perf: limit the inflight requests (#2100)

* docs: add todos Signed-off-by: Ning Yu <[email protected]> * perf(network): limit the inflight requests by size Signed-off-by: Ning Yu <[email protected]> * perf(ReplicaManager): limit the queue size of the `fetchExecutor`s Signed-off-by: Ning Yu <[email protected]> * perf(KafkaApis): limit the queue size of async request handlers Signed-off-by: Ning Yu <[email protected]> * refactor(network): make "queued.max.requests.size.bytes" configurable Signed-off-by: Ning Yu <[email protected]> * style: fix lint Signed-off-by: Ning Yu <[email protected]> * fix(network): limit the min queued request size per queue Signed-off-by: Ning Yu <[email protected]> --------- Signed-off-by: Ning Yu <[email protected]>
AutoMQ · Nov 1, 2024 · a7250cd · a7250cd
1 parent 33e0be0
commit a7250cd
Show file tree

Hide file tree

Showing 8 changed files with 79 additions and 15 deletions.
diff --git a/core/src/main/resources/jmx/rules/broker.yaml b/core/src/main/resources/jmx/rules/broker.yaml
@@ -207,6 +207,13 @@ rules:
         type: gauge
         desc: Size of the request queue
 
+  - bean: kafka.network:type=RequestChannel,name=AvailableRequestSize
+    mapping:
+      Value:
+        metric: kafka.available.request.size
+        type: gauge
+        desc: Remaining permitted request size in the request queue
+
   - bean: kafka.network:type=RequestChannel,name=ResponseQueueSize
     mapping:
       Value:
@@ -370,4 +377,4 @@ rules:
       connection-accept-throttle-time:
         metric: kafka.listener.connection.accept.throttle.time
         type: gauge
-        desc: The average throttle-time pre listener
+        desc: The average throttle-time pre listener
diff --git a/core/src/main/scala/kafka/network/RequestChannel.scala b/core/src/main/scala/kafka/network/RequestChannel.scala
@@ -50,6 +50,10 @@ object RequestChannel extends Logging {
   private val ResponseQueueSizeMetric = "ResponseQueueSize"
   val ProcessorMetricTag = "processor"
 
+  // AutoMQ inject start
+  private val AvailableRequestSizeMetric = "AvailableRequestSize"
+  // AutoMQ inject end
+
   private def isRequestLoggingEnabled: Boolean = requestLogger.underlying.isDebugEnabled
 
   sealed trait BaseRequest
@@ -347,6 +351,9 @@ object RequestChannel extends Logging {
 }
 
 class RequestChannel(val queueSize: Int,
+                     // AutoMQ inject start
+                     val queuedRequestSize: Int,
+                      // AutoMQ inject end
                      val metricNamePrefix: String,
                      time: Time,
                      val metrics: RequestChannel.Metrics) {
@@ -355,12 +362,23 @@ class RequestChannel(val queueSize: Int,
   private val metricsGroup = new KafkaMetricsGroup(this.getClass)
 
   private val requestQueue = new ArrayBlockingQueue[BaseRequest](queueSize)
+  // AutoMQ inject start
+  /**
+   * Queue of requests to be handled, in the order they arrived.
+   * Note: Before any request enters this queue, it needs to acquire {@link multiQueuedRequestSizeSemaphore}
+   */
   private val multiRequestQueue = new java.util.ArrayList[ArrayBlockingQueue[BaseRequest]]()
-
+  /**
+   * Semaphore to limit the total size of requests in the {@link multiRequestQueue}.
+   */
+  private val multiQueuedRequestSizeSemaphore = new java.util.ArrayList[Semaphore]()
+  private val availableRequestSizeMetricName = metricNamePrefix.concat(AvailableRequestSizeMetric)
+  // AutoMQ inject end
   private val processors = new ConcurrentHashMap[Int, Processor]()
   private val requestQueueSizeMetricName = metricNamePrefix.concat(RequestQueueSizeMetric)
   private val responseQueueSizeMetricName = metricNamePrefix.concat(ResponseQueueSizeMetric)
   private val callbackQueue = new ArrayBlockingQueue[BaseRequest](queueSize)
+  // AutoMQ inject start
   private val multiCallbackQueue = new java.util.ArrayList[ArrayBlockingQueue[BaseRequest]]()
   private var notifiedShutdown = false
 
