zaynt4606 commented on code in PR #3858: URL: https://github.com/apache/celeborn/pull/3858#discussion_r4266481576
########## worker/src/main/scala/org/apache/celeborn/service/deploy/worker/metrics/DefaultScaleMetricCollector.scala: ########## @@ -0,0 +1,112 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ + +package org.apache.celeborn.service.deploy.worker.metrics + +import java.lang.management.{ManagementFactory, OperatingSystemMXBean} +import java.util.concurrent.{ScheduledFuture, TimeUnit} + +import scala.util.control.NonFatal + +import org.apache.celeborn.common.CelebornConf +import org.apache.celeborn.common.internal.Logging +import org.apache.celeborn.common.meta.DiskStatus +import org.apache.celeborn.common.metrics.WorkerStats +import org.apache.celeborn.common.util.{DiskUtils, ThreadUtils} +import org.apache.celeborn.service.deploy.worker.Worker + +class DefaultScaleMetricCollector private[metrics] ( + conf: CelebornConf, + operatingSystem: OperatingSystemMXBean) + extends ScaleMetricCollector with Logging { + + def this(conf: CelebornConf) = + this(conf, ManagementFactory.getOperatingSystemMXBean) + + private val memoryWindow = new MetricSlidingWindow(conf.scaleMetricWindowSize) + private val executor = + ThreadUtils.newDaemonSingleThreadScheduledExecutor("scale-metric-collector") + private var scheduledTask: ScheduledFuture[_] = _ + + @volatile private var stopped = false + @volatile private var currentStats: Option[WorkerStats] = None + + override def init(worker: Worker): Unit = synchronized { + require(!stopped && scheduledTask == null, "Metric collector can only be initialized once") + scheduledTask = executor.scheduleWithFixedDelay( + new Runnable { + override def run(): Unit = collectOnce(worker) + }, + conf.scaleMetricInterval, + conf.scaleMetricInterval, + TimeUnit.MILLISECONDS) + } + + override def stop(): Unit = synchronized { + stopped = true + currentStats = None + if (scheduledTask != null) { + scheduledTask.cancel(true) + } + executor.shutdownNow() + } + + override def currentWorkerStats(): Option[WorkerStats] = { + if (stopped) None + else currentStats + } + + private[metrics] def collectOnce(worker: Worker): Unit = { + try { + val metrics = memoryMetrics(worker) ++ systemLoadMetrics() ++ diskMetrics(worker) + currentStats = Some(WorkerStats(metrics)) + } catch { + case NonFatal(e) => + logWarning("Failed to collect worker scaling metrics", e) + } + } + + private def memoryMetrics(worker: Worker): Map[String, Double] = { + val capacity = worker.memoryManager.maxDirectMemory + val used = worker.memoryManager.getMemoryUsage + memoryWindow.update(used) + Map(WorkerStats.NettyMemoryUsedRatio -> (memoryWindow.average / capacity)) + } + + private def systemLoadMetrics(): Map[String, Double] = { + val load = operatingSystem.getSystemLoadAverage + val processors = operatingSystem.getAvailableProcessors + Map(WorkerStats.LastMinuteSystemLoadRatio -> (load / processors)) + } + + private def diskMetrics(worker: Worker): Map[String, Double] = { + val storageManager = worker.storageManager + val disks = storageManager.localDisksSnapshot() + val usableDisks = disks.filter(_.status != DiskStatus.IO_HANG) Review Comment: [P2] Exclude failed disks from the available-capacity metrics This filter only excludes `IO_HANG`, so disks in `READ_OR_WRITE_FAILURE` or `CRITICAL_ERROR` still contribute their last `actualUsableSpace`. `ObservedDevice` changes the status without clearing that value, and `StorageManager.updateDiskInfos()` skips disks in `READ_OR_WRITE_FAILURE`. Storage allocation already rejects these disks through `DiskInfo.isHealthy`. I reproduced both cases on `cf21ba4c`: with one disk whose effective capacity is 1024 bytes and cached usable space is 512 bytes, changing its status to either failure state still reports `DiskRemainingSize = 512` and `DiskUsedRatio = 0.5`, although the disk cannot accept new allocations. This would give future scaling policies an incorrect available-capacity signal. Could we exclude non-allocatable disks from `usableBytes`, using the same health criteria as storage allocation, while retaining the capacity denominator as for `IO_HANG`? Please also add tests for both failure states and recovery to `HEALTHY`. The existing 27 selected tests pass, but the two additional failure-state probes reproduce this issue. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
