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; |
| 15 | import org.onlab.util.KryoNamespace; |
| 16 | import org.onosproject.cluster.ClusterService; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 17 | import org.onosproject.cluster.Leadership; |
| 18 | import org.onosproject.cluster.LeadershipEvent; |
| 19 | import org.onosproject.cluster.LeadershipEventListener; |
| 20 | import org.onosproject.cluster.LeadershipService; |
| 21 | import org.onosproject.cluster.NodeId; |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 22 | import org.onosproject.event.ListenerRegistry; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 23 | import org.onosproject.event.EventDeliveryService; |
| 24 | import org.onosproject.store.cluster.messaging.ClusterCommunicationService; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 25 | import org.onosproject.store.cluster.messaging.MessageSubject; |
| 26 | import org.onosproject.store.serializers.KryoNamespaces; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 27 | import org.onosproject.store.service.ConsistentMap; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 28 | import org.onosproject.store.service.ConsistentMapException; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 29 | import org.onosproject.store.service.Serializer; |
| 30 | import org.onosproject.store.service.StorageService; |
| 31 | import org.onosproject.store.service.Versioned; |
| 32 | import org.slf4j.Logger; |
| 33 | |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 34 | import java.util.ArrayList; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 35 | import java.util.Map; |
| 36 | import java.util.Map.Entry; |
| 37 | import java.util.Objects; |
| 38 | import java.util.Set; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 39 | import java.util.List; |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 40 | import java.util.concurrent.CancellationException; |
| 41 | import java.util.concurrent.CompletableFuture; |
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; |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 46 | import java.util.concurrent.atomic.AtomicBoolean; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 47 | import java.util.stream.Collectors; |
| 48 | |
| 49 | import static com.google.common.base.Preconditions.checkArgument; |
| 50 | import static org.onlab.util.Tools.groupedThreads; |
| 51 | import static org.slf4j.LoggerFactory.getLogger; |
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 | |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 90 | private ListenerRegistry<LeadershipEvent, LeadershipEventListener> |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 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, |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 131 | SERIALIZER::decode, |
| 132 | this::onLeadershipEvent, |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 133 | messageHandlingExecutor); |
| 134 | |
| 135 | deadLockDetectionExecutor.scheduleWithFixedDelay( |
| 136 | this::purgeStaleLocks, 0, DEADLOCK_DETECTION_INTERVAL_SEC, TimeUnit.SECONDS); |
| 137 | leadershipStatusBroadcaster.scheduleWithFixedDelay( |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 138 | this::sendLeadershipStatus, 0, LEADERSHIP_STATUS_UPDATE_INTERVAL_SEC, TimeUnit.SECONDS); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 139 | |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 140 | listenerRegistry = new ListenerRegistry<>(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 141 | eventDispatcher.addSink(LeadershipEvent.class, listenerRegistry); |
| 142 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 143 | log.info("Started"); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 144 | } |
| 145 | |
| 146 | @Deactivate |
| 147 | public void deactivate() { |
| 148 | leaderBoard.forEach((topic, leadership) -> { |
| 149 | if (localNodeId.equals(leadership.leader())) { |
| 150 | withdraw(topic); |
| 151 | } |
| 152 | }); |
| 153 | |
| 154 | eventDispatcher.removeSink(LeadershipEvent.class); |
| 155 | clusterCommunicator.removeSubscriber(LEADERSHIP_EVENT_MESSAGE_SUBJECT); |
| 156 | |
| 157 | messageHandlingExecutor.shutdown(); |
| 158 | retryLeaderLockExecutor.shutdown(); |
| 159 | deadLockDetectionExecutor.shutdown(); |
| 160 | leadershipStatusBroadcaster.shutdown(); |
| 161 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 162 | log.info("Stopped"); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 163 | } |
| 164 | |
| 165 | @Override |
| 166 | public Map<String, Leadership> getLeaderBoard() { |
| 167 | return ImmutableMap.copyOf(leaderBoard); |
| 168 | } |
| 169 | |
| 170 | @Override |
Madan Jampani | fd45d5e | 2015-04-20 13:33:21 -0700 | [diff] [blame] | 171 | public Map<String, List<NodeId>> getCandidates() { |
| 172 | return Maps.toMap(candidateBoard.keySet(), this::getCandidates); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 173 | } |
| 174 | |
| 175 | @Override |
| 176 | public List<NodeId> getCandidates(String path) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 177 | Leadership current = candidateBoard.get(path); |
| 178 | return current == null ? ImmutableList.of() : ImmutableList.copyOf(current.candidates()); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 179 | } |
| 180 | |
| 181 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 182 | public NodeId getLeader(String path) { |