@@ -371,6 +389,14 @@ class RequestChannel(val queueSize: Int,
       requestQueue.size()
     }
   })
+  metricsGroup.newGauge(availableRequestSizeMetricName, () => {
+    multiQueuedRequestSizeSemaphore.stream().mapToInt(s => s.availablePermits()).sum()
+  })
+
+  def this(queueSize: Int, metricNamePrefix: String, time: Time, metrics: RequestChannel.Metrics) {
+    this(queueSize, Integer.MAX_VALUE, metricNamePrefix, time, metrics)
+  }
+  // AutoMQ inject end
 
   metricsGroup.newGauge(responseQueueSizeMetricName, () => {
     processors.values.asScala.foldLeft(0) {(total, processor) =>
@@ -386,14 +412,21 @@ class RequestChannel(val queueSize: Int,
       Map(ProcessorMetricTag -> processor.id.toString).asJava)
   }
 
-  def registerNRequestHandler(count: Int): util.List[BlockingQueue[BaseRequest]] = {
+  // AutoMQ inject start
+  def registerNRequestHandler(count: Int): Unit = {
     val queueSize = math.max(this.queueSize / count, 1)
-    for (i <- 0 until count) {
+    // TODO: maxQueuedRequestSize will be 100 / 8 = 12.5 MiB as a default.
+    //  However, if the request size is too large, it will block at the semaphore.
+    //  Currently, the max request size is 1 MiB (max.request.size) by default, so it is not very problematic.
+    val maxQueuedRequestSize = math.max(this.queuedRequestSize / count, 10 * 1024 * 1024)
+    for (_ <- 0 until count) {
       multiRequestQueue.add(new ArrayBlockingQueue[BaseRequest](queueSize))
+      multiQueuedRequestSizeSemaphore.add(new Semaphore(maxQueuedRequestSize))
       multiCallbackQueue.add(new ArrayBlockingQueue[BaseRequest](queueSize))
     }
     Collections.unmodifiableList(multiRequestQueue)
   }
+  // AutoMQ inject end
 
   def removeProcessor(processorId: Int): Unit = {
     processors.remove(processorId)
@@ -403,6 +436,8 @@ class RequestChannel(val queueSize: Int,
   /** Send a request to be handled, potentially blocking until there is room in the queue for the request */
   def sendRequest(request: RequestChannel.Request): Unit = {
     if (multiRequestQueue.size() != 0) {
+      val requestSizeSemaphore = multiQueuedRequestSizeSemaphore.get(math.abs(request.context.connectionId.hashCode % multiQueuedRequestSizeSemaphore.size()))
+      requestSizeSemaphore.acquire(request.sizeInBytes)
       val requestQueue = multiRequestQueue.get(math.abs(request.context.connectionId.hashCode % multiRequestQueue.size()))
       requestQueue.put(request)
     } else {
@@ -505,21 +540,26 @@ class RequestChannel(val queueSize: Int,
     }
   }
 
+  // AutoMQ inject start
   def receiveRequest(timeout: Long, id: Int): RequestChannel.BaseRequest = {
     val callbackQueue = multiCallbackQueue.get(id)
     val requestQueue = multiRequestQueue.get(id)
+    val requestSizeSemaphore = multiQueuedRequestSizeSemaphore.get(id)
     val callbackRequest = callbackQueue.poll()
     if (callbackRequest != null)
       callbackRequest
     else {
       val request = requestQueue.poll(timeout, TimeUnit.MILLISECONDS)
       request match {
         case WakeupRequest => callbackQueue.poll()
+        case request: Request =>
+          requestSizeSemaphore.release(request.sizeInBytes)
+          request
         case _ => request
       }
     }
   }
-
+  // AutoMQ inject end
   /** Get the next request or block until there is one */
   @Deprecated
   def receiveRequest(): RequestChannel.BaseRequest =

