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; |
Madan Jampani | 3650fc4 | 2015-05-27 17:01:50 -0700 | [diff] [blame] | 5 | import com.google.common.collect.Iterables; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 6 | import com.google.common.collect.Lists; |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 7 | import com.google.common.collect.MapDifference; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 8 | import com.google.common.collect.Maps; |
| 9 | import com.google.common.collect.Sets; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 10 | |
Madan Jampani | 8d37efc | 2015-06-01 17:25:48 -0700 | [diff] [blame] | 11 | import org.apache.commons.lang.math.RandomUtils; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 12 | import org.apache.felix.scr.annotations.Activate; |
| 13 | import org.apache.felix.scr.annotations.Component; |
| 14 | import org.apache.felix.scr.annotations.Deactivate; |
| 15 | import org.apache.felix.scr.annotations.Reference; |
| 16 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
| 17 | import org.apache.felix.scr.annotations.Service; |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 18 | import org.onosproject.cluster.ClusterEvent; |
| 19 | import org.onosproject.cluster.ClusterEvent.Type; |
| 20 | import org.onosproject.cluster.ClusterEventListener; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 21 | import org.onosproject.cluster.ClusterService; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 22 | import org.onosproject.cluster.Leadership; |
| 23 | import org.onosproject.cluster.LeadershipEvent; |
| 24 | import org.onosproject.cluster.LeadershipEventListener; |
| 25 | import org.onosproject.cluster.LeadershipService; |
| 26 | import org.onosproject.cluster.NodeId; |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 27 | import org.onosproject.event.ListenerRegistry; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 28 | import org.onosproject.event.EventDeliveryService; |
| 29 | import org.onosproject.store.cluster.messaging.ClusterCommunicationService; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 30 | import org.onosproject.store.serializers.KryoNamespaces; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 31 | import org.onosproject.store.service.ConsistentMap; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 32 | import org.onosproject.store.service.ConsistentMapException; |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 33 | import org.onosproject.store.service.MapEvent; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 34 | import org.onosproject.store.service.Serializer; |
| 35 | import org.onosproject.store.service.StorageService; |
| 36 | import org.onosproject.store.service.Versioned; |
| 37 | import org.slf4j.Logger; |
| 38 | |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 39 | import java.util.ArrayList; |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 40 | import java.util.Collections; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 41 | import java.util.Map; |
| 42 | import java.util.Map.Entry; |
| 43 | import java.util.Objects; |
| 44 | import java.util.Set; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 45 | import java.util.List; |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 46 | import java.util.concurrent.CancellationException; |
| 47 | import java.util.concurrent.CompletableFuture; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 48 | import java.util.concurrent.Executors; |
| 49 | import java.util.concurrent.ScheduledExecutorService; |
| 50 | import java.util.concurrent.TimeUnit; |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 51 | import java.util.concurrent.atomic.AtomicBoolean; |
Jonathan Hart | e649c75 | 2015-03-03 18:04:25 -0800 | [diff] [blame] | 52 | import java.util.stream.Collectors; |
| 53 | |
| 54 | import static com.google.common.base.Preconditions.checkArgument; |
| 55 | import static org.onlab.util.Tools.groupedThreads; |
| 56 | import static org.slf4j.LoggerFactory.getLogger; |
Ayaka Koshibe | 4a3c239 | 2015-04-22 18:03:13 -0700 | [diff] [blame] | 57 | import static org.onosproject.cluster.ControllerNode.State.ACTIVE; |
| 58 | import static org.onosproject.cluster.ControllerNode.State.INACTIVE; |
| 59 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 60 | /** |
| 61 | * Distributed Lock Manager implemented on top of ConsistentMap. |
