refactor(ingress): 移除 G1 窗口补偿扫描与迟到到达检测
实际负载不足 10 条/秒,G1 属过度防御。删除迟到到达检测机制、 existingIds 接口、PipelineCounters 字段、lateDetect 配置, 以及文档中 CLM-1/CLM-2 声明与 G1 缺口索引。
This commit is contained in:
@@ -64,17 +64,6 @@ class PipelineProps {
|
||||
*/
|
||||
var cutoverWatermark: String? = null
|
||||
|
||||
/**
|
||||
* 迟到到达检测(只读,ACM2-41 阶段 0):复查"已判定为永久空洞的 ID"是否后来真的出现。
|
||||
* 默认 60 秒一轮;设为 0 或负数即关闭。
|
||||
*
|
||||
* 检测只计数与告警,**不入队、不改变任何处理语义**——自动补入队(阶段 1)需要先与库方定案。
|
||||
*/
|
||||
var lateDetectPeriod: Duration = Duration.ofSeconds(60)
|
||||
|
||||
/** 每轮最多复查多少个被放行的空洞 ID。 */
|
||||
var lateDetectBatch: Int = 200
|
||||
|
||||
/**
|
||||
* 普通事件(`KAFKA:msg`)每轮向一个目标领取的条数上限。
|
||||
* 逐条领取会让投递吞吐被"每条一次 DB 往返 + 一轮一次 sleep"压到每秒 1 条。
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
package com.gzzn.omms.msgexchange.infra.metrics
|
||||
|
||||
import jakarta.inject.Singleton
|
||||
import java.util.concurrent.atomic.AtomicLong
|
||||
|
||||
/**
|
||||
* 管道运行期的**进程内**计数。
|
||||
@@ -12,26 +11,4 @@ import java.util.concurrent.atomic.AtomicLong
|
||||
* 注意:计数在重启后归零。需要跨重启的累计值应由指标后端聚合,不在这里做持久化。
|
||||
*/
|
||||
@Singleton
|
||||
class PipelineCounters {
|
||||
private val holeAgedOut = AtomicLong(0)
|
||||
private val lateArrivalDetected = AtomicLong(0)
|
||||
|
||||
/** 水位因空洞超过老化阈值而放行(判定为永久空洞)的次数。 */
|
||||
fun holeAgedOutIncrement() {
|
||||
holeAgedOut.incrementAndGet()
|
||||
}
|
||||
|
||||
fun holeAgedOutCount(): Long = holeAgedOut.get()
|
||||
|
||||
/**
|
||||
* "迟到到达"检测命中的**不同**消息 ID 数(ACM2-41 阶段 0)。
|
||||
*
|
||||
* 含义:该 ID 曾被判定为永久空洞并放行,之后却真的出现在信箱里——即上游提交晚于水位推进。
|
||||
* 阶段 0 只计数与告警,**不会**补入队;因此这个值 > 0 表示"确有迟到发生,需要与库方对契约"。
|
||||
*/
|
||||
fun lateArrivalDetectedIncrement() {
|
||||
lateArrivalDetected.incrementAndGet()
|
||||
}
|
||||
|
||||
fun lateArrivalDetectedCount(): Long = lateArrivalDetected.get()
|
||||
}
|
||||
class PipelineCounters
|
||||
|
||||
@@ -22,7 +22,6 @@ import java.time.Duration
|
||||
* - `msgx.pipeline.backfill.abandoned`:已放弃自动回填的条数(**非 0 需人工对账**)
|
||||
* - `msgx.pipeline.backfill.oldest_unmarked_seconds`:最老一条仍待自动回填的年龄
|
||||
* - `msgx.pipeline.watermark.lag`:水位落后信箱最新 ID 的距离
|
||||
* - `msgx.pipeline.hole.aged_out.total`:永久空洞放行次数
|
||||
* - `msgx.pipeline.job.heartbeat_age_seconds`:距上一次作业 tick 完成的秒数(未跑过为 -1)
|
||||
* - `msgx.pipeline.job.last_failure_age_seconds`:距最近一次作业 tick 失败的秒数(从未失败为 -1)
|
||||
* - `msgx.pipeline.job.ticks.total` / `msgx.pipeline.job.failures.total`:作业 tick 完成/抛错次数
|
||||
@@ -44,7 +43,6 @@ class PipelineMetrics(
