Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 1 | /* |
Ray Milkey | 34c9590 | 2015-04-15 09:47:53 -0700 | [diff] [blame] | 2 | * Copyright 2014-2015 Open Networking Laboratory |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 3 | * |
| 4 | * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | * you may not use this file except in compliance with the License. |
| 6 | * You may obtain a copy of the License at |
| 7 | * |
| 8 | * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | * |
| 10 | * Unless required by applicable law or agreed to in writing, software |
| 11 | * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | * See the License for the specific language governing permissions and |
| 14 | * limitations under the License. |
| 15 | */ |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 16 | package org.onosproject.store.cluster.impl; |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 17 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 18 | import com.google.common.collect.Maps; |
| 19 | import com.hazelcast.config.TopicConfig; |
| 20 | import com.hazelcast.core.IAtomicLong; |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 21 | import com.hazelcast.core.ILock; |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 22 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 23 | import org.apache.felix.scr.annotations.Activate; |
| 24 | import org.apache.felix.scr.annotations.Component; |
| 25 | import org.apache.felix.scr.annotations.Deactivate; |
| 26 | import org.apache.felix.scr.annotations.Reference; |
| 27 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
| 28 | import org.apache.felix.scr.annotations.Service; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 29 | import org.onlab.util.KryoNamespace; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 30 | import org.onosproject.cluster.ClusterService; |
| 31 | import org.onosproject.cluster.Leadership; |
| 32 | import org.onosproject.cluster.LeadershipEvent; |
| 33 | import org.onosproject.cluster.LeadershipEventListener; |
| 34 | import org.onosproject.cluster.LeadershipService; |
| 35 | import org.onosproject.cluster.NodeId; |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 36 | import org.onosproject.event.ListenerRegistry; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 37 | import org.onosproject.event.EventDeliveryService; |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 38 | import org.onosproject.store.cluster.messaging.ClusterCommunicationService; |
| 39 | import org.onosproject.store.cluster.messaging.ClusterMessage; |
| 40 | import org.onosproject.store.cluster.messaging.ClusterMessageHandler; |
| 41 | import org.onosproject.store.cluster.messaging.MessageSubject; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 42 | import org.onosproject.store.hz.StoreService; |
| 43 | import org.onosproject.store.serializers.KryoNamespaces; |
| 44 | import org.onosproject.store.serializers.KryoSerializer; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 45 | import org.slf4j.Logger; |
| 46 | import org.slf4j.LoggerFactory; |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 47 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 48 | import java.util.HashMap; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 49 | import java.util.List; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 50 | import java.util.Map; |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 51 | import java.util.Set; |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 52 | import java.util.concurrent.CompletableFuture; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 53 | import java.util.concurrent.ExecutorService; |
| 54 | import java.util.concurrent.Executors; |
| 55 | import java.util.concurrent.Future; |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 56 | import java.util.stream.Collectors; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 57 | |
| 58 | import static com.google.common.base.Preconditions.checkArgument; |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 59 | import static org.onlab.util.Tools.groupedThreads; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 60 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 61 | /** |
| 62 | * Distributed implementation of LeadershipService that is based on Hazelcast. |
| 63 | * <p> |
| 64 | * The election is eventually-consistent: if there is Hazelcast partitioning, |
| 65 | * and the partitioning is healed, there could be a short window of time |
| 66 | * until the leaders in each partition discover each other. If this happens, |
| 67 | * the leaders release the leadership and run again for election. |
| 68 | * </p> |
| 69 | * <p> |
| 70 | * The leader election is based on Hazelcast's Global Lock, which is stongly |
| 71 | * consistent. In addition, each leader periodically advertises events |
| 72 | * (using a Hazelcast Topic) that it is the elected leader. Those events are |
