fix(任务): 卡死自愈扫描不再被一行的锁等待作废整轮

线上 12:58–13:13 每 110 秒一次 1205:扫描按 updated_at 升序取到的第一行正被另一台
持着行锁,一抛就把本轮已复位成功的 job 的 dispatch 事件一起带走(它们已是 PENDING
却没人派发),该判终态失败的作业也不 finalize。改成循环体逐行 try/catch、派发也逐个包,
只跳过这一行。刻意不加事务:每行本就带 status+updated_at 的幂等 CAS 条件。
This commit is contained in:
2026-09-28 14:11:26 +08:00
parent d70c472cab
commit 3a416f3935
2 changed files with 104 additions and 44 deletions
@@ -342,65 +342,83 @@ public class TaskFileJobService {
Map<Long, TaskFileJobEntity> originalJobsById = jobs.stream()
.collect(Collectors.toMap(TaskFileJobEntity::getId, job -> job, (a, b) -> a));
for (TaskFileJobEntity job : jobs) {
if ("FAILED".equals(job.getStatus())
&& job.getRetryCount() != null && job.getRetryCount() >= MAX_RETRY_COUNT) {
exhaustedJobIds.add(job.getId());
originalExhaustedById.put(job.getId(), job);
continue;
}
if ("PENDING".equals(job.getStatus())) {
// PENDING 僵尸兜底:补发 dispatch 事件(worker claim 幂等),
// 并累加 retryCount 防无限重发;注意不翻转状态,
// 避免与真实 PENDING 竞争 worker 的 claim 条件更新冲突。
try {
if ("FAILED".equals(job.getStatus())
&& job.getRetryCount() != null && job.getRetryCount() >= MAX_RETRY_COUNT) {
exhaustedJobIds.add(job.getId());
originalExhaustedById.put(job.getId(), job);
continue;
}
if ("PENDING".equals(job.getStatus())) {
// PENDING 僵尸兜底:补发 dispatch 事件(worker claim 幂等),
// 并累加 retryCount 防无限重发;注意不翻转状态,
// 避免与真实 PENDING 竞争 worker 的 claim 条件更新冲突。
int retryCount = job.getRetryCount() == null ? 0 : job.getRetryCount();
int nextRetryCount = Math.min(MAX_RETRY_COUNT - 1, retryCount + 1);
int updated = taskFileJobMapper.update(null, new LambdaUpdateWrapper<TaskFileJobEntity>()
.eq(TaskFileJobEntity::getId, job.getId())
.eq(TaskFileJobEntity::getStatus, "PENDING")
.eq(job.getUpdatedAt() != null, TaskFileJobEntity::getUpdatedAt, job.getUpdatedAt())
.isNull(job.getUpdatedAt() == null, TaskFileJobEntity::getUpdatedAt)
.set(TaskFileJobEntity::getRetryCount, nextRetryCount)
.set(TaskFileJobEntity::getUpdatedAt, LocalDateTime.now()));
if (updated > 0) {
eventJobIds.add(job.getId());
}
continue;
}
int retryCount = job.getRetryCount() == null ? 0 : job.getRetryCount();
int nextRetryCount = Math.min(MAX_RETRY_COUNT - 1, retryCount + 1);
int updated = taskFileJobMapper.update(null, new LambdaUpdateWrapper<TaskFileJobEntity>()
int nextRetryCount = Math.min(MAX_RETRY_COUNT, retryCount + 1);
LocalDateTime now = LocalDateTime.now();
LambdaUpdateWrapper<TaskFileJobEntity> update = new LambdaUpdateWrapper<TaskFileJobEntity>()
.eq(TaskFileJobEntity::getId, job.getId())
.eq(TaskFileJobEntity::getStatus, "PENDING")
.eq(TaskFileJobEntity::getStatus, "RUNNING")
.eq(job.getUpdatedAt() != null, TaskFileJobEntity::getUpdatedAt, job.getUpdatedAt())
.isNull(job.getUpdatedAt() == null, TaskFileJobEntity::getUpdatedAt)
.set(TaskFileJobEntity::getRetryCount, nextRetryCount)
.set(TaskFileJobEntity::getUpdatedAt, LocalDateTime.now()));
if (updated > 0) {
eventJobIds.add(job.getId());