|
||||
private val backlogs: BacklogSnapshotProvider,
|
||||
private val cursor: BeanProvider<InboxCursorRepository>,
|
||||
private val mailbox: BeanProvider<CminmsgInboxRepository>,
|
||||
private val counters: PipelineCounters,
|
||||
private val activity: JobActivity,
|
||||
private val clock: Clock,
|
||||
) {
|
||||
@@ -71,15 +69,6 @@ class PipelineMetrics(
|
||||
if (watermark != null && maxId != null) (maxId - watermark).toDouble() else -1.0
|
||||
}.strongReference(true).register(registry)
|
||||
|
||||
Gauge.builder("msgx.pipeline.hole.aged_out.total", counters) { it.holeAgedOutCount().toDouble() }
|
||||
.strongReference(true)
|
||||
.register(registry)
|
||||
|
||||
// 迟到到达检测命中数(阶段 0 只观测):> 0 表示上游提交确实晚于水位推进,需要与库方对契约。
|
||||
Gauge.builder("msgx.pipeline.late_arrival.detected.total", counters) { it.lateArrivalDetectedCount().toDouble() }
|
||||
.strongReference(true)
|
||||
.register(registry)
|
||||
|
||||
// 作业心跳与扫描积压:作业线程是回填的唯一驱动,停摆必须能被 /metrics 与 /health 看见。
|
||||
Gauge.builder("msgx.pipeline.job.heartbeat_age_seconds", activity) { a ->
|
||||
a.snapshot().lastTickAt?.let { Duration.between(it, clock.instant()).seconds.toDouble() } ?: -1.0
|
||||
|
||||
@@ -353,12 +353,6 @@ interface CminmsgInboxRepository {
|
||||
/** 信箱当前最小 ID,空表返回 null;切流播种用它推算 `W = MIN(ID) − 1`。 */
|
||||
fun minId(): Long?
|
||||
|
||||
/**
|
||||
* 批量查询这些 ID 里哪些**当前存在于信箱**。
|
||||
* 只用于"迟到到达"检测(被放行的空洞 ID 后来是否真的出现),不读大字段。
|
||||
*/
|
||||
fun existingIds(msgIds: Collection<Long>): Set<Long>
|
||||
|
||||
/**
|
||||
* 把处理标记写回信箱,并且**只写还是空标记的行**:库里已有值时不覆盖、不回退,
|
||||
* 重复调用没有副作用。结果明确区分本次写入、已有标记和信箱行缺失。
|
||||
|
||||
-14
@@ -86,20 +86,6 @@ class JdbcCminmsgInboxRepository(
|
||||
rs.getLong("min_id").takeIf { !rs.wasNull() }
|
||||
}
|
||||
|
||||
override fun existingIds(msgIds: Collection<Long>): Set<Long> {
|
||||
if (msgIds.isEmpty()) return emptySet()
|
||||
val out = linkedSetOf<Long>()
|
||||
// 分块避免 IN 列表过长(迟到检测一次最多查 late-detect-batch 个)。
|
||||
msgIds.chunked(200).forEach { chunk ->
|
||||
val placeholders = chunk.joinToString(",") { "?" }
|
||||
ds.query(
|
||||
"SELECT CMINMSGS_ID FROM cminmsgs WHERE CMINMSGS_ID IN ($placeholders)",
|
||||
{ ps -> chunk.forEachIndexed { i, id -> ps.setLong(i + 1, id) } },
|
||||
) { rs -> rs.getLong("CMINMSGS_ID") }.forEach { out += it }
|
||||
}
|
||||
return out
|
||||
}
|
||||
|
||||
/**
|
||||
* 只更新还是空标记的行,所以重复调用不会覆盖库里已有的值;
|
||||
* 影响 0 行时再查一次主键,区分“已有标记”和“信箱行缺失”;后者不能记为回填成功。
|
||||
|
||||
