Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 1 | package org.onosproject.store.consistent.impl; |
| 2 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 3 | import com.google.common.collect.ImmutableList; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 4 | import com.google.common.collect.ImmutableMap; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 5 | import com.google.common.collect.Lists; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 6 | import com.google.common.collect.Maps; |
| 7 | import com.google.common.collect.Sets; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 8 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 9 | import org.apache.felix.scr.annotations.Activate; |
| 10 | import org.apache.felix.scr.annotations.Component; |
| 11 | import org.apache.felix.scr.annotations.Deactivate; |
| 12 | import org.apache.felix.scr.annotations.Reference; |
| 13 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
| 14 | import org.apache.felix.scr.annotations.Service; |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 15 | import org.apache.commons.lang3.mutable.MutableBoolean; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 16 | import org.onlab.util.KryoNamespace; |
| 17 | import org.onosproject.cluster.ClusterService; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 18 | import org.onosproject.cluster.Leadership; |
| 19 | import org.onosproject.cluster.LeadershipEvent; |
| 20 | import org.onosproject.cluster.LeadershipEventListener; |
| 21 | import org.onosproject.cluster.LeadershipService; |
| 22 | import org.onosproject.cluster.NodeId; |
| 23 | import org.onosproject.event.AbstractListenerRegistry; |
| 24 | import org.onosproject.event.EventDeliveryService; |
| 25 | import org.onosproject.store.cluster.messaging.ClusterCommunicationService; |
| 26 | import org.onosproject.store.cluster.messaging.ClusterMessage; |
| 27 | import org.onosproject.store.cluster.messaging.ClusterMessageHandler; |
| 28 | import org.onosproject.store.cluster.messaging.MessageSubject; |
| 29 | import org.onosproject.store.serializers.KryoNamespaces; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 30 | import org.onosproject.store.service.ConsistentMap; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 31 | import org.onosproject.store.service.ConsistentMapException; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 32 | import org.onosproject.store.service.Serializer; |
| 33 | import org.onosproject.store.service.StorageService; |
| 34 | import org.onosproject.store.service.Versioned; |
| 35 | import org.slf4j.Logger; |
| 36 | |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 37 | import java.util.Map; |
| 38 | import java.util.Map.Entry; |
| 39 | import java.util.Objects; |
| 40 | import java.util.Set; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 41 | import java.util.List; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 42 | import java.util.concurrent.ExecutorService; |
| 43 | import java.util.concurrent.Executors; |
| 44 | import java.util.concurrent.ScheduledExecutorService; |
| 45 | import java.util.concurrent.TimeUnit; |
| 46 | import java.util.stream.Collectors; |
| 47 | |
| 48 | import static com.google.common.base.Preconditions.checkArgument; |
| 49 | import static org.onlab.util.Tools.groupedThreads; |
| 50 | import static org.slf4j.LoggerFactory.getLogger; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 51 | |
Ayaka Koshibe | 4a3c239 | 2015-04-22 18:03:13 -0700 | [diff] [blame] | 52 | import static org.onosproject.cluster.ControllerNode.State.ACTIVE; |
| 53 | import static org.onosproject.cluster.ControllerNode.State.INACTIVE; |
| 54 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 55 | /** |
| 56 | * Distributed Lock Manager implemented on top of ConsistentMap. |
| 57 | * <p> |
| 58 | * This implementation makes use of cluster manager's failure |
| 59 | * detection capabilities to detect and purge stale locks. |
| 60 | * TODO: Ensure lock safety and liveness. |
| 61 | */ |
Madan Jampani | 7f72c3f | 2015-03-01 17:34:59 -0800 | [diff] [blame] | 62 | @Component(immediate = true, enabled = true) |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 63 | @Service |
| 64 | public class DistributedLeadershipManager implements LeadershipService { |
| 65 | |
| 66 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 67 | protected StorageService storageService; |
| 68 | |