| 73 | * used for two purposes: (1) Discover multi-leader collisions (in case of |
| 74 | * healed Hazelcast partitions), and (2) Inform all listeners who is |
| 75 | * the current leader (e.g., for informational purpose). |
| 76 | * </p> |
| 77 | */ |
Madan Jampani | 7f72c3f | 2015-03-01 17:34:59 -0800 | [diff] [blame] | 78 | @Component(immediate = true, enabled = false) |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 79 | @Service |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 80 | public class HazelcastLeadershipService implements LeadershipService { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 81 | private static final Logger log = |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 82 | LoggerFactory.getLogger(HazelcastLeadershipService.class); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 83 | |
| 84 | private static final KryoSerializer SERIALIZER = new KryoSerializer() { |
| 85 | @Override |
| 86 | protected void setupKryoPool() { |
| 87 | serializerPool = KryoNamespace.newBuilder() |
| 88 | .register(KryoNamespaces.API) |
| 89 | .build() |
| 90 | .populate(1); |
| 91 | } |
| 92 | }; |
| 93 | |
| 94 | private static final long LEADERSHIP_PERIODIC_INTERVAL_MS = 5 * 1000; // 5s |
| 95 | private static final long LEADERSHIP_REMOTE_TIMEOUT_MS = 15 * 1000; // 15s |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 96 | private static final String TOPIC_HZ_ID = "LeadershipService/AllTopics"; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 97 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 98 | // indicates there is no term value yet |
| 99 | private static final long NO_TERM = 0; |
| 100 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 101 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 102 | protected ClusterCommunicationService clusterCommunicator; |
| 103 | |
| 104 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 105 | protected ClusterService clusterService; |
| 106 | |
| 107 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 108 | protected StoreService storeService; |
| 109 | |
| 110 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 111 | protected EventDeliveryService eventDispatcher; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 112 | |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 113 | private ListenerRegistry<LeadershipEvent, LeadershipEventListener> |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 114 | listenerRegistry; |
| 115 | private final Map<String, Topic> topics = Maps.newConcurrentMap(); |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 116 | private NodeId localNodeId; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 117 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 118 | private static final MessageSubject LEADERSHIP_EVENT_MESSAGE_SUBJECT = |
| 119 | new MessageSubject("hz-leadership-events"); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 120 | |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 121 | private ExecutorService messageHandlingExecutor; |
| 122 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 123 | @Activate |
| 124 | protected void activate() { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 125 | localNodeId = clusterService.getLocalNode().id(); |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 126 | listenerRegistry = new ListenerRegistry<>(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 127 | eventDispatcher.addSink(LeadershipEvent.class, listenerRegistry); |
| 128 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 129 | TopicConfig topicConfig = new TopicConfig(); |
| 130 | topicConfig.setGlobalOrderingEnabled(true); |
| 131 | topicConfig.setName(TOPIC_HZ_ID); |
| 132 | storeService.getHazelcastInstance().getConfig().addTopicConfig(topicConfig); |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 133 | |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 134 | messageHandlingExecutor = Executors.newSingleThreadExecutor( |
| 135 | groupedThreads("onos/store/leadership", "message-handler")); |
| 136 | |
| 137 | clusterCommunicator.addSubscriber( |
| 138 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 139 | new InternalLeadershipEventListener(), |
| 140 | messageHandlingExecutor); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 141 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 142 | log.info("Hazelcast Leadership Service started"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 143 | } |
| 144 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 145 | @Deactivate |
| 146 | protected void deactivate() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 147 | eventDispatcher.removeSink(LeadershipEvent.class); |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 148 | messageHandlingExecutor.shutdown(); |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 149 | clusterCommunicator.removeSubscriber(LEADERSHIP_EVENT_MESSAGE_SUBJECT); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 150 | |
| 151 | for (Topic topic : topics.values()) { |
