feat(acm2): 闭合 G-KAFKA-D3 / G-IGNORE / G-EVENT-RETENTION 三个缺口

- G-KAFKA-D3:max-in-flight 默认收敛到 1,KafkaD3Check 启动自检钉住三项联合满足 D3
- G-IGNORE:IgnoreRules 在身份绑定后精确匹配 TYPE 字段,命中写 SKIPPED + 回填意图
- G-EVENT-RETENTION:V10 迁移加 SENT_AT 列,投递原子写,EventCleanupJob 有界清理过期 SENT 行
This commit is contained in:
windyboy
2026-09-13 15:48:16 +08:00
parent d37a9eb38e
commit fc744d0cfd
23 changed files with 540 additions and 35 deletions
@@ -0,0 +1,39 @@
package com.gzzn.omms.msgexchange.config
import io.micronaut.context.annotation.Context
import io.micronaut.context.annotation.Value
import jakarta.inject.Singleton
/**
* D3 启动自检:Kafka 生产者的 acks / enable-idempotence / max-in-flight 三项
* 必须联合满足幂等生产前提。不满足时拒绝启动——静默放行会让幂等保证在运行时失效。
*
* 独立于 `PipelineLifecycle`(受 autostart 开关控制),无条件执行。
*/
@Singleton
@Context
class KafkaD3Check(
@Value("\${kafka.producers.default.acks}") acks: String,
@Value("\${kafka.producers.default.enable-idempotence}") idempotence: String,
@Value("\${kafka.producers.default.max-in-flight-requests-per-connection}") maxInFlight: String,
) {
private val acks: String = acks
private val idempotence: Boolean = idempotence.toBoolean()
private val maxInFlight: Int = maxInFlight.toInt()
fun validate() {
require(acks == "all" || acks == "-1") {
"D3 violation: kafka.producers.default.acks must be 'all', got '$acks'"
}
require(idempotence) {
"D3 violation: kafka.producers.default.enable-idempotence must be true, got $idempotence"
}
require(maxInFlight == 1) {
"D3 violation: kafka.producers.default.max-in-flight-requests-per-connection must be 1, got $maxInFlight"
}
}
init {
validate()
}
}
@@ -82,6 +82,13 @@ class PipelineProps {
*/
var autostart: Boolean = false
/**
* MSG_EVENT 已确认投递行(`STATE='SENT'`)的保留期。
* 清理作业删除 `SENT_AT < now - eventRetention` 的行;
* 取值应大于最长可能的投递重试周期,避免误删刚发出的事件。
*/
var eventRetention: Duration = Duration.ofDays(7)
/** N28attempt ≤ 0(如 FAILED 未递增 attempts 的行)不得抛异常,取下界=首档退避。 */
fun backoffFor(attempt: Int): Long {
val index = (attempt - 1).coerceAtLeast(0)
@@ -130,7 +130,7 @@ class Dispatcher(
// 未知目标不能在"未发送"的情况下被标记为已发;记账后停止本轮。
else -> error("unknown delivery target: $target")
}
msgEvents.markSent(eventId)
msgEvents.markSent(eventId, scheduler.now())
true
} catch (ex: Exception) {
retryOrDead(e, ex.message ?: ex.javaClass.simpleName)
@@ -158,7 +158,7 @@ class Dispatcher(
EventType.UPSERT -> port.sendKafkaSchd("schd", e.partitionKey, e.payloadJson)
}
// 条件确认:读取时刻的代次(EVENT_ID + STATE_VERSION)被新写入覆盖时不标记,留待下一轮重发。
e.eventId?.let { msgEvents.markSentIfVersion(it, e.stateVersion) }
e.eventId?.let { msgEvents.markSentIfVersion(it, e.stateVersion, scheduler.now()) }
} catch (ex: Exception) {
failures.add(e)
}
@@ -29,4 +29,6 @@ data class MsgEvent(
val errorClass: ErrorClass? = null,
val lastError: String? = null,
val createdAt: java.time.Instant,
/** 投递确认时刻(`STATE='SENT'` 时写入);清理作业按此字段判断保留期。 */
val sentAt: java.time.Instant? = null,
)
@@ -15,6 +15,7 @@ import java.util.concurrent.atomic.AtomicLong
class PipelineCounters {
private val srvtSeen = AtomicLong(0)
private val vipfSeen = AtomicLong(0)
private val ignored = AtomicLong(0)
/**
* 入站记录里出现 `SRVT`/`VIPF` 段的条数(`[G-SRVT-VIPF]`)。
@@ -30,4 +31,9 @@ class PipelineCounters {
fun srvtSeenCount(): Long = srvtSeen.get()
fun vipfSeenCount(): Long = vipfSeen.get()
/** US-04 忽略清单命中计数。 */