| 69 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 70 | protected ClusterService clusterService; |
| 71 | |
| 72 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 73 | protected ClusterCommunicationService clusterCommunicator; |
| 74 | |
| 75 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 76 | protected EventDeliveryService eventDispatcher; |
| 77 | |
| 78 | private static final MessageSubject LEADERSHIP_EVENT_MESSAGE_SUBJECT = |
| 79 | new MessageSubject("distributed-leadership-manager-events"); |
| 80 | |
| 81 | private final Logger log = getLogger(getClass()); |
| 82 | private ExecutorService messageHandlingExecutor; |
| 83 | private ScheduledExecutorService retryLeaderLockExecutor; |
| 84 | private ScheduledExecutorService deadLockDetectionExecutor; |
| 85 | private ScheduledExecutorService leadershipStatusBroadcaster; |
| 86 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 87 | private ConsistentMap<String, NodeId> leaderMap; |
| 88 | private ConsistentMap<String, List<NodeId>> candidateMap; |
| 89 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 90 | private AbstractListenerRegistry<LeadershipEvent, LeadershipEventListener> |
| 91 | listenerRegistry; |
| 92 | private final Map<String, Leadership> leaderBoard = Maps.newConcurrentMap(); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 93 | private final Map<String, Leadership> candidateBoard = Maps.newConcurrentMap(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 94 | private NodeId localNodeId; |
| 95 | |
| 96 | private Set<String> activeTopics = Sets.newConcurrentHashSet(); |
| 97 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 98 | private static final int ELECTION_JOIN_ATTEMPT_INTERVAL_SEC = 2; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 99 | private static final int DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC = 2; |
| 100 | private static final int DEADLOCK_DETECTION_INTERVAL_SEC = 2; |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 101 | private static final int LEADERSHIP_STATUS_UPDATE_INTERVAL_SEC = 2; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 102 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 103 | private static final int LEADER_CANDIDATE_POS = 0; |
| 104 | |
| 105 | private static final Serializer SERIALIZER = Serializer.using( |
| 106 | new KryoNamespace.Builder().register(KryoNamespaces.API).build()); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 107 | |
| 108 | @Activate |
| 109 | public void activate() { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 110 | leaderMap = storageService.<String, NodeId>consistentMapBuilder() |
| 111 | .withName("onos-topic-leaders") |
| 112 | .withSerializer(SERIALIZER) |
| 113 | .withPartitionsDisabled().build(); |
| 114 | candidateMap = storageService.<String, List<NodeId>>consistentMapBuilder() |
| 115 | .withName("onos-topic-candidates") |
| 116 | .withSerializer(SERIALIZER) |
| 117 | .withPartitionsDisabled().build(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 118 | |
| 119 | localNodeId = clusterService.getLocalNode().id(); |
| 120 | |
| 121 | messageHandlingExecutor = Executors.newSingleThreadExecutor( |
| 122 | groupedThreads("onos/store/leadership", "message-handler")); |
| 123 | retryLeaderLockExecutor = Executors.newScheduledThreadPool( |
| 124 | 4, groupedThreads("onos/store/leadership", "election-thread-%d")); |
| 125 | deadLockDetectionExecutor = Executors.newSingleThreadScheduledExecutor( |
| 126 | groupedThreads("onos/store/leadership", "dead-lock-detector")); |
| 127 | leadershipStatusBroadcaster = Executors.newSingleThreadScheduledExecutor( |
| 128 | groupedThreads("onos/store/leadership", "peer-updater")); |
| 129 | clusterCommunicator.addSubscriber( |
| 130 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 131 | new InternalLeadershipEventListener(), |
| 132 | messageHandlingExecutor); |
| 133 | |
| 134 | deadLockDetectionExecutor.scheduleWithFixedDelay( |
| 135 | this::purgeStaleLocks, 0, DEADLOCK_DETECTION_INTERVAL_SEC, TimeUnit.SECONDS); |
| 136 | leadershipStatusBroadcaster.scheduleWithFixedDelay( |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 137 | this::sendLeadershipStatus, 0, LEADERSHIP_STATUS_UPDATE_INTERVAL_SEC, TimeUnit.SECONDS); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 138 | |
| 139 | listenerRegistry = new AbstractListenerRegistry<>(); |
| 140 | eventDispatcher.addSink(LeadershipEvent.class, listenerRegistry); |
| 141 | |