| 152 | topic.stop(); |
| 153 | } |
| 154 | topics.clear(); |
| 155 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 156 | log.info("Hazelcast Leadership Service stopped"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 157 | } |
| 158 | |
| 159 | @Override |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 160 | public NodeId getLeader(String path) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 161 | Topic topic = topics.get(path); |
| 162 | if (topic == null) { |
| 163 | return null; |
| 164 | } |
| 165 | return topic.leader(); |
| 166 | } |
| 167 | |
| 168 | @Override |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 169 | public Leadership getLeadership(String path) { |
| 170 | checkArgument(path != null); |
| 171 | Topic topic = topics.get(path); |
| 172 | if (topic != null) { |
| 173 | return new Leadership(topic.topicName(), |
| 174 | topic.leader(), |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 175 | topic.term(), |
| 176 | 0); |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 177 | } |
| 178 | return null; |
| 179 | } |
| 180 | |
| 181 | @Override |
| 182 | public Set<String> ownedTopics(NodeId nodeId) { |
| 183 | checkArgument(nodeId != null); |
| 184 | return topics.values() |
| 185 | .stream() |
| 186 | .filter(topic -> nodeId.equals(topic.leader())) |
| 187 | .map(topic -> topic.topicName) |
| 188 | .collect(Collectors.toSet()); |
| 189 | } |
| 190 | |
| 191 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 192 | public CompletableFuture<Leadership> runForLeadership(String path) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 193 | checkArgument(path != null); |
| 194 | Topic topic = new Topic(path); |
| 195 | Topic oldTopic = topics.putIfAbsent(path, topic); |
| 196 | if (oldTopic == null) { |
| 197 | topic.start(); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 198 | topic.runForLeadership(); |
| 199 | } else { |
| 200 | oldTopic.runForLeadership(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 201 | } |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 202 | return CompletableFuture.completedFuture(getLeadership(path)); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 203 | } |
| 204 | |
| 205 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 206 | public CompletableFuture<Void> withdraw(String path) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 207 | checkArgument(path != null); |
| 208 | Topic topic = topics.get(path); |
| 209 | if (topic != null) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 210 | topics.remove(path, topic); |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 211 | topic.stop(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 212 | } |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 213 | return CompletableFuture.completedFuture(null); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 214 | } |
| 215 | |
| 216 | @Override |
Yuta HIGUCHI | c2bf3d8 | 2014-11-28 18:50:41 -0800 | [diff] [blame] | 217 | public Map<String, Leadership> getLeaderBoard() { |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 218 | Map<String, Leadership> result = new HashMap<>(); |
| 219 | |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 220 | for (Topic topic : topics.values()) { |
| 221 | Leadership leadership = new Leadership(topic.topicName(), |
| 222 | topic.leader(), |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 223 | topic.term(), |
| 224 | 0); |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 225 | result.put(topic.topicName(), leadership); |
| 226 | } |
| 227 | return result; |
Yuta HIGUCHI | c2bf3d8 | 2014-11-28 18:50:41 -0800 | [diff] [blame] | 228 | } |
| 229 | |
| 230 | @Override |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 231 | public void addListener(LeadershipEventListener listener) { |
| 232 | listenerRegistry.addListener(listener); |
| 233 | } |
| 234 | |
| 235 | @Override |
| 236 | public void removeListener(LeadershipEventListener listener) { |
| 237 | listenerRegistry.removeListener(listener); |
| 238 | } |
| 239 | |
| 240 | /** |
| 241 | * Class for keeping per-topic information. |
| 242 | */ |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 243 | private final class Topic { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 244 | private final String topicName; |
| 245 | private volatile boolean isShutdown = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 246 | private volatile boolean isRunningForLeadership = false; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 247 | private volatile long lastLeadershipUpdateMs = 0; |
| 248 | private ExecutorService leaderElectionExecutor; |
| 249 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 250 | private volatile IAtomicLong term; |
| 251 | // This is local state, recording the term number for the last time |
| 252 | // this instance was leader for this topic. The current term could be |
| 253 | // higher if the mastership has changed any times. |