| 183 | Leadership leadership = leaderBoard.get(path); |
| 184 | return leadership != null ? leadership.leader() : null; |
| 185 | } |
| 186 | |
| 187 | @Override |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 188 | public Leadership getLeadership(String path) { |
| 189 | checkArgument(path != null); |
| 190 | return leaderBoard.get(path); |
| 191 | } |
| 192 | |
| 193 | @Override |
| 194 | public Set<String> ownedTopics(NodeId nodeId) { |
| 195 | checkArgument(nodeId != null); |
| 196 | return leaderBoard.entrySet() |
| 197 | .stream() |
| 198 | .filter(entry -> nodeId.equals(entry.getValue().leader())) |
| 199 | .map(Entry::getKey) |
| 200 | .collect(Collectors.toSet()); |
| 201 | } |
| 202 | |
| 203 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 204 | public CompletableFuture<Leadership> runForLeadership(String path) { |
Madan Jampani | 52860be | 2015-02-27 12:52:37 -0800 | [diff] [blame] | 205 | log.debug("Running for leadership for topic: {}", path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 206 | CompletableFuture<Leadership> resultFuture = new CompletableFuture<>(); |
| 207 | doRunForLeadership(path, resultFuture); |
| 208 | return resultFuture; |
| 209 | } |
| 210 | |
| 211 | private void doRunForLeadership(String path, CompletableFuture<Leadership> future) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 212 | try { |
| 213 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 214 | if (candidates != null) { |
| 215 | List<NodeId> candidateList = Lists.newArrayList(candidates.value()); |
| 216 | if (!candidateList.contains(localNodeId)) { |
| 217 | candidateList.add(localNodeId); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 218 | if (candidateMap.replace(path, candidates.version(), candidateList)) { |
| 219 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 220 | publish(new LeadershipEvent( |
| 221 | LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 222 | new Leadership(path, |
| 223 | newCandidates.value(), |
| 224 | newCandidates.version(), |
| 225 | newCandidates.creationTime()))); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 226 | } else { |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 227 | rerunForLeadership(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 228 | return; |
| 229 | } |
| 230 | } |
| 231 | } else { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 232 | List<NodeId> candidateList = ImmutableList.of(localNodeId); |
| 233 | if ((candidateMap.putIfAbsent(path, candidateList) == null)) { |
| 234 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 235 | publish(new LeadershipEvent( |
| 236 | LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 237 | new Leadership(path, |
| 238 | newCandidates.value(), |
| 239 | newCandidates.version(), |
| 240 | newCandidates.creationTime()))); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 241 | } else { |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 242 | rerunForLeadership(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 243 | return; |
| 244 | } |
| 245 | } |
| 246 | log.debug("In the leadership race for topic {} with candidates {}", path, candidates); |
| 247 | activeTopics.add(path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 248 | tryLeaderLock(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 249 | } catch (ConsistentMapException e) { |
| 250 | log.debug("Failed to enter topic leader race for {}. Retrying.", path, e); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 251 | rerunForLeadership(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 252 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 253 | } |
| 254 | |
| 255 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 256 | public CompletableFuture<Void> withdraw(String path) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 257 | activeTopics.remove(path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 258 | CompletableFuture<Void> resultFuture = new CompletableFuture<>(); |
| 259 | doWithdraw(path, resultFuture); |
| 260 | return resultFuture; |
| 261 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 262 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 263 | |
| 264 | private void doWithdraw(String path, CompletableFuture<Void> future) { |
| 265 | if (activeTopics.contains(path)) { |
| 266 | future.completeExceptionally(new CancellationException(String.format("%s is now a active topic", path))); |
| 267 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 268 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 269 | Versioned<NodeId> leader = leaderMap.get(path); |
| 270 | if (leader != null && Objects.equals(leader.value(), localNodeId)) { |
| 271 | if (leaderMap.remove(path, leader.version())) { |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 272 | log.info("Gave up leadership for {}", path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 273 | future.complete(null); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 274 | publish(new LeadershipEvent( |
| 275 | LeadershipEvent.Type.LEADER_BOOTED, |
| 276 | new Leadership(path, |
| 277 | localNodeId, |
| 278 | leader.version(), |
| 279 | leader.creationTime()))); |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 280 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 281 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 282 | // else we are not the current leader, can still be a candidate. |