| 142 | log.info("Started."); |
| 143 | } |
| 144 | |
| 145 | @Deactivate |
| 146 | public void deactivate() { |
| 147 | leaderBoard.forEach((topic, leadership) -> { |
| 148 | if (localNodeId.equals(leadership.leader())) { |
| 149 | withdraw(topic); |
| 150 | } |
| 151 | }); |
| 152 | |
| 153 | eventDispatcher.removeSink(LeadershipEvent.class); |
| 154 | clusterCommunicator.removeSubscriber(LEADERSHIP_EVENT_MESSAGE_SUBJECT); |
| 155 | |
| 156 | messageHandlingExecutor.shutdown(); |
| 157 | retryLeaderLockExecutor.shutdown(); |
| 158 | deadLockDetectionExecutor.shutdown(); |
| 159 | leadershipStatusBroadcaster.shutdown(); |
| 160 | |
| 161 | log.info("Stopped."); |
| 162 | } |
| 163 | |
| 164 | @Override |
| 165 | public Map<String, Leadership> getLeaderBoard() { |
| 166 | return ImmutableMap.copyOf(leaderBoard); |
| 167 | } |
| 168 | |
| 169 | @Override |
Madan Jampani | fd45d5e | 2015-04-20 13:33:21 -0700 | [diff] [blame] | 170 | public Map<String, List<NodeId>> getCandidates() { |
| 171 | return Maps.toMap(candidateBoard.keySet(), this::getCandidates); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 172 | } |
| 173 | |
| 174 | @Override |
| 175 | public List<NodeId> getCandidates(String path) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 176 | Leadership current = candidateBoard.get(path); |
| 177 | return current == null ? ImmutableList.of() : ImmutableList.copyOf(current.candidates()); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 178 | } |
| 179 | |
| 180 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 181 | public NodeId getLeader(String path) { |
| 182 | Leadership leadership = leaderBoard.get(path); |
| 183 | return leadership != null ? leadership.leader() : null; |
| 184 | } |
| 185 | |
| 186 | @Override |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 187 | public Leadership getLeadership(String path) { |
| 188 | checkArgument(path != null); |
| 189 | return leaderBoard.get(path); |
| 190 | } |
| 191 | |
| 192 | @Override |
| 193 | public Set<String> ownedTopics(NodeId nodeId) { |
| 194 | checkArgument(nodeId != null); |
| 195 | return leaderBoard.entrySet() |
| 196 | .stream() |
| 197 | .filter(entry -> nodeId.equals(entry.getValue().leader())) |
| 198 | .map(Entry::getKey) |
| 199 | .collect(Collectors.toSet()); |
| 200 | } |
| 201 | |
| 202 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 203 | public void runForLeadership(String path) { |
Madan Jampani | 52860be | 2015-02-27 12:52:37 -0800 | [diff] [blame] | 204 | log.debug("Running for leadership for topic: {}", path); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 205 | try { |
| 206 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 207 | if (candidates != null) { |
| 208 | List<NodeId> candidateList = Lists.newArrayList(candidates.value()); |
| 209 | if (!candidateList.contains(localNodeId)) { |
| 210 | candidateList.add(localNodeId); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 211 | if (candidateMap.replace(path, candidates.version(), candidateList)) { |
| 212 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
| 213 | notifyCandidateAdded( |
| 214 | path, candidateList, newCandidates.version(), newCandidates.creationTime()); |
| 215 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 216 | rerunForLeadership(path); |
| 217 | return; |
| 218 | } |
| 219 | } |
| 220 | } else { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 221 | List<NodeId> candidateList = ImmutableList.of(localNodeId); |
| 222 | if ((candidateMap.putIfAbsent(path, candidateList) == null)) { |
| 223 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
| 224 | notifyCandidateAdded(path, candidateList, newCandidates.version(), newCandidates.creationTime()); |
| 225 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 226 | rerunForLeadership(path); |
| 227 | return; |
| 228 | } |
| 229 | } |
| 230 | log.debug("In the leadership race for topic {} with candidates {}", path, candidates); |
| 231 | activeTopics.add(path); |
| 232 | tryLeaderLock(path); |
| 233 | } catch (ConsistentMapException e) { |
| 234 | log.debug("Failed to enter topic leader race for {}. Retrying.", path, e); |
| 235 | rerunForLeadership(path); |
| 236 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 237 | } |
| 238 | |
| 239 | @Override |
| 240 | public void withdraw(String path) { |
