fix(batch-image): 修复审计发现的计费死锁、状态机与队列原子性缺陷

修复 PR #3768 批量图像 MVP 合并后审计报告中的全部问题:

结算与计费(高危):
- 所有 SETTLEMENT_* 失败(超冻结/计数非法/manifest 冲突/定价缺失/扣费失败)
  统一计入 retry_count 并在耗尽时释放冻结转 failed,消灭 settling 无限
  requeue 导致的冻结余额永久锁死
- 耗尽出口的释放指纹统一为 RequestHash,与 processor/Cancel/recovery 一致;
  release 遇同 request id 指纹冲突视为幂等成功,治愈历史毒消息
- 管理端校验 hold_multiplier >= discount_multiplier,定价快照对存量脏数据钳制
- 释放前校验 per-job hold claim(dedup+归档表),杜绝幻影释放

索引对账(高危):
- provider 输出与提交 custom_id 集对账:未知条目丢弃并记事件,
  漏项补 PROVIDER_RESULT_MISSING 失败行,保证 success+fail == item_count

提交与恢复(高危):
- 提交前转 uploading 并在 provider.Submit 期间心跳刷新 updated_at;
  恢复扫描改为原子复核(FailStaleUnsubmittedBatchImageJob),
  消灭慢提交被误杀退款而上游任务照常计费的孤儿场景
- 上游任务创建成功但本地状态推进失败时,尽力取消上游并清理输入
- recovery 释放失败时入队交由 worker releaseTerminalHold 兜底重试

队列与并发(中危):
- Enqueue(SetNX+LPush)与 Reserve(BRPop+ZAdd)均改为 Lua 原子脚本,
  消灭崩溃窗口导致 job 脱离队列、被 7 天 inflight 键锁死
- 锁冲突按 LockConflictDelay 重新入队(原直接丢弃需等 10 分钟 stale 恢复)
- 处理期间心跳:active zset 续期(ZAddXX 防幽灵成员)+ 锁 TTL 续期
- ReplaceBatchImageItemsForJob 增加 indexing 状态守卫,防掉队 worker 重写账目

存量回归(中危):
- image-only 定价条目(仅图片价无 token 价)恢复 token 计费 fail-closed,
  不再按 $0 计费;图片计费路径不受影响
- 鉴权余额门槛恢复 balance <= 0 语义,MinimumBalanceReserve 不再作硬 403

加固:
- ZIP max_items 钳制到管理员上限;Submit 补齐 Platform==Gemini 校验;
  gemini downloadUri 跟随前做 host 白名单校验
- 批量客户端改用共享 httpclient(拨号/TLS/响应头超时有界)
- 审计点名的忽略错误(MarkDownloaded/SettlementFailed/AppendEvent 等)改为记日志
This commit is contained in:
shaw committed 2026-07-07 18:53:56 +08:00
1 parent 44ab690a01
commit 80a229bce5
31 files changed
+1180 -90

No files matched your search

@@ -152,6 +152,45 @@ func (r *batchImageRepository) TransitionBatchImageJobStatus(ctx context.Context
return tx.Commit()
}
func (r *batchImageRepository) TouchBatchImageJobSubmitting(ctx context.Context, batchID string) error {
_, err := r.sql.ExecContext(ctx, `
UPDATE batch_image_jobs
SET updated_at = $2
WHERE batch_id = $1
AND status IN ('created', 'uploading')`, batchID, time.Now())
return err
}
func (r *batchImageRepository) FailStaleUnsubmittedBatchImageJob(ctx context.Context, batchID string, cutoff time.Time, code, message string) (bool, error) {
now := time.Now()
res, err := r.sql.ExecContext(ctx, `
UPDATE batch_image_jobs
SET status = 'failed',
last_error_code = $2,
last_error_message = $3,
finished_at = CASE WHEN finished_at IS NULL THEN $4 ELSE finished_at END,
updated_at = $4,
version = version + 1
WHERE batch_id = $1
AND status IN ('created', 'uploading')
AND provider_job_name IS NULL
AND updated_at <= $5`, batchID, code, message, now, cutoff)
if err != nil {
return false, err
}
affected, err := res.RowsAffected()
if err != nil {
return false, err
}
if affected == 0 {
return false, nil
}
return true, appendBatchImageEventWithSQL(ctx, r.sql, batchID, "billing_hold_recovery_failed_unsubmitted", map[string]any{
"batch_id": batchID,
"error_code": code,
})
}
func (r *batchImageRepository) UpdateBatchImageJobProviderOutputRef(ctx context.Context, batchID, providerOutputRef string) error {
res, err := r.sql.ExecContext(ctx, `
UPDATE batch_image_jobs
@@ -410,9 +449,15 @@ func (r *batchImageRepository) ReplaceBatchImageItemsForJob(ctx context.Context,
func (r *batchImageRepository) replaceBatchImageItemsForJobWithSQL(ctx context.Context, sqlq batchImageSQLExecutor, batchID string, items []service.CreateBatchImageItemParams, counts service.BatchImageCounts) error {
var id int64
if err := sqlq.QueryRowContext(ctx, `SELECT id FROM batch_image_jobs WHERE batch_id = $1 FOR UPDATE`, batchID).Scan(&id); err != nil {
var status string
if err := sqlq.QueryRowContext(ctx, `SELECT id, status FROM batch_image_jobs WHERE batch_id = $1 FOR UPDATE`, batchID).Scan(&id, &status); err != nil {
return translatePersistenceError(err, service.ErrBatchImageJobNotFound, nil)
}
// 仅允许 indexing 状态重建 item 表:防止锁过期后掉队的 worker
// 重写已完成/已结算 job 的条目,造成账目与结果漂移。
if status != service.BatchImageJobStatusIndexing {
return service.ErrBatchImageIndexStateConflict
}
promptPreviews, err := r.batchImageItemPromptPreviews(ctx, sqlq, batchID)
if err != nil {
return err