| 283 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 284 | List<NodeId> candidateList = candidates != null |
| 285 | ? Lists.newArrayList(candidates.value()) |
| 286 | : Lists.newArrayList(); |
| 287 | if (!candidateList.remove(localNodeId)) { |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 288 | future.complete(null); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 289 | return; |
| 290 | } |
Ayaka Koshibe | 93451b4 | 2015-04-21 15:12:03 -0700 | [diff] [blame] | 291 | if (candidateMap.replace(path, candidates.version(), candidateList)) { |
Ayaka Koshibe | 67cf7de | 2015-04-17 11:47:27 -0700 | [diff] [blame] | 292 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 293 | future.complete(null); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 294 | publish(new LeadershipEvent( |
| 295 | LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 296 | new Leadership(path, |
| 297 | newCandidates.value(), |
| 298 | newCandidates.version(), |
| 299 | newCandidates.creationTime()))); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 300 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 301 | log.warn("Failed to withdraw from candidates list. Will retry"); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 302 | retryWithdraw(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 303 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 304 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 305 | log.debug("Failed to verify (and clear) any lock this node might be holding for {}", path, e); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 306 | retryWithdraw(path, future); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 307 | } |
| 308 | } |
| 309 | |
| 310 | @Override |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 311 | public boolean stepdown(String path) { |
Madan Jampani | 9bd1f15 | 2015-04-30 23:33:35 -0700 | [diff] [blame] | 312 | if (!activeTopics.contains(path) || !Objects.equals(localNodeId, getLeader(path))) { |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 313 | return false; |
| 314 | } |
| 315 | |
| 316 | try { |
| 317 | Versioned<NodeId> leader = leaderMap.get(path); |
| 318 | if (leader != null && Objects.equals(leader.value(), localNodeId)) { |
| 319 | if (leaderMap.remove(path, leader.version())) { |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 320 | log.info("Stepped down from leadership for {}", path); |
| 321 | publish(new LeadershipEvent( |
| 322 | LeadershipEvent.Type.LEADER_BOOTED, |
| 323 | new Leadership(path, |
| 324 | localNodeId, |
| 325 | leader.version(), |
| 326 | leader.creationTime()))); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 327 | retryLock(path, new CompletableFuture<>()); |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 328 | return true; |
| 329 | } |
| 330 | } |
| 331 | } catch (Exception e) { |
| 332 | log.warn("Error executing stepdown for {}", path, e); |
| 333 | } |
| 334 | return false; |
| 335 | } |
| 336 | |
| 337 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 338 | public void addListener(LeadershipEventListener listener) { |
| 339 | listenerRegistry.addListener(listener); |
| 340 | } |
| 341 | |
| 342 | @Override |
| 343 | public void removeListener(LeadershipEventListener listener) { |
| 344 | listenerRegistry.removeListener(listener); |
| 345 | } |
| 346 | |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 347 | @Override |
| 348 | public boolean makeTopCandidate(String path, NodeId nodeId) { |
| 349 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 350 | if (candidates == null || !candidates.value().contains(nodeId)) { |
| 351 | return false; |
| 352 | } |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 353 | List<NodeId> currentRoster = candidates.value(); |
| 354 | if (nodeId.equals(currentRoster.get(LEADER_CANDIDATE_POS))) { |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 355 | return true; |
| 356 | } |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 357 | List<NodeId> newRoster = new ArrayList<>(currentRoster.size()); |
| 358 | newRoster.add(nodeId); |
| 359 | currentRoster.stream().filter(id -> !nodeId.equals(id)).forEach(newRoster::add); |
| 360 | boolean updated = candidateMap.replace(path, candidates.version(), newRoster); |
| 361 | if (updated) { |
| 362 | Versioned<List<NodeId>> newCandidates = candidateMap.get(path); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 363 | publish(new LeadershipEvent( |
| 364 | LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 365 | new Leadership(path, |
| 366 | newCandidates.value(), |
| 367 | newCandidates.version(), |
| 368 | newCandidates.creationTime()))); |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 369 | } |
| 370 | return updated; |
| 371 | } |
| 372 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 373 | private void tryLeaderLock(String path, CompletableFuture<Leadership> future) { |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 374 | if (!activeTopics.contains(path) || Objects.equals(localNodeId, getLeader(path))) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 375 | return; |
| 376 | } |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 377 | try { |
| 378 | Versioned<List<NodeId>> candidates = candidateMap.get(path); |