@@ -484,8 +484,6 @@ class StubInbox(private val clock: Clock = Clock.systemUTC()) : CminmsgInboxRepo
|
||||
|
||||
override fun minId(): Long? = raws.keys.minOrNull()
|
||||
|
||||
override fun existingIds(msgIds: Collection<Long>): Set<Long> = msgIds.filter { raws.containsKey(it) }.toSet()
|
||||
|
||||
/** 只写还没有标记的行,并区分已有标记与行缺失。 */
|
||||
override fun markProcessedIfUnmarked(msgId: Long, value: String): MailboxMarkResult {
|
||||
if (!raws.containsKey(msgId)) return MailboxMarkResult.MISSING
|
||||
|
||||
@@ -6,7 +6,6 @@ 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 com.gzzn.omms.msgexchange.infra.metrics.PipelineCounters
|
||||
import jakarta.inject.Singleton
|
||||
import java.time.Clock
|
||||
import java.time.Duration
|
||||
@@ -40,7 +39,6 @@ class InboxPoller(
|
||||
private val txManager: PipelineTransactionManager,
|
||||
private val props: PipelineProps,
|
||||
private val clock: Clock,
|
||||
private val counters: PipelineCounters,
|
||||
) {
|
||||
private val log = org.slf4j.LoggerFactory.getLogger(InboxPoller::class.java)
|
||||
|
||||
@@ -50,7 +48,6 @@ class InboxPoller(
|
||||
/** @return 本轮新登记的消息条数(已登记过的行不计入,也不影响水位推进)。 */
|
||||
fun pollOnce(now: Instant): Int {
|
||||
seedCutoverWatermarkIfConfigured(now)
|
||||
detectLateArrivalsIfDue(now)
|
||||
val batch = props.pipeline.claimBatch.coerceAtLeast(1)
|
||||
val watermark = cursor.load()
|
||||
val rows = mailbox.readRange(watermark.committedUpTo, batch)
|
||||
@@ -68,8 +65,6 @@ class InboxPoller(
|
||||
} else {
|
||||
// 缺口等太久了:当成永久缺失跳过,让水位继续往前走
|
||||
committedTo = rows.first { it.msgId > contiguous }.msgId - 1
|
||||
counters.holeAgedOutIncrement()
|
||||
rememberAgedHoles(contiguous + 1, committedTo)
|
||||
log.warn("hole after W={} aged out, watermark advanced to {}", contiguous, committedTo)
|
||||
}
|
||||
}
|
||||
@@ -131,65 +126,6 @@ class InboxPoller(
|
||||
/** "拒绝重新播种"的告警只打一次,避免每轮刷屏。 */
|
||||
private val refusedReseed = AtomicBoolean(false)
|
||||
|
||||
/**
|
||||
* "已被判定为永久空洞"的 ID 监视队列(ACM2-41 阶段 0,只读检测)。
|
||||
*
|
||||
* 为什么只监视这些 ID 就够:水位只会越过两类 ID——连续存在的(已入队)与被判定为永久空洞的。
|
||||
* 因此凡是"水位越过之后才出现在信箱里"的 ID,必然曾经被当作空洞放行过。
|
||||
*
|
||||
* 队列有界([HOLE_WATCH_LIMIT]),单轮复查量受 `late-detect-batch` 约束;
|
||||
* 它是**进程内**状态,重启后丢失,因此检测是尽力而为的观测(阶段 0 不做补入队)。
|
||||
*/
|
||||
private val holeWatch = ArrayDeque<Long>()
|
||||
|
||||
private var lastLateDetectAt: Instant? = null
|
||||
|
||||
/** 记下被放行的空洞 ID(超大空洞只记前一段,避免内存被吃掉)。 */
|
||||
private fun rememberAgedHoles(from: Long, to: Long) {