| 254 | private long myLastLeaderTerm = NO_TERM; |
| 255 | |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 256 | private volatile NodeId leader; |
| 257 | private ILock leaderLock; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 258 | private Future<?> getLockFuture; |
| 259 | private Future<?> periodicProcessingFuture; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 260 | |
| 261 | /** |
| 262 | * Constructor. |
| 263 | * |
| 264 | * @param topicName the topic name |
| 265 | */ |
| 266 | private Topic(String topicName) { |
| 267 | this.topicName = topicName; |
| 268 | } |
| 269 | |
| 270 | /** |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 271 | * Gets the topic name. |
| 272 | * |
| 273 | * @return the topic name |
| 274 | */ |
| 275 | private String topicName() { |
| 276 | return topicName; |
| 277 | } |
| 278 | |
| 279 | /** |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 280 | * Gets the leader for the topic. |
| 281 | * |
| 282 | * @return the leader for the topic |
| 283 | */ |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 284 | private NodeId leader() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 285 | return leader; |
| 286 | } |
| 287 | |
| 288 | /** |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 289 | * Gets the current term for the topic. |
| 290 | * |
| 291 | * @return the term for the topic |
| 292 | */ |
| 293 | private long term() { |
| 294 | if (term == null) { |
| 295 | return NO_TERM; |
| 296 | } |
| 297 | return term.get(); |
| 298 | } |
| 299 | |
| 300 | /** |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 301 | * Starts operation. |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 302 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 303 | private synchronized void start() { |
| 304 | if (!isShutdown) { |
| 305 | // already running |
| 306 | return; |
| 307 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 308 | isShutdown = false; |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 309 | String threadPoolName = "election-" + topicName + "-%d"; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 310 | leaderElectionExecutor = Executors.newScheduledThreadPool(2, |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 311 | groupedThreads("onos/leadership", threadPoolName)); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 312 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 313 | periodicProcessingFuture = |
| 314 | leaderElectionExecutor.submit(new Runnable() { |
| 315 | @Override |
| 316 | public void run() { |
| 317 | doPeriodicProcessing(); |
| 318 | } |
| 319 | }); |
| 320 | } |
| 321 | |
| 322 | /** |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 323 | * Runs for leadership. |
| 324 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 325 | private synchronized void runForLeadership() { |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 326 | if (isRunningForLeadership) { |
| 327 | return; // Nothing to do: already running |
| 328 | } |
| 329 | if (isShutdown) { |
| 330 | start(); |
| 331 | } |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 332 | isRunningForLeadership = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 333 | String lockHzId = "LeadershipService/" + topicName + "/lock"; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 334 | String termHzId = "LeadershipService/" + topicName + "/term"; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 335 | leaderLock = storeService.getHazelcastInstance().getLock(lockHzId); |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 336 | term = storeService.getHazelcastInstance().getAtomicLong(termHzId); |
| 337 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 338 | getLockFuture = leaderElectionExecutor.submit(new Runnable() { |
| 339 | @Override |
| 340 | public void run() { |
| 341 | doLeaderElectionThread(); |
| 342 | } |
| 343 | }); |
| 344 | } |
| 345 | |
| 346 | /** |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 347 | * Stops leadership election for the topic. |
| 348 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 349 | private synchronized void stop() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 350 | isShutdown = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 351 | isRunningForLeadership = false; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 352 | // getLockFuture.cancel(true); |
| 353 | // periodicProcessingFuture.cancel(true); |
| 354 | leaderElectionExecutor.shutdownNow(); |
| 355 | } |
| 356 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 357 | /** |
| 358 | * Received a Leadership Event. |
| 359 | * |
| 360 | * @param leadershipEvent the received Leadership Event |
| 361 | */ |
| 362 | private void receivedLeadershipEvent(LeadershipEvent leadershipEvent) { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 363 | NodeId eventLeaderId = leadershipEvent.subject().leader(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 364 | if (!leadershipEvent.subject().topic().equals(topicName)) { |