| 379 | if (candidates != null) { |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 380 | List<NodeId> activeNodes = candidates.value() |
| 381 | .stream() |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 382 | .filter(n -> clusterService.getState(n) == ACTIVE) |
| 383 | .collect(Collectors.toList()); |
| 384 | if (localNodeId.equals(activeNodes.get(LEADER_CANDIDATE_POS))) { |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 385 | leaderLockAttempt(path, candidates.value(), future); |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 386 | } else { |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 387 | retryLock(path, future); |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 388 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 389 | } else { |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 390 | throw new IllegalStateException("should not be here"); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 391 | } |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 392 | } catch (Exception e) { |
| 393 | log.debug("Failed to fetch candidate information for {}", path, e); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 394 | retryLock(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 395 | } |
| 396 | } |
| 397 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 398 | private void leaderLockAttempt(String path, List<NodeId> candidates, CompletableFuture<Leadership> future) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 399 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 400 | Versioned<NodeId> currentLeader = leaderMap.get(path); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 401 | if (currentLeader != null) { |
| 402 | if (localNodeId.equals(currentLeader.value())) { |
| 403 | log.info("Already has leadership for {}", path); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 404 | // FIXME: candidates can get out of sync. |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 405 | Leadership leadership = new Leadership(path, |
| 406 | localNodeId, |
| 407 | currentLeader.version(), |
| 408 | currentLeader.creationTime()); |
| 409 | future.complete(leadership); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 410 | publish(new LeadershipEvent( |
| 411 | LeadershipEvent.Type.LEADER_ELECTED, |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 412 | leadership)); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 413 | } else { |
| 414 | // someone else has leadership. will retry after sometime. |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 415 | retryLock(path, future); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 416 | } |
| 417 | } else { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 418 | if (leaderMap.putIfAbsent(path, localNodeId) == null) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 419 | log.info("Assumed leadership for {}", path); |
| 420 | // do a get again to get the version (epoch) |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 421 | Versioned<NodeId> newLeader = leaderMap.get(path); |
| 422 | // FIXME: candidates can get out of sync |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 423 | Leadership leadership = new Leadership(path, |
| 424 | newLeader.value(), |
| 425 | newLeader.version(), |
| 426 | newLeader.creationTime()); |
| 427 | future.complete(leadership); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 428 | publish(new LeadershipEvent( |
| 429 | LeadershipEvent.Type.LEADER_ELECTED, |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 430 | leadership)); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 431 | } else { |
| 432 | // someone beat us to it. |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 433 | retryLock(path, future); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 434 | } |
| 435 | } |
| 436 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 437 | log.debug("Attempt to acquire leadership lock for topic {} failed", path, e); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 438 | retryLock(path, future); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 439 | } |
| 440 | } |
| 441 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 442 | private void publish(LeadershipEvent event) { |
| 443 | onLeadershipEvent(event); |
| 444 | clusterCommunicator.broadcast(event, LEADERSHIP_EVENT_MESSAGE_SUBJECT, SERIALIZER::encode); |
| 445 | } |
| 446 | |
| 447 | private void onLeadershipEvent(LeadershipEvent leadershipEvent) { |
| 448 | log.trace("Leadership Event: time = {} type = {} event = {}", |
| 449 | leadershipEvent.time(), leadershipEvent.type(), |
| 450 | leadershipEvent); |
| 451 | |
| 452 | Leadership leadershipUpdate = leadershipEvent.subject(); |
| 453 | LeadershipEvent.Type eventType = leadershipEvent.type(); |
| 454 | String topic = leadershipUpdate.topic(); |
| 455 | |
| 456 | AtomicBoolean updateAccepted = new AtomicBoolean(false); |
| 457 | if (eventType.equals(LeadershipEvent.Type.LEADER_ELECTED)) { |
| 458 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
| 459 | if (currentLeadership == null || currentLeadership.epoch() < leadershipUpdate.epoch()) { |
| 460 | updateAccepted.set(true); |
| 461 | return leadershipUpdate; |
| 462 | } |
| 463 | return currentLeadership; |
| 464 | }); |
| 465 | } else if (eventType.equals(LeadershipEvent.Type.LEADER_BOOTED)) { |
| 466 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
