hadoop-common-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From GitBox <...@apache.org>
Subject [GitHub] [hadoop] nandakumar131 commented on a change in pull request #620: HDDS-1205. Refactor ReplicationManager to handle QUASI_CLOSED contain…
Date Tue, 19 Mar 2019 20:01:11 GMT
nandakumar131 commented on a change in pull request #620: HDDS-1205. Refactor ReplicationManager
to handle QUASI_CLOSED contain…
URL: https://github.com/apache/hadoop/pull/620#discussion_r267072642
 
 

 ##########
 File path: hadoop-hdds/server-scm/src/main/java/org/apache/hadoop/hdds/scm/container/ReplicationManager.java
 ##########
 @@ -0,0 +1,686 @@
+/**
+ * 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
+ * <p>
+ * http://www.apache.org/licenses/LICENSE-2.0
+ * <p>
+ * 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.hadoop.hdds.scm.container;
+
+import com.google.common.annotations.VisibleForTesting;
+import org.apache.hadoop.conf.Configuration;
+import org.apache.hadoop.hdds.protocol.DatanodeDetails;
+import org.apache.hadoop.hdds.protocol.proto.HddsProtos.LifeCycleState;
+import org.apache.hadoop.hdds.protocol.proto
+    .StorageContainerDatanodeProtocolProtos.ContainerReplicaProto.State;
+import org.apache.hadoop.hdds.scm.ScmConfigKeys;
+import org.apache.hadoop.hdds.scm.container.placement.algorithms
+    .ContainerPlacementPolicy;
+import org.apache.hadoop.hdds.scm.events.SCMEvents;
+import org.apache.hadoop.hdds.server.events.EventPublisher;
+import org.apache.hadoop.ozone.lock.LockManager;
+import org.apache.hadoop.ozone.protocol.commands.CloseContainerCommand;
+import org.apache.hadoop.ozone.protocol.commands.CommandForDatanode;
+import org.apache.hadoop.ozone.protocol.commands.DeleteContainerCommand;
+import org.apache.hadoop.ozone.protocol.commands.ReplicateContainerCommand;
+import org.apache.hadoop.util.ExitUtil;
+import org.apache.hadoop.util.Time;
+import org.apache.ratis.util.Preconditions;
+import org.slf4j.Logger;
+import org.slf4j.LoggerFactory;
+
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.Collections;
+import java.util.HashMap;
+import java.util.Iterator;
+import java.util.LinkedHashMap;
+import java.util.List;
+import java.util.Map;
+import java.util.Set;
+import java.util.UUID;
+import java.util.concurrent.TimeUnit;
+import java.util.function.Predicate;
+import java.util.stream.Collectors;
+
+/**
+ * Replication Manager (RM) is the one which is responsible for making sure
+ * that the containers are properly replicated. Replication Manager deals only
+ * with Quasi Closed / Closed container.
+ */
+public class ReplicationManager {
+
+  private static final Logger LOG =
+      LoggerFactory.getLogger(ReplicationManager.class);
+
+  /**
+   * Reference to the ContainerManager.
+   */
+  private final ContainerManager containerManager;
+
+  /**
+   * PlacementPolicy which is used to identify where a container
+   * should be copied.
+   */
+  private final ContainerPlacementPolicy containerPlacement;
+
+  /**
+   * EventPublisher to fire Replicate and Delete container commands.
+   */
+  private final EventPublisher eventPublisher;
+
+  /**
+   * Used for locking a container with its ID while processing it.
+   */
+  private final LockManager<ContainerID> lockManager;
+
+  /**
+   * This is used to track container replication commands which are issued
+   * by ReplicationManager and not yet complete.
+   */
+  private final Map<ContainerID, List<InflightAction>> inflightReplication;
+
+  /**
+   * This is used to track container deletion commands which are issued
+   * by ReplicationManager and not yet complete.
+   */
+  private final Map<ContainerID, List<InflightAction>> inflightDeletion;
+
+  /**
+   * ReplicationMonitor thread is the one which wakes up at configured
+   * interval and processes all the containers.
+   */
+  private final Thread replicationMonitor;
+
+  /**
+   * The frequency in which ReplicationMonitor thread should run.
+   */
+  private final long interval;
+
+  /**
+   * Timeout for container replication & deletion command issued by
+   * ReplicationManager.
+   */
+  private final long eventTimeout;
+
+  /**
+   * Flag used to check if ReplicationMonitor thread is running or not.
+   */
+  private volatile boolean running;
+
+  /**
+   * Constructs ReplicationManager instance with the given configuration.
