refactor(ingress): 信箱边界层契约重构——发现与处理标记解耦、回填事实并入 PROC_STATE

收报扫描不再以 DATE_PROCESSED 为谓词:终态而未回填的行(解码失败死信等)会永久占据
有限批次,累积到 claim-batch 后收报整体停摆(US-01 条目 3 / message-lifecycle §5.3)。

ingress/InboxPoller.kt:按 ID 区间升序有界读取(ID > W),水位落 INBOX_CURSOR 并与入队
同一 PG 事务推进(中断后重扫补建);遇空洞即停,空洞超过 pipeline.max-commit-delay 判定
为永久并放行——否则水位永久停摆于一次自增回滚留下的空位。删除 InboxEnqueue(改由
insertIfAbsent 幂等入队)与 poller 内的回填扫描(消除 ingress→jobs 反向依赖)。

infra/persistence:端口按事实重画为 readRange/maxId/markProcessedIfUnmarked,标记 UPDATE
带 DATE_PROCESSED IS NULL 守卫,只把空标写为已处理(§11 单调,重复执行无副作用)。
回填事实并入 PROC_STATE(RECEIVED_AT/BACKFILL_AT/NEXT_AT/ATTEMPTS/ERROR),BACKFILL_TODO
随 V2 迁移下线;终态与回填意图是同一条 UPDATE,由处理器在自己的业务事务内落库,
message-lifecycle §4 登记的两个崩溃窗口(提交后回填前崩溃、待办二次落账失败)不再是缺口。

processing/BackfillService.kt(取代 BackfillSweepJob):终态提交后立即尝试一次,失败按
30s→15min 指数退避重试;扫描条件「终态 + 未确认标记 +(已到期 或 接收时间早于 NOW − R)」
使 §5.2 的超期期限 R 覆盖退避,中间态永不补写。死信同样可补写——回填只需消息 ID,
不再依赖 META。Pump 改用可注入 Clock。

infra/health:InboxLifecycleHealthIndicator 输出积压条数、最老未处理信龄、未回填终态数与
水位滞后(OPS-2 / §5.3 验收)。预计消化时长需吞吐采样,留待接入指标注册表时补。

配置:pipeline.max-commit-delay / overdue-backfill / backfill-batch、mailbox.processed-value
(Q2/Q6/Q7 未书面确认前取保守初值,不得为提速下调)。

不变量回归测试:死信不阻断后续发现、水位遇空洞即停与老化放行、终态+意图同事务、
超期 R 覆盖退避、中间态不补写、标记单调;InboxLifecycleJdbcSqlTest 以 H2 的 PostgreSQL
兼容模式直连验证上述 SQL 语义(不依赖 docker)。libs.h2 由 testRuntimeOnly 提为
testImplementation 以支持该用例。