| 467 | if (currentLeadership == null || currentLeadership.epoch() <= leadershipUpdate.epoch()) { |
| 468 | updateAccepted.set(true); |
| 469 | return null; |
| 470 | } |
| 471 | return currentLeadership; |
| 472 | }); |
| 473 | } else if (eventType.equals(LeadershipEvent.Type.CANDIDATES_CHANGED)) { |
| 474 | candidateBoard.compute(topic, (k, currentInfo) -> { |
| 475 | if (currentInfo == null || currentInfo.epoch() < leadershipUpdate.epoch()) { |
| 476 | updateAccepted.set(true); |
| 477 | return leadershipUpdate; |
| 478 | } |
| 479 | return currentInfo; |
| 480 | }); |
| 481 | } else { |
| 482 | throw new IllegalStateException("Unknown event type."); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 483 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 484 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 485 | if (updateAccepted.get()) { |
| 486 | eventDispatcher.post(leadershipEvent); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 487 | } |
| 488 | } |
| 489 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 490 | private void rerunForLeadership(String path, CompletableFuture<Leadership> future) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 491 | retryLeaderLockExecutor.schedule( |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 492 | () -> doRunForLeadership(path, future), |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 493 | ELECTION_JOIN_ATTEMPT_INTERVAL_SEC, |
| 494 | TimeUnit.SECONDS); |
| 495 | } |
| 496 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 497 | private void retryLock(String path, CompletableFuture<Leadership> future) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 498 | retryLeaderLockExecutor.schedule( |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 499 | () -> tryLeaderLock(path, future), |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 500 | DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC, |
| 501 | TimeUnit.SECONDS); |
| 502 | } |
| 503 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 504 | private void retryWithdraw(String path, CompletableFuture<Void> future) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 505 | retryLeaderLockExecutor.schedule( |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 506 | () -> doWithdraw(path, future), |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 507 | DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC, |
| 508 | TimeUnit.SECONDS); |
| 509 | } |
| 510 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 511 | private void purgeStaleLocks() { |
| 512 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 513 | leaderMap.entrySet() |
| 514 | .stream() |
Ayaka Koshibe | 4a3c239 | 2015-04-22 18:03:13 -0700 | [diff] [blame] | 515 | .filter(e -> clusterService.getState(e.getValue().value()) == INACTIVE) |
| 516 | .filter(e -> activeTopics.contains(e.getKey())) |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 517 | .forEach(entry -> { |
| 518 | String path = entry.getKey(); |
| 519 | NodeId nodeId = entry.getValue().value(); |
| 520 | long epoch = entry.getValue().version(); |
| 521 | long creationTime = entry.getValue().creationTime(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 522 | try { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 523 | if (leaderMap.remove(path, epoch)) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 524 | log.info("Purged stale lock held by {} for {}", nodeId, path); |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 525 | publish(new LeadershipEvent( |
| 526 | LeadershipEvent.Type.LEADER_BOOTED, |
| 527 | new Leadership(path, nodeId, epoch, creationTime))); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 528 | } |
| 529 | } catch (Exception e) { |
| 530 | log.warn("Failed to purge stale lock held by {} for {}", nodeId, path, e); |
| 531 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 532 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 533 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 534 | log.debug("Failed cleaning up stale locks", e); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 535 | } |
| 536 | } |
| 537 | |
| 538 | private void sendLeadershipStatus() { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 539 | try { |
| 540 | leaderBoard.forEach((path, leadership) -> { |
| 541 | if (leadership.leader().equals(localNodeId)) { |
| 542 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.LEADER_ELECTED, leadership); |
Madan Jampani | 2bfa94c | 2015-04-11 05:03:49 -0700 | [diff] [blame] | 543 | clusterCommunicator.broadcast(event, |
| 544 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 545 | SERIALIZER::encode); |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 546 | } |
| 547 | }); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 548 | candidateBoard.forEach((path, leadership) -> { |
| 549 | LeadershipEvent event = new LeadershipEvent(LeadershipEvent.Type.CANDIDATES_CHANGED, leadership); |
| 550 | clusterCommunicator.broadcast(event, |
| 551 | LEADERSHIP_EVENT_MESSAGE_SUBJECT, |
| 552 | SERIALIZER::encode); |
| 553 | }); |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 554 | } catch (Exception e) { |
| 555 | log.debug("Failed to send leadership updates", e); |
| 556 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 557 | } |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 558 | } |