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; |
| 36 | import org.onosproject.event.AbstractListenerRegistry; |
| 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; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 52 | import java.util.concurrent.ExecutorService; |
| 53 | import java.util.concurrent.Executors; |
| 54 | import java.util.concurrent.Future; |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 55 | import java.util.stream.Collectors; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 56 | |
| 57 | import static com.google.common.base.Preconditions.checkArgument; |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 58 | import static org.onlab.util.Tools.groupedThreads; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 59 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 60 | /** |
| 61 | * Distributed implementation of LeadershipService that is based on Hazelcast. |
| 62 | * <p> |
| 63 | * The election is eventually-consistent: if there is Hazelcast partitioning, |
| 64 | * and the partitioning is healed, there could be a short window of time |
| 65 | * until the leaders in each partition discover each other. If this happens, |
| 66 | * the leaders release the leadership and run again for election. |
| 67 | * </p> |
| 68 | * <p> |
| 69 | * The leader election is based on Hazelcast's Global Lock, which is stongly |
| 70 | * consistent. In addition, each leader periodically advertises events |
| 71 | * (using a Hazelcast Topic) that it is the elected leader. Those events are |
| 72 | * used for two purposes: (1) Discover multi-leader collisions (in case of |
| 73 | * healed Hazelcast partitions), and (2) Inform all listeners who is |
| 74 | * the current leader (e.g., for informational purpose). |
| 75 | * </p> |
| 76 | */ |
Madan Jampani | 7f72c3f | 2015-03-01 17:34:59 -0800 | [diff] [blame] | 77 | @Component(immediate = true, enabled = false) |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 78 | @Service |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 79 | public class HazelcastLeadershipService implements LeadershipService { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 80 | private static final Logger log = |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 81 | LoggerFactory.getLogger(HazelcastLeadershipService.class); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 82 | |
| 83 | private static final KryoSerializer SERIALIZER = new KryoSerializer() { |
| 84 | @Override |
| 85 | protected void setupKryoPool() { |
| 86 | serializerPool = KryoNamespace.newBuilder() |
| 87 | .register(KryoNamespaces.API) |
| 88 | .build() |
| 89 | .populate(1); |
| 90 | } |
| 91 | }; |
| 92 | |
| 93 | private static final long LEADERSHIP_PERIODIC_INTERVAL_MS = 5 * 1000; // 5s |
| 94 | private static final long LEADERSHIP_REMOTE_TIMEOUT_MS = 15 * 1000; // 15s |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 95 | private static final String TOPIC_HZ_ID = "LeadershipService/AllTopics"; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 96 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 97 | // indicates there is no term value yet |
| 98 | private static final long NO_TERM = 0; |
| 99 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 100 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 101 | protected ClusterCommunicationService clusterCommunicator; |
| 102 | |
| 103 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 104 | protected ClusterService clusterService; |
| 105 | |
| 106 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 107 | protected StoreService storeService; |
| 108 | |
| 109 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 110 | protected EventDeliveryService eventDispatcher; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 111 | |
| 112 | private AbstractListenerRegistry<LeadershipEvent, LeadershipEventListener> |
| 113 | listenerRegistry; |
| 114 | private final Map<String, Topic> topics = Maps.newConcurrentMap(); |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 115 | private NodeId localNodeId; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 116 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 117 | private static final MessageSubject LEADERSHIP_EVENT_MESSAGE_SUBJECT = |
| 118 | new MessageSubject("hz-leadership-events"); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 119 | |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 120 | private ExecutorService messageHandlingExecutor; |
| 121 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 122 | @Activate |
| 123 | protected void activate() { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 124 | localNodeId = clusterService.getLocalNode().id(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 125 | listenerRegistry = new AbstractListenerRegistry<>(); |
| 126 | eventDispatcher.addSink(LeadershipEvent.class, listenerRegistry); |
| 127 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 128 | TopicConfig topicConfig = new TopicConfig(); |
| 129 | topicConfig.setGlobalOrderingEnabled(true); |
| 130 | topicConfig.setName(TOPIC_HZ_ID); |