+   *
+   * @param conf OzoneConfiguration
+   * @param containerManager ContainerManager
+   * @param containerPlacement ContainerPlacementPolicy
+   * @param eventPublisher EventPublisher
+   */
+  public ReplicationManager(final Configuration conf,
+                            final ContainerManager containerManager,
+                            final ContainerPlacementPolicy containerPlacement,
+                            final EventPublisher eventPublisher) {
+    this.containerManager = containerManager;
+    this.containerPlacement = containerPlacement;
+    this.eventPublisher = eventPublisher;
+    this.lockManager = new LockManager<>(conf);
+    this.inflightReplication = new HashMap<>();
+    this.inflightDeletion = new HashMap<>();
+    this.replicationMonitor = new Thread(this::run);
+    this.replicationMonitor.setName("ReplicationMonitor");
+    this.interval = conf.getTimeDuration(
+        ScmConfigKeys.HDDS_SCM_REPLICATION_THREAD_INTERVAL,
+        ScmConfigKeys.HDDS_SCM_REPLICATION_THREAD_INTERVAL_DEFAULT,
+        TimeUnit.MILLISECONDS);
+    this.eventTimeout = conf.getTimeDuration(
+        ScmConfigKeys.HDDS_SCM_REPLICATION_EVENT_TIMEOUT,
+        ScmConfigKeys.HDDS_SCM_REPLICATION_EVENT_TIMEOUT_DEFAULT,
+        TimeUnit.MILLISECONDS);
+    this.running = false;
+  }
+
+  /**
+   * Starts Replication Monitor thread.
+   */
+  public void start() {
+    LOG.info("Starting Replication Monitor Thread.");
+    running = true;
+    replicationMonitor.start();
+  }
+
+  /**
+   * Process all the containers immediately.
+   */
+  @VisibleForTesting
+  void processContainersNow() {
+    containerManager.getContainerIDs()
+        .parallelStream().forEach(this::processContainer);
+  }
+
+  /**
+   * Stops Replication Monitor thread.
+   */
+  public void stop() {
+    LOG.info("Stopping Replication Monitor Thread.");
+    running = false;
+    replicationMonitor.interrupt();
+  }
+
+  /**
+   * ReplicationMonitor thread runnable. This wakes up at configured
+   * interval and processes all the containers in the system.
+   */
+  private void run() {
+    try {
+      while (running) {
+        try {
+          final long start = Time.monotonicNow();
+          final List<ContainerID> containerIds =
+              containerManager.getContainerIDs();
+          containerIds.parallelStream().forEach(this::processContainer);
+          LOG.debug("Replication Monitor Thread took {} milliseconds for" +
+                  " processing {} containers.", Time.monotonicNow() - start,
+              containerIds.size());
+          Thread.sleep(interval);
+        } catch (InterruptedException ex) {
+          // Wakeup and process the containers.
+          LOG.debug("Replication Monitor Thread got interrupt exception.");
+        }
+      }
+    } catch (Throwable t) {
+      // When we get runtime exception, we should terminate SCM.
+      LOG.error("Exception in Replication Monitor Thread.", t);
+      ExitUtil.terminate(1, t);
+    }
+  }
+
+  /**
+   * Process the given container.
+   *
+   * @param id ContainerID
+   */
+  private void processContainer(ContainerID id) {
+    lockManager.lock(id);
+    try {
+      final ContainerInfo container = containerManager.getContainer(id);
+      final Set<ContainerReplica> replicas = containerManager
+          .getContainerReplicas(container.containerID());
+      final LifeCycleState state = container.getState();
+
+      /*
+       * We don't take any action if the container is in OPEN state.
+       */
+      if (state == LifeCycleState.OPEN) {
+        return;
+      }
+
+      /*
+       * If the container is in CLOSING state, the replicas can either
+       * be in OPEN or in CLOSING state. In both of this cases
+       * we have to resend close container command to the datanodes.
+       */
+      if (state == LifeCycleState.CLOSING) {
+        replicas.forEach(replica -> sendCloseCommand(
+            container, replica.getDatanodeDetails(), false));
+        return;
+      }
+
+      /*
+       * Before processing the container we have to reconcile the
+       * inflightReplication and inflightDeletion actions.
+       *
+       * We remove the entry from inflightReplication and inflightDeletion
+       * list, if the operation is completed or if it has timed out.
+       */
+      updateInflightAction(container, inflightReplication,
+          action -> replicas.stream()
+              .anyMatch(r -> r.getDatanodeDetails().equals(action.datanode)));
+
+      updateInflightAction(container, inflightDeletion,
+          action -> replicas.stream()
+              .noneMatch(r -> r.getDatanodeDetails().equals(action.datanode)));
+
+
+      /*
+       * If the container is in QUASI_CLOSED state, check and close the
+       * container if possible.