| 62 | * <p> |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 63 | * This implementation makes use of ClusterService's failure |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 64 | * detection capabilities to detect and purge stale locks. |
| 65 | * TODO: Ensure lock safety and liveness. |
| 66 | */ |
Madan Jampani | 7f72c3f | 2015-03-01 17:34:59 -0800 | [diff] [blame] | 67 | @Component(immediate = true, enabled = true) |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 68 | @Service |
| 69 | public class DistributedLeadershipManager implements LeadershipService { |
| 70 | |
| 71 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 72 | protected StorageService storageService; |
| 73 | |
| 74 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 75 | protected ClusterService clusterService; |
| 76 | |
| 77 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 78 | protected ClusterCommunicationService clusterCommunicator; |
| 79 | |
| 80 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
| 81 | protected EventDeliveryService eventDispatcher; |
| 82 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 83 | private final Logger log = getLogger(getClass()); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 84 | private ScheduledExecutorService electionRunner; |
| 85 | private ScheduledExecutorService lockExecutor; |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 86 | private ScheduledExecutorService staleLeadershipPurgeExecutor; |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 87 | private ScheduledExecutorService leadershipRefresher; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 88 | |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 89 | private ConsistentMap<String, NodeId> leaderMap; |
| 90 | private ConsistentMap<String, List<NodeId>> candidateMap; |
| 91 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 92 | private ListenerRegistry<LeadershipEvent, LeadershipEventListener> listenerRegistry; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 93 | private final Map<String, Leadership> leaderBoard = Maps.newConcurrentMap(); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 94 | private final Map<String, Leadership> candidateBoard = Maps.newConcurrentMap(); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 95 | private final ClusterEventListener clusterEventListener = new InternalClusterEventListener(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 96 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 97 | private NodeId localNodeId; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 98 | private Set<String> activeTopics = Sets.newConcurrentHashSet(); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 99 | private Map<String, CompletableFuture<Leadership>> pendingFutures = Maps.newConcurrentMap(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 100 | |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 101 | // The actual delay is randomly chosen from the interval [0, WAIT_BEFORE_RETRY_MILLIS) |
Madan Jampani | 8d37efc | 2015-06-01 17:25:48 -0700 | [diff] [blame] | 102 | private static final int WAIT_BEFORE_RETRY_MILLIS = 150; |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 103 | private static final int DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC = 2; |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 104 | private static final int LEADERSHIP_REFRESH_INTERVAL_SEC = 2; |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 105 | private static final int DELAY_BETWEEN_STALE_LEADERSHIP_PURGE_ATTEMPTS_SEC = 2; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 106 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 107 | private final AtomicBoolean staleLeadershipPurgeScheduled = new AtomicBoolean(false); |
| 108 | |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 109 | private static final Serializer SERIALIZER = Serializer.using(KryoNamespaces.API); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 110 | |
| 111 | @Activate |
| 112 | public void activate() { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 113 | leaderMap = storageService.<String, NodeId>consistentMapBuilder() |
| 114 | .withName("onos-topic-leaders") |
| 115 | .withSerializer(SERIALIZER) |
| 116 | .withPartitionsDisabled().build(); |
| 117 | candidateMap = storageService.<String, List<NodeId>>consistentMapBuilder() |