fun ignoredAdd(count: Int = 1) { ignored.addAndGet(count.toLong()) }
fun ignoredCount(): Long = ignored.get()
}
@@ -28,6 +28,7 @@ import java.time.Duration
* - `msgx.pipeline.job.last_sweep_selected`:上一轮回填扫描选中的待办条数(扫描积压)
* - `msgx.pipeline.codec.srvt_seen.total` / `msgx.pipeline.codec.vipf_seen.total`:入站记录里出现
* `SRVT`/`VIPF` 段的条数(尚未落明细表,`[G-SRVT-VIPF]`> 0 表示真实流量确有该段)
* - `msgx.pipeline.processing.ignored.total`:命中 US-04 忽略清单的报文条数
*
* 取数统一走 [BacklogSnapshotProvider]30 秒 TTL),因此指标抓取不会打穿数据库。
* 无法取数时以 `NaN` 上报(Micrometer 的惯例表示"本次无值"),而不是伪造 0。
@@ -101,6 +102,10 @@ class PipelineMetrics(
Gauge.builder("msgx.pipeline.codec.vipf_seen.total", counters) { it.vipfSeenCount().toDouble() }
.strongReference(true)
.register(registry)
Gauge.builder("msgx.pipeline.processing.ignored.total", counters) { it.ignoredCount().toDouble() }
.strongReference(true)
.register(registry)
}
private fun backlogGauge(name: String, value: (com.gzzn.omms.msgexchange.infra.persistence.Backlog) -> Double) {
@@ -170,7 +170,7 @@ interface MsgEventRepository {
/** 领取待发的整态事件:写入侧已保证每个 `FLID` 只有一行,读端不再去重合并。 */
fun mergePendingSchd(now: Instant, limit: Int): List<MsgEvent>
fun markSent(eventId: Long)
fun markSent(eventId: Long, now: Instant)
/**
* 条件确认:只在该行仍是本批读取到的代次(`EVENT_ID` + `STATE_VERSION`)且尚未发出时标记 `SENT`。
@@ -178,13 +178,20 @@ interface MsgEventRepository {
*
* @return 受影响行数(0 或 1
*/
fun markSentIfVersion(eventId: Long, stateVersion: Long): Int
fun markSentIfVersion(eventId: Long, stateVersion: Long, now: Instant): Int
fun markAllSent(eventIds: List<Long>)
fun markAllSent(eventIds: List<Long>, now: Instant)
fun scheduleRetry(eventId: Long, nextAttemptAt: Instant, attempts: Int)
fun markDead(eventId: Long, errorClass: ErrorClass, lastError: String, attempts: Int? = null)
/**
* 删除已确认投递且超过保留期的事件行:`STATE='SENT' AND SENT_AT < cutoff`。
* 带 `STATE='SENT'` 条件,避免与 upsertSchd 的并发写入冲突。
* @return 实际删除的行数
*/
fun deleteExpiredSent(cutoff: Instant, limit: Int): Int
}
/**
@@ -473,6 +473,7 @@ class JdbcMsgEventRepository(
next_attempt_at = NULL,
error_class = NULL,
last_error = NULL,
sent_at = NULL,
created_at = EXCLUDED.created_at
WHERE EXCLUDED.state_version > msg_event.state_version
OR (EXCLUDED.state_version = msg_event.state_version AND EXCLUDED.event_type = 'TOMBSTONE')
@@ -519,22 +520,26 @@ class JdbcMsgEventRepository(
::mapEvent,
)
override fun markSent(eventId: Long) {
ds.update("UPDATE msg_event SET state = 'SENT' WHERE event_id = ?", { ps -> ps.setLong(1, eventId) })
override fun markSent(eventId: Long, now: Instant) {
ds.update(
"UPDATE msg_event SET state = 'SENT', sent_at = ? WHERE event_id = ?",
{ ps -> ps.setTimestamp(1, now.toSqlTimestamp()); ps.setLong(2, eventId) },
)
}
/** 条件确认:仅当行仍是本批读到的代次且尚未发出时才标记,被新代次覆盖则影响 0 行。 */
override fun markSentIfVersion(eventId: Long, stateVersion: Long): Int =
override fun markSentIfVersion(eventId: Long, stateVersion: Long, now: Instant): Int =
ds.update(
"UPDATE msg_event SET state = 'SENT' WHERE event_id = ? AND state_version = ? AND state = 'PENDING'",
{ ps -> ps.setLong(1, eventId); ps.setLong(2, stateVersion) },
"UPDATE msg_event SET state = 'SENT', sent_at = ? WHERE event_id = ? AND state_version = ? AND state = 'PENDING'",
{ ps -> ps.setTimestamp(1, now.toSqlTimestamp()); ps.setLong(2, eventId); ps.setLong(3, stateVersion) },