| 241 | activeTopics.remove(path); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 242 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 243 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 244 | Versioned<NodeId> leader = leaderMap.get(path); |
| 245 | if (leader != null && Objects.equals(leader.value(), localNodeId)) { |
| 246 | if (leaderMap.remove(path, leader.version())) { |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 247 | log.info("Gave up leadership for {}", path); |
| 248 | notifyRemovedLeader(path, localNodeId, leader.version(), leader.creationTime()); |
| 249 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 250 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 251 | // else we are not the current leader, can still be a candidate. |
| 252 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 253 | List<NodeId> candidateList = candidates != null |
| 254 | ? Lists.newArrayList(candidates.value()) |
| 255 | : Lists.newArrayList(); |
| 256 | if (!candidateList.remove(localNodeId)) { |
| 257 | return; |
| 258 | } |
Ayaka Koshibe | 93451b4 | 2015-04-21 15:12:03 -0700 | [diff] [blame] | 259 | if (candidateMap.replace(path, candidates.version(), candidateList)) { |
Ayaka Koshibe | 67cf7de | 2015-04-17 11:47:27 -0700 | [diff] [blame] | 260 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
| 261 | notifyCandidateRemoved(path, candidates.version(), candidates.creationTime(), newCandidates); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 262 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 263 | log.warn("Failed to withdraw from candidates list. Will retry"); |
| 264 | retryWithdraw(path); |
| 265 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 266 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 267 | log.debug("Failed to verify (and clear) any lock this node might be holding for {}", path, e); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 268 | retryWithdraw(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 269 | } |
| 270 | } |
| 271 | |
| 272 | @Override |
| 273 | public void addListener(LeadershipEventListener listener) { |
| 274 | listenerRegistry.addListener(listener); |
| 275 | } |
| 276 | |
| 277 | @Override |
| 278 | public void removeListener(LeadershipEventListener listener) { |
| 279 | listenerRegistry.removeListener(listener); |
| 280 | } |
| 281 | |
| 282 | private void tryLeaderLock(String path) { |
| 283 | if (!activeTopics.contains(path)) { |
| 284 | return; |
| 285 | } |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 286 | try { |
| 287 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 288 | if (candidates != null) { |
| 289 | List<NodeId> activeNodes = candidates.value().stream() |
| 290 | .filter(n -> clusterService.getState(n) == ACTIVE) |
| 291 | .collect(Collectors.toList()); |
| 292 | if (localNodeId.equals(activeNodes.get(LEADER_CANDIDATE_POS))) { |
| 293 | leaderLockAttempt(path, candidates.value()); |
| 294 | } else { |
| 295 | retryLock(path); |
| 296 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 297 | } else { |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 298 | throw new IllegalStateException("should not be here"); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 299 | } |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 300 | } catch (Exception e) { |
| 301 | log.debug("Failed to fetch candidate information for {}", path, e); |
| 302 | retryLock(path); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 303 | } |
| 304 | } |
| 305 | |
| 306 | private void leaderLockAttempt(String path, List<NodeId> candidates) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 307 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 308 | Versioned<NodeId> currentLeader = leaderMap.get(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 309 | if (currentLeader != null) { |
| 310 | if (localNodeId.equals(currentLeader.value())) { |
| 311 | log.info("Already has leadership for {}", path); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 312 | // FIXME: candidates can get out of sync. |
| 313 | notifyNewLeader( |
| 314 | path, localNodeId, candidates, currentLeader.version(), currentLeader.creationTime()); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 315 | } else { |
| 316 | // someone else has leadership. will retry after sometime. |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 317 | retryLock(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 318 | } |