| 131 | storeService.getHazelcastInstance().getConfig().addTopicConfig(topicConfig); |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 132 | |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 133 | messageHandlingExecutor = Executors.newSingleThreadExecutor( |
| 134 | groupedThreads("onos/store/leadership", "message-handler")); |
| 135 | |
| 136 | clusterCommunicator.addSubscriber( |
| 137 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 138 | new InternalLeadershipEventListener(), |
| 139 | messageHandlingExecutor); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 140 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 141 | log.info("Hazelcast Leadership Service started"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 142 | } |
| 143 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 144 | @Deactivate |
| 145 | protected void deactivate() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 146 | eventDispatcher.removeSink(LeadershipEvent.class); |
Madan Jampani | 2af244a | 2015-02-22 13:12:01 -0800 | [diff] [blame] | 147 | messageHandlingExecutor.shutdown(); |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 148 | clusterCommunicator.removeSubscriber(LEADERSHIP_EVENT_MESSAGE_SUBJECT); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 149 | |
| 150 | for (Topic topic : topics.values()) { |
| 151 | topic.stop(); |
| 152 | } |
| 153 | topics.clear(); |
| 154 | |
Jonathan Hart | 949c284 | 2014-11-28 23:44:09 -0800 | [diff] [blame] | 155 | log.info("Hazelcast Leadership Service stopped"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 156 | } |
| 157 | |
| 158 | @Override |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 159 | public NodeId getLeader(String path) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 160 | Topic topic = topics.get(path); |
| 161 | if (topic == null) { |
| 162 | return null; |
| 163 | } |
| 164 | return topic.leader(); |
| 165 | } |
| 166 | |
| 167 | @Override |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 168 | public Leadership getLeadership(String path) { |
| 169 | checkArgument(path != null); |
| 170 | Topic topic = topics.get(path); |
| 171 | if (topic != null) { |
| 172 | return new Leadership(topic.topicName(), |
| 173 | topic.leader(), |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 174 | topic.term(), |
| 175 | 0); |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 176 | } |
| 177 | return null; |
| 178 | } |
| 179 | |
| 180 | @Override |
| 181 | public Set<String> ownedTopics(NodeId nodeId) { |
| 182 | checkArgument(nodeId != null); |
| 183 | return topics.values() |
| 184 | .stream() |
| 185 | .filter(topic -> nodeId.equals(topic.leader())) |
| 186 | .map(topic -> topic.topicName) |
| 187 | .collect(Collectors.toSet()); |
| 188 | } |
| 189 | |
| 190 | @Override |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 191 | public void runForLeadership(String path) { |
| 192 | checkArgument(path != null); |
| 193 | Topic topic = new Topic(path); |
| 194 | Topic oldTopic = topics.putIfAbsent(path, topic); |
| 195 | if (oldTopic == null) { |
| 196 | topic.start(); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 197 | topic.runForLeadership(); |
| 198 | } else { |
| 199 | oldTopic.runForLeadership(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 200 | } |
| 201 | } |
| 202 | |
| 203 | @Override |
| 204 | public void withdraw(String path) { |
| 205 | checkArgument(path != null); |
| 206 | Topic topic = topics.get(path); |
| 207 | if (topic != null) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 208 | topics.remove(path, topic); |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 209 | topic.stop(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 210 | } |
| 211 | } |
| 212 | |
| 213 | @Override |
Yuta HIGUCHI | c2bf3d8 | 2014-11-28 18:50:41 -0800 | [diff] [blame] | 214 | public Map<String, Leadership> getLeaderBoard() { |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 215 | Map<String, Leadership> result = new HashMap<>(); |
| 216 | |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 217 | for (Topic topic : topics.values()) { |
| 218 | Leadership leadership = new Leadership(topic.topicName(), |
| 219 | topic.leader(), |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 220 | topic.term(), |
| 221 | 0); |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 222 | result.put(topic.topicName(), leadership); |
| 223 | } |
| 224 | return result; |
Yuta HIGUCHI | c2bf3d8 | 2014-11-28 18:50:41 -0800 | [diff] [blame] | 225 | } |
| 226 | |
| 227 | @Override |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 228 | public void addListener(LeadershipEventListener listener) { |
| 229 | listenerRegistry.addListener(listener); |
| 230 | } |
| 231 | |
| 232 | @Override |
| 233 | public void removeListener(LeadershipEventListener listener) { |
| 234 | listenerRegistry.removeListener(listener); |
| 235 | } |
| 236 | |
| 237 | /** |
| 238 | * Class for keeping per-topic information. |
| 239 | */ |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 240 | private final class Topic { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 241 | private final String topicName; |