| 118 | .withName("onos-topic-candidates") |
| 119 | .withSerializer(SERIALIZER) |
| 120 | .withPartitionsDisabled().build(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 121 | |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 122 | leaderMap.addListener(event -> { |
| 123 | log.debug("Received {}", event); |
| 124 | LeadershipEvent.Type leadershipEventType = null; |
| 125 | if (event.type() == MapEvent.Type.INSERT || event.type() == MapEvent.Type.UPDATE) { |
| 126 | leadershipEventType = LeadershipEvent.Type.LEADER_ELECTED; |
| 127 | } else if (event.type() == MapEvent.Type.REMOVE) { |
| 128 | leadershipEventType = LeadershipEvent.Type.LEADER_BOOTED; |
| 129 | } |
| 130 | onLeadershipEvent(new LeadershipEvent( |
| 131 | leadershipEventType, |
| 132 | new Leadership(event.key(), |
| 133 | event.value().value(), |
| 134 | event.value().version(), |
| 135 | event.value().creationTime()))); |
| 136 | }); |
| 137 | |
| 138 | candidateMap.addListener(event -> { |
| 139 | log.debug("Received {}", event); |
| 140 | if (event.type() != MapEvent.Type.INSERT && event.type() != MapEvent.Type.UPDATE) { |
| 141 | log.error("Entries must not be removed from candidate map"); |
| 142 | return; |
| 143 | } |
| 144 | onLeadershipEvent(new LeadershipEvent( |
| 145 | LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 146 | new Leadership(event.key(), |
| 147 | event.value().value(), |
| 148 | event.value().version(), |
| 149 | event.value().creationTime()))); |
| 150 | }); |
| 151 | |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 152 | localNodeId = clusterService.getLocalNode().id(); |
| 153 | |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 154 | electionRunner = Executors.newSingleThreadScheduledExecutor( |
| 155 | groupedThreads("onos/store/leadership", "election-runner")); |
| 156 | lockExecutor = Executors.newScheduledThreadPool( |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 157 | 4, groupedThreads("onos/store/leadership", "election-thread-%d")); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 158 | staleLeadershipPurgeExecutor = Executors.newSingleThreadScheduledExecutor( |
| 159 | groupedThreads("onos/store/leadership", "stale-leadership-evictor")); |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 160 | leadershipRefresher = Executors.newSingleThreadScheduledExecutor( |
| 161 | groupedThreads("onos/store/leadership", "refresh-thread")); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 162 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 163 | clusterService.addListener(clusterEventListener); |
| 164 | |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 165 | electionRunner.scheduleWithFixedDelay( |
| 166 | this::electLeaders, 0, DELAY_BETWEEN_LEADER_LOCK_ATTEMPTS_SEC, TimeUnit.SECONDS); |
| 167 | |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 168 | leadershipRefresher.scheduleWithFixedDelay( |
| 169 | this::refreshLeaderBoard, 0, LEADERSHIP_REFRESH_INTERVAL_SEC, TimeUnit.SECONDS); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 170 | |
Simon Hunt | ff66374 | 2015-05-14 13:33:05 -0700 | [diff] [blame] | 171 | listenerRegistry = new ListenerRegistry<>(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 172 | eventDispatcher.addSink(LeadershipEvent.class, listenerRegistry); |
| 173 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 174 | log.info("Started"); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 175 | } |
| 176 | |
| 177 | @Deactivate |
| 178 | public void deactivate() { |
| 179 | leaderBoard.forEach((topic, leadership) -> { |
| 180 | if (localNodeId.equals(leadership.leader())) { |
| 181 | withdraw(topic); |
| 182 | } |
| 183 | }); |
| 184 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 185 | clusterService.removeListener(clusterEventListener); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 186 | eventDispatcher.removeSink(LeadershipEvent.class); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 187 | |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 188 | electionRunner.shutdown(); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 189 | lockExecutor.shutdown(); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 190 | staleLeadershipPurgeExecutor.shutdown(); |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 191 | leadershipRefresher.shutdown(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 192 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 193 | log.info("Stopped"); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 194 | } |