)
override fun markAllSent(eventIds: List<Long>) {
override fun markAllSent(eventIds: List<Long>, now: Instant) {
if (eventIds.isEmpty()) return
val placeholders = eventIds.joinToString(",") { "?" }
ds.update("UPDATE msg_event SET state = 'SENT' WHERE event_id IN ($placeholders)") { ps ->
eventIds.forEachIndexed { i, id -> ps.setLong(i + 1, id) }
ds.update("UPDATE msg_event SET state = 'SENT', sent_at = ? WHERE event_id IN ($placeholders)") { ps ->
ps.setTimestamp(1, now.toSqlTimestamp())
eventIds.forEachIndexed { i, id -> ps.setLong(i + 2, id) }
}
}
@@ -557,6 +562,21 @@ class JdbcMsgEventRepository(
)
}
override fun deleteExpiredSent(cutoff: Instant, limit: Int): Int =
ds.update(
"""
DELETE FROM msg_event WHERE event_id IN (
SELECT event_id FROM msg_event
WHERE state = 'SENT' AND sent_at IS NOT NULL AND sent_at < ?
ORDER BY event_id LIMIT ?
)
""".trimIndent(),
{ ps ->
ps.setTimestamp(1, cutoff.toSqlTimestamp())
ps.setInt(2, limit)
},
)
private fun mapEvent(rs: ResultSet) = MsgEvent(
eventId = rs.getLong("event_id"),
target = rs.getString("target"),
@@ -570,6 +590,7 @@ class JdbcMsgEventRepository(
errorClass = rs.getString("error_class")?.let(ErrorClass::valueOf),
lastError = rs.getString("last_error"),
createdAt = rs.getInstant("created_at") ?: clock.instant(),
sentAt = rs.getInstant("sent_at"),
)
}
@@ -262,7 +262,7 @@ class StubMsgEvents : MsgEventRepository {
val id = ids.incrementAndGet()
rows[id] = e.copy(
eventId = id, state = EventStatus.PENDING, attempts = 0, nextAttemptAt = null,
errorClass = null, lastError = null,
errorClass = null, lastError = null, sentAt = null,
)
return id
}
@@ -278,19 +278,19 @@ class StubMsgEvents : MsgEventRepository {
.sortedBy { it.eventId!! }
.take(limit)
override fun markSent(eventId: Long) {
rows[eventId] = (rows[eventId] ?: return).copy(state = EventStatus.SENT)
override fun markSent(eventId: Long, now: Instant) {
rows[eventId] = (rows[eventId] ?: return).copy(state = EventStatus.SENT, sentAt = now)
}
override fun markSentIfVersion(eventId: Long, stateVersion: Long): Int {
override fun markSentIfVersion(eventId: Long, stateVersion: Long, now: Instant): Int {
val row = rows[eventId] ?: return 0
if (row.state != EventStatus.PENDING || row.stateVersion != stateVersion) return 0
rows[eventId] = row.copy(state = EventStatus.SENT)
rows[eventId] = row.copy(state = EventStatus.SENT, sentAt = now)
return 1
}
override fun markAllSent(eventIds: List<Long>) {
eventIds.forEach { markSent(it) }
override fun markAllSent(eventIds: List<Long>, now: Instant) {
eventIds.forEach { markSent(it, now) }
}
override fun scheduleRetry(eventId: Long, nextAttemptAt: Instant, attempts: Int) {
@@ -303,6 +303,16 @@ class StubMsgEvents : MsgEventRepository {
attempts = attempts ?: rows[eventId]?.attempts ?: 0,
)
}
override fun deleteExpiredSent(cutoff: Instant, limit: Int): Int {
val ids = rows.values
.filter { it.state == EventStatus.SENT && it.sentAt != null && it.sentAt < cutoff }
.sortedBy { it.eventId!! }
.take(limit)
.map { it.eventId!! }
ids.forEach { rows.remove(it) }
return ids.size
}
}
@Singleton
@@ -0,0 +1,30 @@
package com.gzzn.omms.msgexchange.jobs
import com.gzzn.omms.msgexchange.config.PipelineProps
import com.gzzn.omms.msgexchange.infra.persistence.MsgEventRepository