diff --git a/core/src/main/scala/kafka/network/SocketServer.scala b/core/src/main/scala/kafka/network/SocketServer.scala
@@ -97,7 +97,10 @@ class SocketServer(val config: KafkaConfig,
   private val memoryPool = if (config.queuedMaxBytes > 0) new SimpleMemoryPool(config.queuedMaxBytes, config.socketRequestMaxBytes, false, memoryPoolSensor) else MemoryPool.NONE
   // data-plane
   private[network] val dataPlaneAcceptors = new ConcurrentHashMap[EndPoint, DataPlaneAcceptor]()
-  val dataPlaneRequestChannel = new RequestChannel(maxQueuedRequests, DataPlaneAcceptor.MetricPrefix, time, apiVersionManager.newRequestMetrics)
+  // AutoMQ inject start
+  // val dataPlaneRequestChannel = new RequestChannel(maxQueuedRequests, DataPlaneAcceptor.MetricPrefix, time, apiVersionManager.newRequestMetrics)
+   val dataPlaneRequestChannel = new RequestChannel(maxQueuedRequests, config.queuedMaxRequestSize, DataPlaneAcceptor.MetricPrefix, time, apiVersionManager.newRequestMetrics)
+  // AutoMQ inject end
   // control-plane
   private[network] var controlPlaneAcceptorOpt: Option[ControlPlaneAcceptor] = None
   val controlPlaneRequestChannelOpt: Option[RequestChannel] = config.controlPlaneListenerName.map(_ =>

diff --git a/core/src/main/scala/kafka/server/KafkaConfig.scala b/core/src/main/scala/kafka/server/KafkaConfig.scala
@@ -100,7 +100,7 @@ object KafkaConfig {
   AutoBalancerControllerConfig.CONFIG_DEF.configKeys().values().forEach(key => configDef.define(key))
   AutoBalancerMetricsReporterConfig.CONFIG_DEF.configKeys().values().forEach(key => configDef.define(key))
   // AutoMQ inject end
-  
+
 
   def configNames: Seq[String] = configDef.names.asScala.toBuffer.sorted
   private[server] def defaultValues: Map[String, _] = configDef.defaultValues.asScala
@@ -435,6 +435,9 @@ class KafkaConfig private(doLog: Boolean, val props: util.Map[_, _])
   val connectionsMaxIdleMs = getLong(SocketServerConfigs.CONNECTIONS_MAX_IDLE_MS_CONFIG)
   val failedAuthenticationDelayMs = getInt(SocketServerConfigs.FAILED_AUTHENTICATION_DELAY_MS_CONFIG)
   val queuedMaxRequests = getInt(SocketServerConfigs.QUEUED_MAX_REQUESTS_CONFIG)
+  // AutoMQ inject start
+  val queuedMaxRequestSize = getInt(SocketServerConfigs.QUEUED_MAX_REQUESTS_SIZE_BYTES_CONFIG)
+  // AutoMQ inject end
   val queuedMaxBytes = getLong(SocketServerConfigs.QUEUED_MAX_BYTES_CONFIG)
   def numNetworkThreads = getInt(SocketServerConfigs.NUM_NETWORK_THREADS_CONFIG)
 

diff --git a/core/src/main/scala/kafka/server/KafkaRequestHandler.scala b/core/src/main/scala/kafka/server/KafkaRequestHandler.scala
@@ -216,7 +216,7 @@ class KafkaRequestHandlerPool(
   this.logIdent = "[" + logAndThreadNamePrefix + " Kafka Request Handler on Broker " + brokerId + "], "
   val runnables = new mutable.ArrayBuffer[KafkaRequestHandler](numThreads)
 
-  var multiRequestQueue = requestChannel.registerNRequestHandler(numThreads)
+  requestChannel.registerNRequestHandler(numThreads)
 
   for (i <- 0 until numThreads) {
     createHandler(i)

diff --git a/core/src/main/scala/kafka/server/streamaspect/ElasticKafkaApis.scala b/core/src/main/scala/kafka/server/streamaspect/ElasticKafkaApis.scala
@@ -1,6 +1,7 @@
 package kafka.server.streamaspect
 