| 195 | |
| 196 | @Override |
| 197 | public Map<String, Leadership> getLeaderBoard() { |
| 198 | return ImmutableMap.copyOf(leaderBoard); |
| 199 | } |
| 200 | |
| 201 | @Override |
Madan Jampani | fd45d5e | 2015-04-20 13:33:21 -0700 | [diff] [blame] | 202 | public Map<String, List<NodeId>> getCandidates() { |
| 203 | return Maps.toMap(candidateBoard.keySet(), this::getCandidates); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 204 | } |
| 205 | |
| 206 | @Override |
| 207 | public List<NodeId> getCandidates(String path) { |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 208 | Leadership current = candidateBoard.get(path); |
| 209 | return current == null ? ImmutableList.of() : ImmutableList.copyOf(current.candidates()); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 210 | } |
| 211 | |
| 212 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 213 | public NodeId getLeader(String path) { |
| 214 | Leadership leadership = leaderBoard.get(path); |
| 215 | return leadership != null ? leadership.leader() : null; |
| 216 | } |
| 217 | |
| 218 | @Override |
Madan Jampani | 5961051 | 2015-02-25 15:25:43 -0800 | [diff] [blame] | 219 | public Leadership getLeadership(String path) { |
| 220 | checkArgument(path != null); |
| 221 | return leaderBoard.get(path); |
| 222 | } |
| 223 | |
| 224 | @Override |
| 225 | public Set<String> ownedTopics(NodeId nodeId) { |
| 226 | checkArgument(nodeId != null); |
| 227 | return leaderBoard.entrySet() |
| 228 | .stream() |
| 229 | .filter(entry -> nodeId.equals(entry.getValue().leader())) |
| 230 | .map(Entry::getKey) |
| 231 | .collect(Collectors.toSet()); |
| 232 | } |
| 233 | |
| 234 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 235 | public CompletableFuture<Leadership> runForLeadership(String path) { |
Madan Jampani | 52860be | 2015-02-27 12:52:37 -0800 | [diff] [blame] | 236 | log.debug("Running for leadership for topic: {}", path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 237 | CompletableFuture<Leadership> resultFuture = new CompletableFuture<>(); |
| 238 | doRunForLeadership(path, resultFuture); |
| 239 | return resultFuture; |
| 240 | } |
| 241 | |
| 242 | private void doRunForLeadership(String path, CompletableFuture<Leadership> future) { |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 243 | try { |
Madan Jampani | 346d4f5 | 2015-05-04 11:09:39 -0700 | [diff] [blame] | 244 | Versioned<List<NodeId>> candidates = candidateMap.computeIf(path, |
| 245 | currentList -> currentList == null || !currentList.contains(localNodeId), |
| 246 | (topic, currentList) -> { |
| 247 | if (currentList == null) { |
| 248 | return ImmutableList.of(localNodeId); |
| 249 | } else { |
| 250 | List<NodeId> newList = Lists.newLinkedList(); |
| 251 | newList.addAll(currentList); |
| 252 | newList.add(localNodeId); |
| 253 | return newList; |
| 254 | } |
| 255 | }); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 256 | log.debug("In the leadership race for topic {} with candidates {}", path, candidates); |
| 257 | activeTopics.add(path); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 258 | Leadership leadership = electLeader(path, candidates.value()); |
| 259 | if (leadership == null) { |
| 260 | pendingFutures.put(path, future); |
| 261 | } else { |
| 262 | future.complete(leadership); |
| 263 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 264 | } catch (ConsistentMapException e) { |
| 265 | log.debug("Failed to enter topic leader race for {}. Retrying.", path, e); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 266 | rerunForLeadership(path, future); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 267 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 268 | } |
| 269 | |
| 270 | @Override |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 271 | public CompletableFuture<Void> withdraw(String path) { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 272 | activeTopics.remove(path); |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 273 | CompletableFuture<Void> resultFuture = new CompletableFuture<>(); |
| 274 | doWithdraw(path, resultFuture); |
| 275 | return resultFuture; |