| 319 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 320 | if (leaderMap.putIfAbsent(path, localNodeId) == null) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 321 | log.info("Assumed leadership for {}", path); |
| 322 | // do a get again to get the version (epoch) |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 323 | Versioned<NodeId> newLeader = leaderMap.get(path); |
| 324 | // FIXME: candidates can get out of sync |
| 325 | notifyNewLeader(path, localNodeId, candidates, newLeader.version(), newLeader.creationTime()); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 326 | } else { |
| 327 | // someone beat us to it. |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 328 | retryLock(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 329 | } |
| 330 | } |
| 331 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 332 | log.debug("Attempt to acquire leadership lock for topic {} failed", path, e); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 333 | retryLock(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 334 | } |
| 335 | } |
| 336 | |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 337 | private void notifyCandidateAdded( |
| 338 | String path, List<NodeId> candidates, long epoch, long electedTime) { |
| 339 | Leadership newInfo = new Leadership(path, candidates, epoch, electedTime); |
| 340 | final MutableBoolean updated = new MutableBoolean(false); |
| 341 | candidateBoard.compute(path, (k, current) -> { |
| 342 | if (current == null || current.epoch() < newInfo.epoch()) { |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 343 | log.debug("updating candidateboard with {}", newInfo); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 344 | updated.setTrue(); |
| 345 | return newInfo; |
| 346 | } |
| 347 | return current; |
| 348 | }); |
| 349 | // maybe rethink types of candidates events |
| 350 | if (updated.booleanValue()) { |
| 351 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.CANDIDATES_CHANGED, newInfo); |
| 352 | notifyPeers(event); |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | private void notifyCandidateRemoved( |
Ayaka Koshibe | 67cf7de | 2015-04-17 11:47:27 -0700 | [diff] [blame] | 357 | String path, long oldEpoch, long oldTime, Versioned<List<NodeId>> candidates) { |
| 358 | Leadership newInfo = (candidates == null) |
| 359 | ? new Leadership(path, ImmutableList.of(), oldEpoch, oldTime) |
| 360 | : new Leadership(path, candidates.value(), candidates.version(), candidates.creationTime()); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 361 | final MutableBoolean updated = new MutableBoolean(false); |
Ayaka Koshibe | 67cf7de | 2015-04-17 11:47:27 -0700 | [diff] [blame] | 362 | |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 363 | candidateBoard.compute(path, (k, current) -> { |
Ayaka Koshibe | 93451b4 | 2015-04-21 15:12:03 -0700 | [diff] [blame] | 364 | if (current != null && current.epoch() < newInfo.epoch()) { |
| 365 | updated.setTrue(); |
| 366 | return newInfo; |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 367 | } |
| 368 | return current; |
| 369 | }); |
| 370 | // maybe rethink types of candidates events |
| 371 | if (updated.booleanValue()) { |
Ayaka Koshibe | 67cf7de | 2015-04-17 11:47:27 -0700 | [diff] [blame] | 372 | log.debug("updated candidateboard with removal: {}", newInfo); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 373 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.CANDIDATES_CHANGED, newInfo); |
| 374 | notifyPeers(event); |
| 375 | } |
| 376 | } |
| 377 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 378 | private void notifyNewLeader(String path, NodeId leader, |
| 379 | List<NodeId> candidates, long epoch, long electedTime) { |
| 380 | Leadership newLeadership = new Leadership(path, leader, candidates, epoch, electedTime); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 381 | final MutableBoolean updatedLeader = new MutableBoolean(false); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 382 | log.debug("candidates for new Leadership {}", candidates); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 383 | leaderBoard.compute(path, (k, currentLeader) -> { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 384 | if (currentLeader == null || currentLeader.epoch() < epoch) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 385 | log.debug("updating leaderboard with new {}", newLeadership); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 386 | updatedLeader.setTrue(); |