验证:gradle clean test --offline → 78 tests / 0 failures / 1 skipped
(PG Testcontainers 集成用例在本机无 docker 时按既有约定 assumeTrue 跳过)。
This commit is contained in:
windyboy
2026-09-10 10:59:58 +08:00
parent ffd3abd655
commit 09b53c77bb
32 changed files with 1402 additions and 548 deletions
@@ -2,18 +2,34 @@ package com.gzzn.omms.msgexchange.ingress
import com.gzzn.omms.msgexchange.config.PipelineProps
import com.gzzn.omms.msgexchange.infra.persistence.CminmsgInboxRepository
import com.gzzn.omms.msgexchange.jobs.BackfillSweepJob
import com.gzzn.omms.msgexchange.infra.persistence.InboxCursorRepository
import com.gzzn.omms.msgexchange.infra.persistence.MailboxRow
import com.gzzn.omms.msgexchange.infra.persistence.PipelineTransactionManager
import com.gzzn.omms.msgexchange.infra.persistence.ProcStateRepository
import jakarta.inject.Singleton
import java.time.Duration
import java.time.Instant
/**
* ACMA-8 流程 1 · 主路径:JDBC 轮询共享 MySQL CMINMSGS`DATE_PROCESSED IS NULL`
* 发现新信后入队自有 PG。与 legacy `MsgExchangeRunner.getNewMsgsAfterId(0L)` 同语义。
* 收报(docs/message-lifecycle.md §5.1):共享 MySQL 信箱按 **ID 区间**升序有界读取
* 在自有 PG 建 `PROC_STATE(PENDING)`,并把消费水位推进到连续上界;入队与水位推进在同一
* PG 事务内提交,中断即重扫补建(§4 第一行)。收报层不解析业务载荷、不写处理标记。
*
* 三条红线:
* - **扫描谓词不含处理标记**:标记只用于回填与库方清除,不参与消息发现;否则已入队而未
* 回填的行会永久占据批次,死信累积到批大小时收报整体停摆;
* - **水位遇空洞即停**:不越过空洞入队(越过后较小 ID 迟到即 FIFO 越序,architecture §5);
* - **空洞老化**:超过最大提交时延(Q2 承诺)的空洞判定为永久并放行,否则水位会永久停摆于
* 一次自增回滚留下的空位。
*
* 空转代价为每轮一次区间 SELECT。
*/
@Singleton
class InboxPoller(
private val inbox: CminmsgInboxRepository,
private val enqueue: InboxEnqueue,
private val backfillSweep: BackfillSweepJob,
private val mailbox: CminmsgInboxRepository,
private val procState: ProcStateRepository,
private val cursor: InboxCursorRepository,
private val txManager: PipelineTransactionManager,
private val props: PipelineProps,
) {
private val log = org.slf4j.LoggerFactory.getLogger(InboxPoller::class.java)
@@ -21,31 +37,40 @@ class InboxPoller(
@Volatile
private var running = false
/** legacy 现役:afterId=0,每轮扫全部未处理行;PROC_STATE 判重防重复入队。 */
fun pollOnce(): Int {
/** @return 本轮新建的入队条数 */
fun pollOnce(now: Instant = Instant.now()): Int {
val batch = props.pipeline.claimBatch.coerceAtLeast(1)
val ids = inbox.pollUnprocessed(afterId = 0L, limit = batch)
var enqueued = 0
for (id in ids) {
if (enqueue.enqueue(id)) {
enqueued++
log.info("polled msgId={}", id)
}
}
sweepBackfillTodos()
return enqueued
}
val watermark = cursor.load()
val rows = mailbox.readRange(watermark.committedUpTo, batch)
if (rows.isEmpty()) return 0
/** v2 §5:每轮心跳顺带对账回填补偿待办(重启即恢复;空表只花一次索引 SELECT)。 */
private fun sweepBackfillTodos() {
try {
val outcome = backfillSweep.sweep()
if (outcome.inspected > 0) {
log.info("backfill sweep inspected={} succeeded={} failed={}", outcome.inspected, outcome.succeeded, outcome.failed)
// 连续上界;读取区间内出现空洞时,只推进到连续部分,空洞之后的行暂不入队(防较小 ID 迟到被越过)
val contiguous = contiguousUpTo(watermark.committedUpTo, rows) ?: watermark.committedUpTo
var committedTo = contiguous
var holeSince: Instant? = null
if (rows.last().msgId > contiguous) {
val since = watermark.holeSince ?: now
if (Duration.between(since, now) < props.pipeline.maxCommitDelay) {
holeSince = since
} else {
// 空洞老化:超过最大提交时延仍缺席即判永久(Q2),放行水位,否则永久停摆
committedTo = rows.first { it.msgId > contiguous }.msgId - 1
log.warn("hole after W={} aged out, watermark advanced to {}", contiguous, committedTo)
}
} catch (e: Exception) {
log.error("backfill sweep tick failed", e)
}
val enqueued = txManager.inTransaction {
var n = 0
rows.takeWhile { it.msgId <= committedTo }.forEach { row ->
if (procState.insertIfAbsent(row.msgId, row.receivedAt)) n++
}
cursor.save(InboxCursorRepository.Cursor(committedTo, holeSince))
n
}
if (enqueued > 0) {
log.info("polled {} new messages, W {} -> {}", enqueued, watermark.committedUpTo, committedTo)
}
return enqueued
}
fun loop() {
@@ -70,7 +95,19 @@ class InboxPoller(
running = false
}
private fun sleepQuietly(d: java.time.Duration) {
/** 连续上界:从 W+1 起 ID 逐 1 相邻的最后一个;首个空位即停(rows 为升序且覆盖该区间)。 */
private fun contiguousUpTo(from: Long, rows: List<MailboxRow>): Long? {
var expected = from + 1
var last: Long? = null
for (row in rows) {
if (row.msgId != expected) break
last = row.msgId
expected++
}
return last
}
private fun sleepQuietly(d: Duration) {
try {
Thread.sleep(d.toMillis().coerceAtLeast(1))
} catch (_: InterruptedException) {