| 242 | private volatile boolean isShutdown = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 243 | private volatile boolean isRunningForLeadership = false; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 244 | private volatile long lastLeadershipUpdateMs = 0; |
| 245 | private ExecutorService leaderElectionExecutor; |
| 246 | |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 247 | private volatile IAtomicLong term; |
| 248 | // This is local state, recording the term number for the last time |
| 249 | // this instance was leader for this topic. The current term could be |
| 250 | // higher if the mastership has changed any times. |
| 251 | private long myLastLeaderTerm = NO_TERM; |
| 252 | |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 253 | private volatile NodeId leader; |
| 254 | private ILock leaderLock; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 255 | private Future<?> getLockFuture; |
| 256 | private Future<?> periodicProcessingFuture; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 257 | |
| 258 | /** |
| 259 | * Constructor. |
| 260 | * |
| 261 | * @param topicName the topic name |
| 262 | */ |
| 263 | private Topic(String topicName) { |
| 264 | this.topicName = topicName; |
| 265 | } |
| 266 | |
| 267 | /** |
Pavlin Radoslavov | 187dff6 | 2014-12-02 07:21:35 -0800 | [diff] [blame] | 268 | * Gets the topic name. |
| 269 | * |
| 270 | * @return the topic name |
| 271 | */ |
| 272 | private String topicName() { |
| 273 | return topicName; |
| 274 | } |
| 275 | |
| 276 | /** |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 277 | * Gets the leader for the topic. |
| 278 | * |
| 279 | * @return the leader for the topic |
| 280 | */ |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 281 | private NodeId leader() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 282 | return leader; |
| 283 | } |
| 284 | |
| 285 | /** |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 286 | * Gets the current term for the topic. |
| 287 | * |
| 288 | * @return the term for the topic |
| 289 | */ |
| 290 | private long term() { |
| 291 | if (term == null) { |
| 292 | return NO_TERM; |
| 293 | } |
| 294 | return term.get(); |
| 295 | } |
| 296 | |
| 297 | /** |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 298 | * Starts operation. |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 299 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 300 | private synchronized void start() { |
| 301 | if (!isShutdown) { |
| 302 | // already running |
| 303 | return; |
| 304 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 305 | isShutdown = false; |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 306 | String threadPoolName = "election-" + topicName + "-%d"; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 307 | leaderElectionExecutor = Executors.newScheduledThreadPool(2, |
Thomas Vachuska | 6f94ded | 2015-02-21 14:02:38 -0800 | [diff] [blame] | 308 | groupedThreads("onos/leadership", threadPoolName)); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 309 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 310 | periodicProcessingFuture = |
| 311 | leaderElectionExecutor.submit(new Runnable() { |
| 312 | @Override |
| 313 | public void run() { |
| 314 | doPeriodicProcessing(); |
| 315 | } |
| 316 | }); |
| 317 | } |
| 318 | |
| 319 | /** |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 320 | * Runs for leadership. |
| 321 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 322 | private synchronized void runForLeadership() { |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 323 | if (isRunningForLeadership) { |
| 324 | return; // Nothing to do: already running |
| 325 | } |
| 326 | if (isShutdown) { |
| 327 | start(); |
| 328 | } |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 329 | isRunningForLeadership = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 330 | String lockHzId = "LeadershipService/" + topicName + "/lock"; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 331 | String termHzId = "LeadershipService/" + topicName + "/term"; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 332 | leaderLock = storeService.getHazelcastInstance().getLock(lockHzId); |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 333 | term = storeService.getHazelcastInstance().getAtomicLong(termHzId); |
| 334 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 335 | getLockFuture = leaderElectionExecutor.submit(new Runnable() { |
| 336 | @Override |
| 337 | public void run() { |
| 338 | doLeaderElectionThread(); |
| 339 | } |
| 340 | }); |
| 341 | } |
| 342 | |
| 343 | /** |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 344 | * Stops leadership election for the topic. |
| 345 | */ |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 346 | private synchronized void stop() { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 347 | isShutdown = true; |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 348 | isRunningForLeadership = false; |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 349 | // getLockFuture.cancel(true); |
| 350 | // periodicProcessingFuture.cancel(true); |
| 351 | leaderElectionExecutor.shutdownNow(); |
| 352 | } |
| 353 | |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 354 | /** |
| 355 | * Received a Leadership Event. |
