Appearance
消费者崩溃与重投（consumer-crash）
本页结论：消费者在「业务已提交、ACK 未发出」之间崩溃，RabbitMQ 会重投同一条消息；幂等表让这次重投只产生一次
duplicate_skipped，而不会重复落库——这正是 at-least-once 下业务必须预期重复的含义。
适用场景
- 理解 at-least-once 投递的具体作用范围：Broker 保证「至少投一次」，不保证「只投一次」。
- 验证规格 §5.4 的幂等消费基准实现：DB 提交后才 ACK，崩溃窗口由幂等表兜底。
- 这是 Phase 1 的退出条件实验：从零启动 RabbitMQ，复现「崩溃 → 重投 → 幂等拦截」。
崩溃窗口在哪里
手动 ACK 的消费流程存在一个无法消除的窗口：
如果第 2 步之后、第 3 步之前进程终止，Broker 认为消息从未被确认，必然重投。反过来，若先 ACK 后写库，则存在「ACK 已发、业务未提交」的丢消息窗口。顺序只能是：业务事务提交 → ACK，重复交给幂等表处理。
实验步骤
bash
npm run lab -- rabbitmq consumer-crash编排分两轮：
- 第一轮：Producer 发送 3 条消息；Consumer 处理完第 1 条（
business_committed）后，在 ACK 前执行崩溃注入（Runtime.halt(137)），进程退出码 137。lab 断言崩溃确实发生——失败注入必须被观察到，而不是「恰好没触发」。 - 第二轮：lab 重启 Consumer。第 1 条消息带
redelivered=true再次到达，幂等表命中，status=duplicate_skipped并安全 ACK；随后正常处理第 2、3 条。
两轮共用同一个 SQLite 幂等库（.lab/ 目录），模拟同一业务实例重启。
断言
| 断言 | 期望 | 说明 |
|---|---|---|
| crashExitCode | 137 | 崩溃注入确实发生 |
| receivedTotal | 4 | 3 条唯一消息 + 1 条重投 |
| redeliveredCount | 1 | 仅 order-1001 被重投 |
| duplicatesObserved | 1 | 幂等表观察到 1 次重复 |
| duplicatesApplied | 0 | 重复未产生任何业务写入 |
| business_rows | 3 | orders 表恰好 3 行 |
提交快照
verifiedrabbitmq / consumer-crash
| 镜像 | rabbitmq:4.1.4-management@sha256:294b01e1796a8acede4619f32a1c394fae1f8021e57986ea01aad38dc2a4f502 |
| 捕获时间 | 2026-08-19T07:55:38.494Z |
| 耗时 / 退出码 | 14776 ms / exit 0 |
断言
| crashExitCode | 137 |
| crashAfterBusinessCommit | 1 |
| receivedTotal | 4 |
| redeliveredCount | 1 |
| duplicatesObserved | 1 |
| duplicatesApplied | 0 |
| uniqueMessageIds | 3 |
| business_rows | 3 |
| queueDepthAfter | 0 |
| consumerExitCode | 0 |
归一化日志
[producer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash routingKey=orders.crash durationMs=<ms> status=confirmed [producer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-2 eventType=order.created schemaVersion=1 aggregateId=order-1002 traceId=trace-2 correlationId=order-1002 destination=orders.crash routingKey=orders.crash durationMs=<ms> status=confirmed [producer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-3 eventType=order.created schemaVersion=1 aggregateId=order-1003 traceId=trace-3 correlationId=order-1003 destination=orders.crash routingKey=orders.crash durationMs=<ms> status=confirmed [producer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash destination=orders.crash confirmed=3 status=done [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash consumer=consumer-1 attempt=1 redelivered=false status=received [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash attempt=1 status=business_committed [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash status=crash_injected [assert] crashExitCode=137 PASS [assert] crashAfterBusinessCommit=1 PASS [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash consumer=consumer-1 attempt=1 redelivered=true status=received [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-1 eventType=order.created schemaVersion=1 aggregateId=order-1001 traceId=trace-1 correlationId=order-1001 destination=orders.crash attempt=1 status=duplicate_skipped [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-2 eventType=order.created schemaVersion=1 aggregateId=order-1002 traceId=trace-2 correlationId=order-1002 destination=orders.crash consumer=consumer-1 attempt=1 redelivered=false status=received [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-2 eventType=order.created schemaVersion=1 aggregateId=order-1002 traceId=trace-2 correlationId=order-1002 destination=orders.crash attempt=1 status=business_committed [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-3 eventType=order.created schemaVersion=1 aggregateId=order-1003 traceId=trace-3 correlationId=order-1003 destination=orders.crash consumer=consumer-1 attempt=1 redelivered=false status=received [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash messageId=mid-3 eventType=order.created schemaVersion=1 aggregateId=order-1003 traceId=trace-3 correlationId=order-1003 destination=orders.crash attempt=1 status=business_committed [consumer] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash queue=orders.crash received=3 status=done [inspect] timestamp=<ts> level=INFO service=order-service product=rabbitmq lab=consumer-crash business_rows=3 processed_rows=3 status=snapshot [assert] receivedTotal=4 PASS [assert] redeliveredCount=1 PASS [assert] duplicatesObserved=1 PASS [assert] duplicatesApplied=0 PASS [assert] uniqueMessageIds=3 PASS [assert] business_rows=3 PASS [assert] queueDepthAfter=0 PASS [assert] consumerExitCode=0 PASS
npm run lab -- rabbitmq consumer-crash快照中关键三行：
text
[consumer] ... messageId=mid-1 ... status=business_committed # 第一轮：业务已提交
[consumer] ... messageId=mid-1 ... status=crash_injected # ACK 前崩溃（exit 137）
[consumer] ... messageId=mid-1 ... redelivered=true ... status=duplicate_skipped # 第二轮：幂等拦截常见误区
- 「ACK 之后才落库」——顺序反了，崩溃窗口变成丢消息窗口。
- 「重投是小概率，可以不管」——at-least-once 意味着业务必须预期重复，而不是「偶尔可能重复」。
- 「幂等表只是优化」——它是崩溃窗口的正确性组件，不可省略。
- 「
redelivered=true就跳过」——重投消息可能第一次就没处理完，必须走完整幂等流程。
官方资料与版本说明
- RabbitMQ 4.1.4，
amqp-client5.34.0。 - Acknowledgement Modes 与 redelivery：https://www.rabbitmq.com/docs/confirms#acknowledgement-modes（checkedAt: 2026-08-19）