| 365 | return; // Not our topic: ignore |
| 366 | } |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 367 | if (eventLeaderId.equals(localNodeId)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 368 | return; // My own message: ignore |
| 369 | } |
| 370 | |
| 371 | synchronized (this) { |
| 372 | switch (leadershipEvent.type()) { |
| 373 | case LEADER_ELECTED: |
| 374 | // FALLTHROUGH |
| 375 | case LEADER_REELECTED: |
| 376 | // |
| 377 | // Another leader: if we are also a leader, then give up |
| 378 | // leadership and run for re-election. |
| 379 | // |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 380 | if ((leader != null) && leader.equals(localNodeId)) { |
| 381 | if (getLockFuture != null) { |
| 382 | getLockFuture.cancel(true); |
| 383 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 384 | } else { |
| 385 | // Just update the current leader |
| 386 | leader = leadershipEvent.subject().leader(); |
| 387 | lastLeadershipUpdateMs = System.currentTimeMillis(); |
| 388 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 389 | break; |
| 390 | case LEADER_BOOTED: |
| 391 | // Remove the state for the current leader |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 392 | if ((leader != null) && eventLeaderId.equals(leader)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 393 | leader = null; |
| 394 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 395 | break; |
| 396 | default: |
| 397 | break; |
| 398 | } |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | private void doPeriodicProcessing() { |
| 403 | |
| 404 | while (!isShutdown) { |
| 405 | |
| 406 | // |
| 407 | // Periodic tasks: |
| 408 | // (a) Advertise ourselves as the leader |
| 409 | // OR |
| 410 | // (b) Expire a stale (remote) leader |
| 411 | // |
| 412 | synchronized (this) { |
| 413 | LeadershipEvent leadershipEvent; |
| 414 | if (leader != null) { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 415 | if (leader.equals(localNodeId)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 416 | // |
| 417 | // Advertise ourselves as the leader |
| 418 | // |
| 419 | leadershipEvent = new LeadershipEvent( |
| 420 | LeadershipEvent.Type.LEADER_REELECTED, |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 421 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 422 | // Dispatch to all instances |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 423 | |
| 424 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 425 | leadershipEvent, |
| 426 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 427 | SERIALIZER::encode); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 428 | } else { |
| 429 | // |
| 430 | // Test if time to expire a stale leader |
| 431 | // |
| 432 | long delta = System.currentTimeMillis() - |
| 433 | lastLeadershipUpdateMs; |
| 434 | if (delta > LEADERSHIP_REMOTE_TIMEOUT_MS) { |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 435 | log.debug("Topic {} leader {} booted due to heartbeat timeout", |
| 436 | topicName, leader); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 437 | leadershipEvent = new LeadershipEvent( |
| 438 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 439 | new Leadership(topicName, leader, myLastLeaderTerm, 0)); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 440 | // Dispatch only to the local listener(s) |
| 441 | eventDispatcher.post(leadershipEvent); |
| 442 | leader = null; |
| 443 | } |
| 444 | } |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | // Sleep before re-advertising |
| 449 | try { |
| 450 | Thread.sleep(LEADERSHIP_PERIODIC_INTERVAL_MS); |
| 451 | } catch (InterruptedException e) { |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 452 | log.debug("Leader Election periodic thread interrupted"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 453 | } |
| 454 | } |
| 455 | } |
| 456 | |
| 457 | /** |
| 458 | * Performs the leader election by using Hazelcast. |
| 459 | */ |
| 460 | private void doLeaderElectionThread() { |
| 461 | |
| 462 | while (!isShutdown) { |
| 463 | LeadershipEvent leadershipEvent; |
| 464 | // |
| 465 | // Try to acquire the lock and keep it until the instance is |
| 466 | // shutdown. |
| 467 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 468 | log.debug("Leader Election begin for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 469 | topicName); |
| 470 | try { |
| 471 | // Block until it becomes the leader |
| 472 | leaderLock.lockInterruptibly(); |
| 473 | } catch (InterruptedException e) { |
| 474 | // |
| 475 | // Thread interrupted. Either shutdown or run for |
| 476 | // re-election. |
| 477 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 478 | log.debug("Election interrupted for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 479 | topicName); |
| 480 | continue; |
| 481 | } |
| 482 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 483 | try { |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 484 | synchronized (this) { |