| 356 | * |
| 357 | * @param leadershipEvent the received Leadership Event |
| 358 | */ |
| 359 | private void receivedLeadershipEvent(LeadershipEvent leadershipEvent) { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 360 | NodeId eventLeaderId = leadershipEvent.subject().leader(); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 361 | if (!leadershipEvent.subject().topic().equals(topicName)) { |
| 362 | return; // Not our topic: ignore |
| 363 | } |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 364 | if (eventLeaderId.equals(localNodeId)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 365 | return; // My own message: ignore |
| 366 | } |
| 367 | |
| 368 | synchronized (this) { |
| 369 | switch (leadershipEvent.type()) { |
| 370 | case LEADER_ELECTED: |
| 371 | // FALLTHROUGH |
| 372 | case LEADER_REELECTED: |
| 373 | // |
| 374 | // Another leader: if we are also a leader, then give up |
| 375 | // leadership and run for re-election. |
| 376 | // |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 377 | if ((leader != null) && leader.equals(localNodeId)) { |
| 378 | if (getLockFuture != null) { |
| 379 | getLockFuture.cancel(true); |
| 380 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 381 | } else { |
| 382 | // Just update the current leader |
| 383 | leader = leadershipEvent.subject().leader(); |
| 384 | lastLeadershipUpdateMs = System.currentTimeMillis(); |
| 385 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 386 | break; |
| 387 | case LEADER_BOOTED: |
| 388 | // Remove the state for the current leader |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 389 | if ((leader != null) && eventLeaderId.equals(leader)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 390 | leader = null; |
| 391 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 392 | break; |
| 393 | default: |
| 394 | break; |
| 395 | } |
| 396 | } |
| 397 | } |
| 398 | |
| 399 | private void doPeriodicProcessing() { |
| 400 | |
| 401 | while (!isShutdown) { |
| 402 | |
| 403 | // |
| 404 | // Periodic tasks: |
| 405 | // (a) Advertise ourselves as the leader |
| 406 | // OR |
| 407 | // (b) Expire a stale (remote) leader |
| 408 | // |
| 409 | synchronized (this) { |
| 410 | LeadershipEvent leadershipEvent; |
| 411 | if (leader != null) { |
Madan Jampani | 8d21c79 | 2014-12-01 16:31:07 -0800 | [diff] [blame] | 412 | if (leader.equals(localNodeId)) { |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 413 | // |
| 414 | // Advertise ourselves as the leader |
| 415 | // |
| 416 | leadershipEvent = new LeadershipEvent( |
| 417 | LeadershipEvent.Type.LEADER_REELECTED, |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 418 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Pavlin Radoslavov | bcb454c | 2014-12-02 11:47:47 -0800 | [diff] [blame] | 419 | // Dispatch to all instances |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 420 | |
| 421 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 422 | leadershipEvent, |
| 423 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 424 | SERIALIZER::encode); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 425 | } else { |
| 426 | // |
| 427 | // Test if time to expire a stale leader |
| 428 | // |
| 429 | long delta = System.currentTimeMillis() - |
| 430 | lastLeadershipUpdateMs; |
| 431 | if (delta > LEADERSHIP_REMOTE_TIMEOUT_MS) { |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 432 | log.debug("Topic {} leader {} booted due to heartbeat timeout", |
| 433 | topicName, leader); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 434 | leadershipEvent = new LeadershipEvent( |
| 435 | LeadershipEvent.Type.LEADER_BOOTED, |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 436 | new Leadership(topicName, leader, myLastLeaderTerm, 0)); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 437 | // Dispatch only to the local listener(s) |
| 438 | eventDispatcher.post(leadershipEvent); |
| 439 | leader = null; |
| 440 | } |
| 441 | } |
| 442 | } |
| 443 | } |
| 444 | |
| 445 | // Sleep before re-advertising |
| 446 | try { |
| 447 | Thread.sleep(LEADERSHIP_PERIODIC_INTERVAL_MS); |
| 448 | } catch (InterruptedException e) { |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 449 | log.debug("Leader Election periodic thread interrupted"); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 450 | } |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | /** |
| 455 | * Performs the leader election by using Hazelcast. |
| 456 | */ |
| 457 | private void doLeaderElectionThread() { |
| 458 | |
| 459 | while (!isShutdown) { |
| 460 | LeadershipEvent leadershipEvent; |
| 461 | // |
| 462 | // Try to acquire the lock and keep it until the instance is |
| 463 | // shutdown. |
| 464 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 465 | log.debug("Leader Election begin for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 466 | topicName); |
| 467 | try { |
| 468 | // Block until it becomes the leader |
| 469 | leaderLock.lockInterruptibly(); |
| 470 | } catch (InterruptedException e) { |
| 471 | // |
| 472 | // Thread interrupted. Either shutdown or run for |
| 473 | // re-election. |