| 387 | return newLeadership; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 388 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 389 | return currentLeader; |
| 390 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 391 | |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 392 | if (updatedLeader.booleanValue()) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 393 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.LEADER_ELECTED, newLeadership); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 394 | notifyPeers(event); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 395 | } |
| 396 | } |
| 397 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 398 | private void notifyPeers(LeadershipEvent event) { |
| 399 | eventDispatcher.post(event); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 400 | clusterCommunicator.broadcast(event, |
| 401 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 402 | SERIALIZER::encode); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 403 | } |
| 404 | |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 405 | private void notifyRemovedLeader(String path, NodeId leader, long epoch, long electedTime) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 406 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 407 | Leadership oldLeadership = new Leadership( |
| 408 | path, leader, candidates.value(), epoch, electedTime); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 409 | final MutableBoolean updatedLeader = new MutableBoolean(false); |
| 410 | leaderBoard.compute(path, (k, currentLeader) -> { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 411 | if (currentLeader != null && currentLeader.epoch() == oldLeadership.epoch()) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 412 | updatedLeader.setTrue(); |
| 413 | return null; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 414 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 415 | return currentLeader; |
| 416 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 417 | |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 418 | if (updatedLeader.booleanValue()) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 419 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.LEADER_BOOTED, oldLeadership); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 420 | notifyPeers(event); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 421 | } |
| 422 | } |
| 423 | |
| 424 | private class InternalLeadershipEventListener implements ClusterMessageHandler { |
| 425 | |
| 426 | @Override |
| 427 | public void handle(ClusterMessage message) { |
| 428 | LeadershipEvent leadershipEvent = |
| 429 | SERIALIZER.decode(message.payload()); |
| 430 | |
Jonathan Hart | b59ac4e | 2015-03-24 18:00:34 -0700 | [diff] [blame] | 431 | log.trace("Leadership Event: time = {} type = {} event = {}", |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 432 | leadershipEvent.time(), leadershipEvent.type(), |
| 433 | leadershipEvent); |
| 434 | |
| 435 | Leadership leadershipUpdate = leadershipEvent.subject(); |
| 436 | LeadershipEvent.Type eventType = leadershipEvent.type(); |
| 437 | String topic = leadershipUpdate.topic(); |
| 438 | |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 439 | MutableBoolean updateAccepted = new MutableBoolean(false); |
| 440 | if (eventType.equals(LeadershipEvent.Type.LEADER_ELECTED)) { |
| 441 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 442 | if (currentLeadership == null || currentLeadership.epoch() < leadershipUpdate.epoch()) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 443 | updateAccepted.setTrue(); |
| 444 | return leadershipUpdate; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 445 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 446 | return currentLeadership; |
| 447 | }); |
| 448 | } else if (eventType.equals(LeadershipEvent.Type.LEADER_BOOTED)) { |
| 449 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
Ayaka Koshibe | 941f860 | 2015-04-15 14:17:08 -0700 | [diff] [blame] | 450 | if (currentLeadership == null || currentLeadership.epoch() == leadershipUpdate.epoch()) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 451 | updateAccepted.setTrue(); |
| 452 | return null; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 453 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 454 | return currentLeadership; |
| 455 | }); |
| 456 | } else if (eventType.equals(LeadershipEvent.Type.CANDIDATES_CHANGED)) { |