import jakarta.inject.Singleton
import java.time.Instant
/**
* MSG_EVENT 有界清理:删除已确认投递(`STATE='SENT'`)且 `SENT_AT` 超过保留期的行。
*
* 每轮 JobRunner tick 跑一次,删除量有上限(`CLEANUP_BATCH`),避免单次长事务。
* 只删 `STATE='SENT'` 的行——upsertSchd 写入的待发行状态是 `PENDING`,不会被误删。
*/
@Singleton
class EventCleanupJob(
private val msgEvents: MsgEventRepository,
private val props: PipelineProps,
) {
data class CleanupOutcome(val deleted: Int)
fun run(now: Instant): CleanupOutcome {
val cutoff = now.minus(props.pipeline.eventRetention)
val deleted = msgEvents.deleteExpiredSent(cutoff, CLEANUP_BATCH)
return CleanupOutcome(deleted)
}
private companion object {
const val CLEANUP_BATCH = 500
}
}
@@ -1,6 +1,7 @@
package com.gzzn.omms.msgexchange.jobs
import com.gzzn.omms.msgexchange.config.OperationDayProps
import com.gzzn.omms.msgexchange.config.PipelineProps
import com.gzzn.omms.msgexchange.infra.metrics.JobActivity
import com.gzzn.omms.msgexchange.processing.BackfillService
import jakarta.inject.Singleton
@@ -23,8 +24,10 @@ import java.time.ZoneId
class JobRunner(
private val backfill: BackfillService,
private val historySweep: HistorySweepJob,
private val eventCleanup: EventCleanupJob,
private val clock: Clock,
private val activity: JobActivity,
private val props: PipelineProps,
operationDayProps: OperationDayProps,
) {
private val log = org.slf4j.LoggerFactory.getLogger(JobRunner::class.java)
@@ -43,7 +46,7 @@ class JobRunner(
running = true
activity.started()
thread = Thread.ofPlatform().name("msgx-jobs").daemon(true).start { loop() }
log.info("job runner started (backfill {}s, history daily 03:30 {})", TICK_PERIOD.seconds, zone)
log.info("job runner started (backfill {}s, history daily 03:30 {}, event cleanup {})", TICK_PERIOD.seconds, zone, props.pipeline.eventRetention)
}
fun stop() {
@@ -75,6 +78,10 @@ class JobRunner(
try {
val selected = backfill.sweep(startedAt)
maybeHistorySweep()
val cleanupOutcome = eventCleanup.run(startedAt)
if (cleanupOutcome.deleted > 0) {
log.info("event cleanup: deleted {} expired SENT rows", cleanupOutcome.deleted)
}
activity.tickFinished(startedAt, (System.nanoTime() - startedNanos) / 1_000_000, selected)
} catch (e: InterruptedException) {
throw e
@@ -0,0 +1,24 @@
package com.gzzn.omms.msgexchange.processing
/**
* US-04 忽略清单:基线规则为 `LDM-*`、`REGN-*`、`RSTA-*`、`EROR-*`。
*
* 匹配按 `TYPE` 前缀做(`TYPE-*` 表示该 TYPE 下所有 STYP 均忽略);
* `MetaFields.type` 已由 codec 统一大写,这里直接用大写常量比较。
* 返回命中的规则标签(如 `LDM-*`),未命中返回 null。
*/
object IgnoreRules {
private val rules: List<Pair<String, String>> = listOf(
"LDM" to "LDM-*",
"REGN" to "REGN-*",
"RSTA" to "RSTA-*",
"EROR" to "EROR-*",
)
fun match(type: String): String? {
for ((prefix, label) in rules) {
if (type == prefix) return label
}
return null
}
}
@@ -211,6 +211,15 @@ class MessageProcessor(
}
}
// US-04:忽略清单命中 → SKIPPED + 回填意图,不取 PIPELINE_LOCK、不写航班表、不创建 MSG_EVENT
val ignoreRule = IgnoreRules.match(decoded.meta.type)
if (ignoreRule != null) {
log.info("ignored:{} -> SKIPPED msgId={}", ignoreRule, head.msgId)
counters.ignoredAdd()
procState.markTerminal(head.msgId, ProcStatus.SKIPPED, lastError = "ignored:$ignoreRule", now = clock.instant())
return
}
// 按报文类型分派:日计划走 SCHD,其余走 FLOP / FDEL / ADFT;报文缺载荷直接判为非法报文的死信
val result: ApplyResult = when (val kind = decoded.kind) {
is MsgKind.Schd -> {