+       */
+      if (state == LifeCycleState.QUASI_CLOSED) {
+        forceCloseContainer(container, replicas);
+      }
+
+      /*
+       * We don't have to take any action if the container is healthy.
+       *
+       * According to ReplicationMonitor container is considered healthy if
+       * the container is either in QUASI_CLOSED or in CLOSED state and has
+       * exact number of replicas in the same state.
+       */
+      if (isContainerHealthy(container, replicas)) {
+        return;
+      }
+
+      /*
+       * Check if the container if under replicated and take appropriate
+       * action.
+       */
+      if (isContainerUnderReplicated(container, replicas)) {
+        handleUnderReplicatedContainer(container, replicas);
+        return;
+      }
+
+      /*
+       * Check if the container if over replicated and take appropriate
+       * action.
+       */
+      if (isContainerOverReplicated(container, replicas)) {
+        handleOverReplicatedContainer(container, replicas);
+        return;
+      }
+
+      /*
+       * The container is neither under nor over replicated and the container
+       * is not healthy. This means that the container has unhealthy/corrupted
+       * replica.
+       */
+      handleInconsistentContainer(container, replicas);
+
+    } catch (ContainerNotFoundException ex) {
+      LOG.warn("Missing container {}.", id);
+    } finally {
+      lockManager.unlock(id);
+    }
+  }
+
+  /**
+   * Reconciles the InflightActions for a given container.
+   *
+   * @param container Container to update
+   * @param inflightActions inflightReplication (or) inflightDeletion
+   * @param filter filter to check if the operation is completed
+   */
+  private void updateInflightAction(final ContainerInfo container,
+      final Map<ContainerID, List<InflightAction>> inflightActions,
+      final Predicate<InflightAction> filter) {
+    final ContainerID id = container.containerID();
+    final long deadline = Time.monotonicNow() - eventTimeout;
+    if (inflightActions.containsKey(id)) {
+      final List<InflightAction> actions = inflightActions.get(id);
+      actions.removeIf(action -> action.time < deadline);
+      actions.removeIf(filter);
+      if (actions.isEmpty()) {
+        inflightActions.remove(id);
+      }
+    }
+  }
+
+  /**
+   * Returns true if the container is healthy according to ReplicationMonitor.
+   *
+   * According to ReplicationMonitor container is considered healthy if
+   * it has exact number of replicas in the same state as the container.
+   *
+   * @param container Container to check
+   * @param replicas Set of ContainerReplicas
+   * @return true if the container is healthy, false otherwise
+   */
+  private boolean isContainerHealthy(final ContainerInfo container,
+                                     final Set<ContainerReplica> replicas) {
+    return container.getReplicationFactor().getNumber() == replicas.size() &&
+        replicas.stream().allMatch(
+            r -> compareState(container.getState(), r.getState()));
+  }
+
+  /**
+   * Checks if the container is under replicated or not.
+   *
+   * @param container Container to check
+   * @param replicas Set of ContainerReplicas
+   * @return true if the container is under replicated, false otherwise
+   */
+  private boolean isContainerUnderReplicated(final ContainerInfo container,
+      final Set<ContainerReplica> replicas) {
+    return container.getReplicationFactor().getNumber() >
+        getReplicaCount(container.containerID(), replicas);
+  }
+
+  /**
+   * Checks if the container is over replicated or not.
+   *
+   * @param container Container to check
+   * @param replicas Set of ContainerReplicas
+   * @return true if the container if over replicated, false otherwise
+   */
+  private boolean isContainerOverReplicated(final ContainerInfo container,
+      final Set<ContainerReplica> replicas) {
+    return container.getReplicationFactor().getNumber() <
+        getReplicaCount(container.containerID(), replicas);
+  }
+
+  /**
+   * Returns the replication count of the given container. This also
+   * considers inflight replication and deletion.
+   *
+   * @param id ContainerID
+   * @param replicas Set of existing replicas
+   * @return number of estimated replicas for this container
+   */
+  private int getReplicaCount(final ContainerID id,
+                              final Set<ContainerReplica> replicas) {
+    return replicas.size()
+        + inflightReplication.getOrDefault(id, Collections.emptyList()).size()
+        - inflightDeletion.getOrDefault(id, Collections.emptyList()).size();
+  }
+
+  /**
+   * Force close the container replica(s) if possible.
+   *
+   * <p>
+   * If <50% of container replicas are in QUASI_CLOSED state and all
+   * the other replica are either in OPEN or CLOSING state, do nothing.