|
||||
val cap = props.pipeline.lateDetectBatch.coerceAtLeast(1) * 2
|
||||
var id = from
|
||||
var added = 0
|
||||
while (id <= to && added < cap && holeWatch.size < HOLE_WATCH_LIMIT) {
|
||||
holeWatch.addLast(id)
|
||||
id++
|
||||
added++
|
||||
}
|
||||
}
|
||||
|
||||
/**
|
||||
* 只读的迟到检测:复查监视队列里的 ID 是否**真的出现在信箱里**。
|
||||
* 命中即计数 + 告警,**不入队**(补入队属阶段 1,需先与库方就提交契约定案)。
|
||||
*/
|
||||
private fun detectLateArrivalsIfDue(now: Instant) {
|
||||
val period = props.pipeline.lateDetectPeriod
|
||||
if (period.isZero || period.isNegative) return
|
||||
val last = lastLateDetectAt
|
||||
if (last != null && Duration.between(last, now) < period) return
|
||||
lastLateDetectAt = now
|
||||
if (holeWatch.isEmpty()) return
|
||||
|
||||
val batch = props.pipeline.lateDetectBatch.coerceAtLeast(1)
|
||||
val probe = ArrayList<Long>(minOf(batch, holeWatch.size))
|
||||
repeat(minOf(batch, holeWatch.size)) { probe += holeWatch.removeFirst() }
|
||||
|
||||
val present = try {
|
||||
mailbox.existingIds(probe)
|
||||
} catch (e: Exception) {
|
||||
log.warn("late-arrival detection skipped: {}", e.message)
|
||||
probe.forEach { holeWatch.addLast(it) } // 查不动就把监视放回去,别丢
|
||||
return
|
||||
}
|
||||
if (present.isNotEmpty()) {
|
||||
present.forEach { counters.lateArrivalDetectedIncrement() }
|
||||
log.error(
|
||||
"LATE ARRIVAL: {} message(s) appeared below the watermark after being aged out: {}",
|
||||
present.size, present.sorted().take(20),
|
||||
)
|
||||
}
|
||||
// 已命中的不再监视(计数表示"不同 ID");仍未出现的轮转回队尾继续看。
|
||||
probe.filter { it !in present }.forEach { holeWatch.addLast(it) }
|
||||
}
|
||||
|
||||
fun loop() {
|
||||
running = true
|
||||
log.info("inbox poller loop started")
|
||||
@@ -236,8 +172,4 @@ class InboxPoller(
|
||||
}
|
||||
}
|
||||
|
||||
private companion object {
|
||||
/** 监视队列上限:防止超大空洞把内存吃光(阶段 0 是尽力而为的观测,不追求全覆盖)。 */
|
||||
const val HOLE_WATCH_LIMIT = 4096
|
||||
}
|
||||
}
|
||||
|
||||
@@ -24,8 +24,6 @@ msgx:
|
||||
backfill-max-attempts: 100 # 单行重试的告警阈值;放弃判据是 R 超期,不是次数(design「回填」)
|
||||
# 一次性切流播种:默认(注释掉)不播种。min=读现存全部 | zero=从 0 按空洞规则 | max=跳过可见存量 | <id>
|
||||
# cutover-watermark: min
|
||||
late-detect-period: 60s # 迟到到达检测(只读,阶段 0):复查被放行的空洞 ID 是否后来真的出现;0=关闭
|
||||
late-detect-batch: 200 # 每轮最多复查多少个空洞 ID
|
||||
autostart: false # U07:启动即拉起 Pump/Dispatcher 循环;需真实仓储或 msgx.stubs=true 才开启(dev 见 application-dev.yml)
|
||||
schd:
|
||||
flush-period: 3s # KEEP 现役推送节律
|
||||
|
||||
Reference in New Issue
Block a user