8 @@ public class TaskHeartbeatPositionService {
private static final String KEY_PREFIX = "task:heartbeat:position:";
private static final String STALL_KEY_PREFIX = "task:progress:stall:";
+ /** 影子判官的独立快照键:影子的观测时刻不得改变正式判死线的「未变化起点」。 */
+ private static final String SHADOW_STALL_KEY_PREFIX = "task:progress:shadow-stall:";
private static final Duration POSITION_TTL = Duration.ofHours(24);
private static final int POSITION_LABEL_MAX_LEN = 60;
/** 计数字符串的最长位数:9 位以内 parseInt 必不溢出,所以这一处不需要 catch。 */
@@ -303,7 +305,57 @@ public class TaskHeartbeatPositionService {
if (position == null || position.isBlank()) {
return StallState.UNKNOWN;
}
- String stallKey = STALL_KEY_PREFIX + taskId;
+ return evaluateAgainstSnapshot(taskId, STALL_KEY_PREFIX + taskId, position, staleWindowMinutes);
+ }
+
+ /**
+ * 影子判官专用:把「计数 + 文本 + 阶段」压成指纹判断推进,快照写独立键。
+ *
+ *
与正式线只比文本不同:计数变了也是推进(长页面里文本可能长时间不变),
+ * 阶段变化同样算推进。影子每 5 分钟扫一次,若与正式线共用快照键,影子可能比正式线早几分钟
+ * 看到变化,从而把正式线的未变化起点提前、间接影响封口时刻;所以这里只写 shadow 键。
+ */
+ public StallState evaluateProgressForShadow(Long taskId, long staleWindowMinutes) {
+ if (taskId == null || taskId <= 0 || staleWindowMinutes <= 0) {
+ return StallState.UNKNOWN;
+ }
+ String fingerprint = readProgressFingerprint(taskId);
+ if (fingerprint == null) {
+ return StallState.UNKNOWN;
+ }
+ return evaluateAgainstSnapshot(taskId, SHADOW_STALL_KEY_PREFIX + taskId,
+ fingerprint, staleWindowMinutes);
+ }
+
+ /**
+ * 影子判官的推进指纹:位置哈希一次读出,计数/文本/阶段任一变化都算推进。
+ * 返回 null 表示这个任务没有任何位置信息(老客户端不传位置),不能作为判死证据。
+ */
+ private String readProgressFingerprint(Long taskId) {
+ Map