 import com.automq.stream.s3.metrics.TimerUtil
+import com.automq.stream.utils.threads.S3StreamThreadPoolMonitor
 import com.yammer.metrics.core.Histogram
 import kafka.automq.zonerouter.{ClientIdMetadata, NoopProduceRouter, ProduceRouter}
 import kafka.coordinator.transaction.TransactionCoordinator
@@ -29,7 +30,7 @@ import org.apache.kafka.common.requests.s3.AutomqZoneRouterRequest
 import org.apache.kafka.common.requests.{AbstractResponse, DeleteTopicsRequest, DeleteTopicsResponse, FetchRequest, FetchResponse, ProduceRequest, ProduceResponse, RequestUtils}
 import org.apache.kafka.common.resource.Resource.CLUSTER_NAME
 import org.apache.kafka.common.resource.ResourceType.{CLUSTER, TOPIC, TRANSACTIONAL_ID}
-import org.apache.kafka.common.utils.{ThreadUtils, Time}
+import org.apache.kafka.common.utils.Time
 import org.apache.kafka.coordinator.group.GroupCoordinator
 import org.apache.kafka.server.ClientMetricsManager
 import org.apache.kafka.server.authorizer.Authorizer
@@ -40,7 +41,7 @@ import org.apache.kafka.storage.internals.log.{FetchIsolation, FetchParams, Fetc
 import java.util
 import java.util.{Collections, Optional}
 import java.util.concurrent.atomic.AtomicLong
-import java.util.concurrent.{ExecutorService, Executors, TimeUnit}
+import java.util.concurrent.{ExecutorService, TimeUnit}
 import java.util.stream.IntStream
 import scala.annotation.nowarn
 import scala.collection.{Map, Seq, mutable}
@@ -76,8 +77,8 @@ class ElasticKafkaApis(
   tokenManager: DelegationTokenManager,
   apiVersionManager: ApiVersionManager,
   clientMetricsManager: Option[ClientMetricsManager],
-  val asyncHandleExecutor: ExecutorService = Executors.newSingleThreadExecutor(ThreadUtils.createThreadFactory("kafka-apis-async-handle-executor-%d", true)),
-  val listOffsetHandleExecutor: ExecutorService = Executors.newSingleThreadExecutor(ThreadUtils.createThreadFactory("kafka-apis-list-offset-handle-executor-%d", true))
+  val deleteTopicHandleExecutor: ExecutorService = S3StreamThreadPoolMonitor.createAndMonitor(1, 1, 0L, TimeUnit.MILLISECONDS, "kafka-apis-delete-topic-handle-executor", true, 1000),
+  val listOffsetHandleExecutor: ExecutorService = S3StreamThreadPoolMonitor.createAndMonitor(1, 1, 0L, TimeUnit.MILLISECONDS, "kafka-apis-list-offset-handle-executor", true, 1000)
 ) extends KafkaApis(requestChannel, metadataSupport, replicaManager, groupCoordinator, txnCoordinator,
   autoTopicCreationManager, brokerId, config, configRepository, metadataCache, metrics, authorizer, quotas,
   fetchManager, brokerTopicStats, clusterId, time, tokenManager, apiVersionManager, clientMetricsManager) {
@@ -131,7 +132,7 @@ class ElasticKafkaApis(
           response.asInstanceOf[DeleteTopicsResponse].data().responses().forEach(result => {
             if (result.errorCode() == Errors.NONE.code()) {
               if (!metadataCache.autoMQVersion().isTopicCleanupByControllerSupported) {
-                asyncHandleExecutor.submit(new Runnable {
+                deleteTopicHandleExecutor.submit(new Runnable {
                   override def run(): Unit = {
                     topicNameToPartitionEpochsMap.get(result.name()).foreach(partitionEpochs => {
                       ElasticLogManager.destroyLog(new TopicPartition(result.name(), partitionEpochs._1), result.topicId(), partitionEpochs._2)

diff --git a/core/src/main/scala/kafka/server/streamaspect/ElasticReplicaManager.scala b/core/src/main/scala/kafka/server/streamaspect/ElasticReplicaManager.scala
@@ -2,6 +2,7 @@ package kafka.server.streamaspect
 