| 485 | // |
| 486 | // This instance is now the leader |
| 487 | // |
| 488 | log.info("Leader Elected for topic {}", topicName); |
| 489 | |
| 490 | updateTerm(); |
| 491 | |
| 492 | leader = localNodeId; |
| 493 | leadershipEvent = new LeadershipEvent( |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 494 | LeadershipEvent.Type.LEADER_ELECTED, |
| 495 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 496 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 497 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 498 | leadershipEvent, |
| 499 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 500 | SERIALIZER::encode); |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 501 | } |
| 502 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 503 | // Sleep forever until interrupted |
| 504 | Thread.sleep(Long.MAX_VALUE); |
| 505 | } catch (InterruptedException e) { |
| 506 | // |
| 507 | // Thread interrupted. Either shutdown or run for |
| 508 | // re-election. |
| 509 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 510 | log.debug("Leader Interrupted for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 511 | topicName); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 512 | |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 513 | } finally { |
| 514 | synchronized (this) { |
| 515 | // If we reach here, we should release the leadership |
| 516 | log.debug("Leader Lock Released for topic {}", topicName); |
| 517 | if ((leader != null) && |
| 518 | leader.equals(localNodeId)) { |
| 519 | leader = null; |
| 520 | } |
| 521 | leadershipEvent = new LeadershipEvent( |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 522 | LeadershipEvent.Type.LEADER_BOOTED, |
| 523 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 524 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 525 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 526 | leadershipEvent, |
| 527 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 528 | SERIALIZER::encode); |
| 529 | |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 530 | if (leaderLock.isLockedByCurrentThread()) { |
| 531 | leaderLock.unlock(); |
| 532 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 533 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 534 | } |
| 535 | } |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 536 | isRunningForLeadership = false; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 537 | } |
| 538 | |
| 539 | // Globally guarded by the leadership lock for this term |
| 540 | // Locally guarded by synchronized (this) |
| 541 | private void updateTerm() { |
| 542 | long oldTerm = term.get(); |
| 543 | long newTerm = term.incrementAndGet(); |
| 544 | myLastLeaderTerm = newTerm; |
| 545 | log.debug("Topic {} updated term from {} to {}", topicName, |
| 546 | oldTerm, newTerm); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 547 | } |
| 548 | } |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 549 | |
| 550 | private class InternalLeadershipEventListener implements ClusterMessageHandler { |
| 551 | |
| 552 | @Override |
| 553 | public void handle(ClusterMessage message) { |
| 554 | LeadershipEvent leadershipEvent = |
| 555 | SERIALIZER.decode(message.payload()); |
| 556 | |
| 557 | log.trace("Leadership Event: time = {} type = {} event = {}", |
| 558 | leadershipEvent.time(), leadershipEvent.type(), |
| 559 | leadershipEvent); |
| 560 | // |
| 561 | // If there is no entry for the topic, then create a new one to |
| 562 | // keep track of the leadership, but don't run for leadership itself. |
| 563 | // |
| 564 | String topicName = leadershipEvent.subject().topic(); |
| 565 | Topic topic = topics.get(topicName); |
| 566 | if (topic == null) { |
| 567 | topic = new Topic(topicName); |
| 568 | Topic oldTopic = topics.putIfAbsent(topicName, topic); |
| 569 | if (oldTopic == null) { |
| 570 | // encountered new topic, start periodic processing |
| 571 | topic.start(); |
| 572 | } else { |
| 573 | topic = oldTopic; |
| 574 | } |
| 575 | } |
| 576 | topic.receivedLeadershipEvent(leadershipEvent); |
| 577 | eventDispatcher.post(leadershipEvent); |
| 578 | } |
| 579 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 580 | |
| 581 | @Override |
Madan Jampani | fd45d5e | 2015-04-20 13:33:21 -0700 | [diff] [blame] | 582 | public Map<String, List<NodeId>> getCandidates() { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 583 | return null; |
| 584 | } |
| 585 | |
| 586 | @Override |
| 587 | public List<NodeId> getCandidates(String path) { |
| 588 | return null; |
| 589 | } |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 590 | |
| 591 | @Override |
| 592 | public boolean stepdown(String path) { |
| 593 | throw new UnsupportedOperationException(); |
| 594 | } |
| 595 | |
| 596 | @Override |
| 597 | public boolean makeTopCandidate(String path, NodeId nodeId) { |
| 598 | throw new UnsupportedOperationException(); |
| 599 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 600 | } |