| 474 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 475 | log.debug("Election interrupted for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 476 | topicName); |
| 477 | continue; |
| 478 | } |
| 479 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 480 | try { |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 481 | synchronized (this) { |
| 482 | // |
| 483 | // This instance is now the leader |
| 484 | // |
| 485 | log.info("Leader Elected for topic {}", topicName); |
| 486 | |
| 487 | updateTerm(); |
| 488 | |
| 489 | leader = localNodeId; |
| 490 | leadershipEvent = new LeadershipEvent( |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 491 | LeadershipEvent.Type.LEADER_ELECTED, |
| 492 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 493 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 494 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 495 | leadershipEvent, |
| 496 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 497 | SERIALIZER::encode); |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 498 | } |
| 499 | |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 500 | // Sleep forever until interrupted |
| 501 | Thread.sleep(Long.MAX_VALUE); |
| 502 | } catch (InterruptedException e) { |
| 503 | // |
| 504 | // Thread interrupted. Either shutdown or run for |
| 505 | // re-election. |
| 506 | // |
Pavlin Radoslavov | 2e57b3b | 2014-11-30 18:36:05 -0800 | [diff] [blame] | 507 | log.debug("Leader Interrupted for topic {}", |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 508 | topicName); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 509 | |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 510 | } finally { |
| 511 | synchronized (this) { |
| 512 | // If we reach here, we should release the leadership |
| 513 | log.debug("Leader Lock Released for topic {}", topicName); |
| 514 | if ((leader != null) && |
| 515 | leader.equals(localNodeId)) { |
| 516 | leader = null; |
| 517 | } |
| 518 | leadershipEvent = new LeadershipEvent( |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 519 | LeadershipEvent.Type.LEADER_BOOTED, |
| 520 | new Leadership(topicName, localNodeId, myLastLeaderTerm, 0)); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 521 | |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 522 | clusterCommunicator.broadcastIncludeSelf( |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 523 | leadershipEvent, |
| 524 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 525 | SERIALIZER::encode); |
| 526 | |
HIGUCHI Yuta | 26647a6 | 2015-02-27 10:57:59 -0800 | [diff] [blame] | 527 | if (leaderLock.isLockedByCurrentThread()) { |
| 528 | leaderLock.unlock(); |
| 529 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 530 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 531 | } |
| 532 | } |
Yuta HIGUCHI | 8f06827 | 2015-02-18 22:28:32 +0900 | [diff] [blame] | 533 | isRunningForLeadership = false; |
Jonathan Hart | 82b9fec | 2015-02-03 16:18:54 -0800 | [diff] [blame] | 534 | } |
| 535 | |
| 536 | // Globally guarded by the leadership lock for this term |
| 537 | // Locally guarded by synchronized (this) |
| 538 | private void updateTerm() { |
| 539 | long oldTerm = term.get(); |
| 540 | long newTerm = term.incrementAndGet(); |
| 541 | myLastLeaderTerm = newTerm; |
| 542 | log.debug("Topic {} updated term from {} to {}", topicName, |
| 543 | oldTerm, newTerm); |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 544 | } |
| 545 | } |
Madan Jampani | 97cf7c4 | 2015-02-19 16:33:37 -0800 | [diff] [blame] | 546 | |
| 547 | private class InternalLeadershipEventListener implements ClusterMessageHandler { |
| 548 | |
| 549 | @Override |
| 550 | public void handle(ClusterMessage message) { |
| 551 | LeadershipEvent leadershipEvent = |
| 552 | SERIALIZER.decode(message.payload()); |
| 553 | |
| 554 | log.trace("Leadership Event: time = {} type = {} event = {}", |
| 555 | leadershipEvent.time(), leadershipEvent.type(), |
| 556 | leadershipEvent); |
| 557 | // |
| 558 | // If there is no entry for the topic, then create a new one to |
| 559 | // keep track of the leadership, but don't run for leadership itself. |
| 560 | // |
| 561 | String topicName = leadershipEvent.subject().topic(); |
| 562 | Topic topic = topics.get(topicName); |
| 563 | if (topic == null) { |
| 564 | topic = new Topic(topicName); |
| 565 | Topic oldTopic = topics.putIfAbsent(topicName, topic); |
| 566 | if (oldTopic == null) { |
| 567 | // encountered new topic, start periodic processing |
| 568 | topic.start(); |
| 569 | } else { |
| 570 | topic = oldTopic; |
| 571 | } |
| 572 | } |
| 573 | topic.receivedLeadershipEvent(leadershipEvent); |
| 574 | eventDispatcher.post(leadershipEvent); |
| 575 | } |
| 576 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 577 | |
| 578 | @Override |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 579 | public Map<String, Leadership> getCandidates() { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 580 | return null; |
| 581 | } |
| 582 | |
| 583 | @Override |
| 584 | public List<NodeId> getCandidates(String path) { |
| 585 | return null; |
| 586 | } |
Pavlin Radoslavov | c91eebe | 2014-11-25 18:45:46 -0800 | [diff] [blame] | 587 | } |