 import com.automq.stream.api.exceptions.FastReadFailFastException
 import com.automq.stream.utils.FutureUtil
+import com.automq.stream.utils.threads.S3StreamThreadPoolMonitor
 import kafka.cluster.Partition
 import kafka.log.remote.RemoteLogManager
 import kafka.log.streamaspect.{ElasticLogManager, PartitionStatusTracker, ReadHint}
@@ -83,8 +84,8 @@ class ElasticReplicaManager(
   brokerEpochSupplier: () => Long = () => -1,
   addPartitionsToTxnManager: Option[AddPartitionsToTxnManager] = None,
   directoryEventHandler: DirectoryEventHandler = DirectoryEventHandler.NOOP,
-  private val fastFetchExecutor: ExecutorService = Executors.newFixedThreadPool(4, ThreadUtils.createThreadFactory("kafka-apis-fast-fetch-executor-%d", true)),
-  private val slowFetchExecutor: ExecutorService = Executors.newFixedThreadPool(12, ThreadUtils.createThreadFactory("kafka-apis-slow-fetch-executor-%d", true)),
+  private val fastFetchExecutor: ExecutorService = S3StreamThreadPoolMonitor.createAndMonitor(4, 4, 0L, TimeUnit.MILLISECONDS, "kafka-apis-fast-fetch-executor", true, 10000),
+  private val slowFetchExecutor: ExecutorService = S3StreamThreadPoolMonitor.createAndMonitor(12, 12, 0L, TimeUnit.MILLISECONDS, "kafka-apis-slow-fetch-executor", true, 10000),
   private val partitionMetricsCleanerExecutor: ScheduledExecutorService = Executors.newSingleThreadScheduledExecutor(ThreadUtils.createThreadFactory("kafka-partition-metrics-cleaner", true)),
 ) extends ReplicaManager(config, metrics, time, scheduler, logManager, remoteLogManager, quotaManagers, metadataCache,
   logDirFailureChannel, alterPartitionManager, brokerTopicStats, isShuttingDown, zkClient, delayedProducePurgatoryParam,

diff --git a/server/src/main/java/org/apache/kafka/network/SocketServerConfigs.java b/server/src/main/java/org/apache/kafka/network/SocketServerConfigs.java
@@ -158,6 +158,12 @@ public class SocketServerConfigs {
     public static final int QUEUED_MAX_REQUESTS_DEFAULT = 500;
     public static final String QUEUED_MAX_REQUESTS_DOC = "The number of queued requests allowed for data-plane, before blocking the network threads";
 
+    // AutoMQ inject start
+    public static final String QUEUED_MAX_REQUESTS_SIZE_BYTES_CONFIG = "queued.max.requests.size.bytes";
+    public static final int QUEUED_MAX_REQUESTS_SIZE_BYTES_DEFAULT = 100 * 1024 * 1024;
+    public static final String QUEUED_MAX_REQUESTS_SIZE_BYTES_DOC = "The number of queued requests size in total allowed for data-plane, before blocking the network threads";
+    // AutoMQ inject end
+
     public static final String QUEUED_MAX_BYTES_CONFIG = "queued.max.request.bytes";
     public static final int QUEUED_MAX_REQUEST_BYTES_DEFAULT = -1;
     public static final String QUEUED_MAX_REQUEST_BYTES_DOC = "The number of queued bytes allowed before no more requests are read";
@@ -182,6 +188,9 @@ public class SocketServerConfigs {
             .define(CONNECTIONS_MAX_IDLE_MS_CONFIG, LONG, CONNECTIONS_MAX_IDLE_MS_DEFAULT, MEDIUM, CONNECTIONS_MAX_IDLE_MS_DOC)
             .define(FAILED_AUTHENTICATION_DELAY_MS_CONFIG, INT, FAILED_AUTHENTICATION_DELAY_MS_DEFAULT, atLeast(0), LOW, FAILED_AUTHENTICATION_DELAY_MS_DOC)
             .define(QUEUED_MAX_REQUESTS_CONFIG, INT, QUEUED_MAX_REQUESTS_DEFAULT, atLeast(1), HIGH, QUEUED_MAX_REQUESTS_DOC)
+            // AutoMQ inject start
+            .define(QUEUED_MAX_REQUESTS_SIZE_BYTES_CONFIG, INT, QUEUED_MAX_REQUESTS_SIZE_BYTES_DEFAULT, atLeast(1024 * 1024), HIGH, QUEUED_MAX_REQUESTS_SIZE_BYTES_DOC)
+            // AutoMQ inject end
             .define(QUEUED_MAX_BYTES_CONFIG, LONG, QUEUED_MAX_REQUEST_BYTES_DEFAULT, MEDIUM, QUEUED_MAX_REQUEST_BYTES_DOC)
             .define(NUM_NETWORK_THREADS_CONFIG, INT, NUM_NETWORK_THREADS_DEFAULT, atLeast(1), HIGH, NUM_NETWORK_THREADS_DOC);
 }