.set(TaskFileJobEntity::getUpdatedAt, now);
if (nextRetryCount >= MAX_RETRY_COUNT) {
int updated = taskFileJobMapper.update(null, update
.set(TaskFileJobEntity::getStatus, "FAILED")
.set(TaskFileJobEntity::getErrorMessage, "文件生成任务运行超时,已达到最大重试次数")
.set(TaskFileJobEntity::getFinishedAt, now)
.set(TaskFileJobEntity::getTerminalCallbackAt, null));
if (updated > 0) {
exhaustedJobIds.add(job.getId());
}
continue;
}
continue;
}
int retryCount = job.getRetryCount() == null ? 0 : job.getRetryCount();
int nextRetryCount = Math.min(MAX_RETRY_COUNT, retryCount + 1);
LocalDateTime now = LocalDateTime.now();
LambdaUpdateWrapper<TaskFileJobEntity> update = new LambdaUpdateWrapper<TaskFileJobEntity>()
.eq(TaskFileJobEntity::getId, job.getId())
.eq(TaskFileJobEntity::getStatus, "RUNNING")
.eq(job.getUpdatedAt() != null, TaskFileJobEntity::getUpdatedAt, job.getUpdatedAt())
.isNull(job.getUpdatedAt() == null, TaskFileJobEntity::getUpdatedAt)
.set(TaskFileJobEntity::getRetryCount, nextRetryCount)
.set(TaskFileJobEntity::getUpdatedAt, now);
if (nextRetryCount >= MAX_RETRY_COUNT) {
int updated = taskFileJobMapper.update(null, update
.set(TaskFileJobEntity::getStatus, "FAILED")
.set(TaskFileJobEntity::getErrorMessage, "文件生成任务运行超时,已达到最大重试次数")
.set(TaskFileJobEntity::getFinishedAt, now)
.set(TaskFileJobEntity::getStatus, "PENDING")
.set(TaskFileJobEntity::getErrorMessage, "文件生成任务运行超时,已重新排队")
.set(TaskFileJobEntity::getFinishedAt, null)
.set(TaskFileJobEntity::getTerminalCallbackAt, null));
if (updated > 0) {
exhaustedJobIds.add(job.getId());
reset++;
eventJobIds.add(job.getId());
}
continue;
}
int updated = taskFileJobMapper.update(null, update
.set(TaskFileJobEntity::getStatus, "PENDING")
.set(TaskFileJobEntity::getErrorMessage, "文件生成任务运行超时,已重新排队")
.set(TaskFileJobEntity::getFinishedAt, null)
.set(TaskFileJobEntity::getTerminalCallbackAt, null));
if (updated > 0) {
reset++;
eventJobIds.add(job.getId());
} catch (RuntimeException ex) {
// 这一行为什么必须单独吞:它正被别的事务占着时会等满 innodb_lock_wait_timeout
// 再抛 1205。整轮向外抛的代价不是这一行没修好,而是同批其余 199 行这轮全不看、
// 本轮已复位成功的 job 也发不出 dispatch(它们在 400 行之后才发布),
// 于是被锁的行每轮都排在最前(ORDER BY updated_at ASC)——线上实测每 110 秒
// 重复一次「等 50 秒 + 整轮作废」,卡住的任务在这段时间里根本没人管。
// jobId 必须打出来:它是下一轮定位持锁方的唯一线索。
log.warn("[task-file-job] stuck reset 单行失败已跳过,继续处理其余行 jobId={} scanned={} taskId={} moduleType={} status={} retryCount={} err={}",
job.getId(), jobs.size(), job.getTaskId(), job.getModuleType(), job.getStatus(),
job.getRetryCount(), ex.getMessage());
}
}
if (!eventJobIds.isEmpty()) {
Map<Long, TaskFileJobEntity> refreshedById = refreshJobsByIds(eventJobIds);
for (Long jobId : eventJobIds) {
publishDispatchEvent(refreshedById.get(jobId));
try {
publishDispatchEvent(refreshedById.get(jobId));
} catch (RuntimeException ex) {
// 派发监听器是同步执行的,一个 job 派发失败不能带走同批其余 job
log.warn("[task-file-job] stuck reset 派发事件发布失败,不影响其余 job jobId={} err={}",
jobId, ex.getMessage());
}
}
}
if (!exhaustedJobIds.isEmpty()) {