| 457 | candidateBoard.compute(topic, (k, currentInfo) -> { |
Ayaka Koshibe | 93451b4 | 2015-04-21 15:12:03 -0700 | [diff] [blame] | 458 | if (currentInfo == null || currentInfo.epoch() < leadershipUpdate.epoch()) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 459 | updateAccepted.setTrue(); |
| 460 | return leadershipUpdate; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 461 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 462 | return currentInfo; |
| 463 | }); |
| 464 | } else { |
| 465 | throw new IllegalStateException("Unknown event type."); |
| 466 | } |
| 467 | |
| 468 | if (updateAccepted.booleanValue()) { |
| 469 | eventDispatcher.post(leadershipEvent); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 470 | } |
| 471 | } |
| 472 | } |
| 473 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 474 | private void rerunForLeadership(String path) { |
| 475 | retryLeaderLockExecutor.schedule( |
| 476 | () -> runForLeadership(path), |
| 477 | ELECTION_JOIN_ATTEMPT_INTERVAL_SEC, |
| 478 | TimeUnit.SECONDS); |
| 479 | } |
| 480 | |
| 481 | private void retryLock(String path) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 482 | retryLeaderLockExecutor.schedule( |
| 483 | () -> tryLeaderLock(path), |
| 484 | DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC, |
| 485 | TimeUnit.SECONDS); |
| 486 | } |
| 487 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 488 | private void retryWithdraw(String path) { |
| 489 | retryLeaderLockExecutor.schedule( |
| 490 | () -> withdraw(path), |
| 491 | DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC, |
| 492 | TimeUnit.SECONDS); |
| 493 | } |
| 494 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 495 | private void purgeStaleLocks() { |
| 496 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 497 | leaderMap.entrySet() |
| 498 | .stream() |
Ayaka Koshibe | 4a3c239 | 2015-04-22 18:03:13 -0700 | [diff] [blame] | 499 | .filter(e -> clusterService.getState(e.getValue().value()) == INACTIVE) |
| 500 | .filter(e -> activeTopics.contains(e.getKey())) |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 501 | .forEach(entry -> { |
| 502 | String path = entry.getKey(); |
| 503 | NodeId nodeId = entry.getValue().value(); |
| 504 | long epoch = entry.getValue().version(); |
| 505 | long creationTime = entry.getValue().creationTime(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 506 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 507 | if (leaderMap.remove(path, epoch)) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 508 | log.info("Purged stale lock held by {} for {}", nodeId, path); |
Madan Jampani | 30a57f8 | 2015-03-02 12:19:41 -0800 | [diff] [blame] | 509 | notifyRemovedLeader(path, nodeId, epoch, creationTime); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 510 | } |
| 511 | } catch (Exception e) { |
| 512 | log.warn("Failed to purge stale lock held by {} for {}", nodeId, path, e); |
| 513 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 514 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 515 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 516 | log.debug("Failed cleaning up stale locks", e); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 517 | } |
| 518 | } |
| 519 | |
| 520 | private void sendLeadershipStatus() { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 521 | try { |
| 522 | leaderBoard.forEach((path, leadership) -> { |
| 523 | if (leadership.leader().equals(localNodeId)) { |
| 524 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.LEADER_ELECTED, leadership); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 525 | clusterCommunicator.broadcast(event, |
| 526 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 527 | SERIALIZER::encode); |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 528 | } |
| 529 | }); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 530 | candidateBoard.forEach((path, leadership) -> { |
| 531 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.CANDIDATES_CHANGED, leadership); |
| 532 | clusterCommunicator.broadcast(event, |
| 533 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 534 | SERIALIZER::encode); |
| 535 | }); |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 536 | } catch (Exception e) { |
| 537 | log.debug("Failed to send leadership updates", e); |
| 538 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 539 | } |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 540 | } |