+   * We cannot identify the correct replica since we don't have quorum
+   * yet.
+   * </p>
+   *
+   * <p>
+   * If >50% (quorum) of replicas are in QUASI_CLOSED state, try to identify
+   * the latest container replica using originNodeId and sequenceId.
+   * Force close those replica(s) which have the latest sequenceId.
+   * </p>
+   *
+   * @param container ContainerInfo
+   * @param replicas Set of ContainerReplicas
+   */
+  private void forceCloseContainer(final ContainerInfo container,
+                                   final Set<ContainerReplica> replicas) {
+    Preconditions.assertTrue(container.getState() ==
+        LifeCycleState.QUASI_CLOSED);
+    final int replicationFactor = container.getReplicationFactor().getNumber();
+    final List<ContainerReplica> quasiClosedReplicas = replicas.stream()
+        .filter(r -> r.getState() == State.QUASI_CLOSED)
+        .collect(Collectors.toList());
+    final long uniqueQuasiClosedReplicaCount = quasiClosedReplicas
+        .stream()
+        .map(ContainerReplica::getOriginDatanodeId)
+        .distinct()
+        .count();
+    if (uniqueQuasiClosedReplicaCount > (replicationFactor / 2)) {
+      // Quorum of unique replica has been QUASI_CLOSED
+      final Long sequenceId = quasiClosedReplicas.stream()
+          .map(ContainerReplica::getSequenceId)
+          .max(Long::compare)
+          .orElse(-1L);
+      quasiClosedReplicas.stream()
+          .filter(r -> sequenceId != -1L)
+          .filter(replica -> replica.getSequenceId().equals(sequenceId))
+          .forEach(replica -> sendCloseCommand(
+              container, replica.getDatanodeDetails(), true));
+    }
+  }
+
+  /**
+   * If the given container is under replicated, identify a new set of
+   * datanode(s) to replicate the container using ContainerPlacementPolicy
+   * and send replicate container command to the identified datanode(s).
+   *
+   * @param container ContainerInfo
+   * @param replicas Set of ContainerReplicas
+   */
+  private void handleUnderReplicatedContainer(final ContainerInfo container,
+      final Set<ContainerReplica> replicas) {
+    try {
+      final ContainerID id = container.containerID();
+      final List<DatanodeDetails> deletionInFlight = inflightDeletion
+          .getOrDefault(id, Collections.emptyList())
+          .stream()
+          .map(action -> action.datanode)
+          .collect(Collectors.toList());
+      final List<DatanodeDetails> source = replicas.stream()
+          .filter(r ->
+              r.getState() == State.QUASI_CLOSED ||
+              r.getState() == State.CLOSED)
+          .filter(r -> !deletionInFlight.contains(r.getDatanodeDetails()))
+          .sorted((r1, r2) -> r2.getSequenceId().compareTo(r1.getSequenceId()))
+          .map(ContainerReplica::getDatanodeDetails)
+          .collect(Collectors.toList());
+      if (source.size() > 0) {
+        final int deficit = container.getReplicationFactor().getNumber() -
+            getReplicaCount(id, replicas);
+        final List<DatanodeDetails> selectedDatanodes = containerPlacement
+            .chooseDatanodes(source, deficit, container.getUsedBytes());
+        inflightReplication.computeIfAbsent(id, k -> new ArrayList<>());
+        for (DatanodeDetails datanode : selectedDatanodes) {
+          sendReplicateCommand(container, datanode, source);
+        }
+      } else {
+        LOG.warn("Cannot replicate container {}, no healthy replica found.",
+            container.containerID());
+      }
+    } catch (IOException ex) {
+      LOG.warn("Exception while replicating container {}.",
+          container.getContainerID(), ex);
+    }
+  }
+
+  /**
+   * If the given container is over replicated, identify the datanode(s)
+   * to delete the container and send delete container command to the
+   * identified datanode(s).
+   *
+   * @param container ContainerInfo
+   * @param replicas Set of ContainerReplicas
+   */
+  private void handleOverReplicatedContainer(final ContainerInfo container,
+      final Set<ContainerReplica> replicas) {
+
+    final ContainerID id = container.containerID();
+    final int replicationFactor = container.getReplicationFactor().getNumber();
+    // Dont consider inflight replication while calculating deficit here.
+    final int deficit = replicas.size() - replicationFactor -
 
 Review comment:
   Addressed

----------------------------------------------------------------
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.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services

---------------------------------------------------------------------
To unsubscribe, e-mail: common-issues-unsubscribe@hadoop.apache.org
For additional commands, e-mail: common-issues-help@hadoop.apache.org


Mime
View raw message