| 276 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 277 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 278 | |
| 279 | private void doWithdraw(String path, CompletableFuture<Void> future) { |
| 280 | if (activeTopics.contains(path)) { |
| 281 | future.completeExceptionally(new CancellationException(String.format("%s is now a active topic", path))); |
| 282 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 283 | try { |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 284 | leaderMap.computeIf(path, |
| 285 | localNodeId::equals, |
| 286 | (topic, leader) -> null); |
| 287 | candidateMap.computeIf(path, |
| 288 | candidates -> candidates != null && candidates.contains(localNodeId), |
| 289 | (topic, candidates) -> candidates.stream() |
| 290 | .filter(nodeId -> !localNodeId.equals(nodeId)) |
| 291 | .collect(Collectors.toList())); |
Madan Jampani | 052ad65 | 2015-06-10 13:20:32 -0700 | [diff] [blame] | 292 | future.complete(null); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 293 | } catch (Exception e) { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 294 | 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] | 295 | retryWithdraw(path, future); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 296 | } |
| 297 | } |
| 298 | |
| 299 | @Override |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 300 | public boolean stepdown(String path) { |
Madan Jampani | 9bd1f15 | 2015-04-30 23:33:35 -0700 | [diff] [blame] | 301 | if (!activeTopics.contains(path) || !Objects.equals(localNodeId, getLeader(path))) { |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 302 | return false; |
| 303 | } |
| 304 | |
| 305 | try { |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 306 | return leaderMap.computeIf(path, |
| 307 | localNodeId::equals, |
| 308 | (topic, leader) -> null) == null; |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 309 | } catch (Exception e) { |
| 310 | log.warn("Error executing stepdown for {}", path, e); |
| 311 | } |
| 312 | return false; |
| 313 | } |
| 314 | |
| 315 | @Override |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 316 | public void addListener(LeadershipEventListener listener) { |
| 317 | listenerRegistry.addListener(listener); |
| 318 | } |
| 319 | |
| 320 | @Override |
| 321 | public void removeListener(LeadershipEventListener listener) { |
| 322 | listenerRegistry.removeListener(listener); |
| 323 | } |
| 324 | |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 325 | @Override |
| 326 | public boolean makeTopCandidate(String path, NodeId nodeId) { |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 327 | Versioned<List<NodeId>> candidateList = candidateMap.computeIf(path, |
Madan Jampani | 3650fc4 | 2015-05-27 17:01:50 -0700 | [diff] [blame] | 328 | candidates -> candidates != null && |
| 329 | candidates.contains(nodeId) && |
| 330 | !nodeId.equals(Iterables.getFirst(candidates, null)), |
Madan Jampani | 346d4f5 | 2015-05-04 11:09:39 -0700 | [diff] [blame] | 331 | (topic, candidates) -> { |
| 332 | List<NodeId> updatedCandidates = new ArrayList<>(candidates.size()); |
| 333 | updatedCandidates.add(nodeId); |
| 334 | candidates.stream().filter(id -> !nodeId.equals(id)).forEach(updatedCandidates::add); |
| 335 | return updatedCandidates; |
| 336 | }); |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 337 | List<NodeId> candidates = candidateList != null ? candidateList.value() : Collections.emptyList(); |
| 338 | return candidates.size() > 0 && nodeId.equals(candidates.get(0)); |
Madan Jampani | 1af8e13 | 2015-04-30 16:41:18 -0700 | [diff] [blame] | 339 | } |
| 340 | |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 341 | private Leadership electLeader(String path, List<NodeId> candidates) { |
| 342 | Leadership currentLeadership = getLeadership(path); |
| 343 | if (currentLeadership != null) { |
| 344 | return currentLeadership; |
| 345 | } else { |
| 346 | NodeId topCandidate = candidates |
| 347 | .stream() |
| 348 | .filter(n -> clusterService.getState(n) == ACTIVE) |
| 349 | .findFirst() |
| 350 | .orElse(null); |
| 351 | try { |
| 352 | Versioned<NodeId> leader = localNodeId.equals(topCandidate) |
| 353 | ? leaderMap.computeIfAbsent(path, p -> localNodeId) : leaderMap.get(path); |
| 354 | if (leader != null) { |
| 355 | Leadership newLeadership = new Leadership(path, |
| 356 | leader.value(), |
| 357 | leader.version(), |
| 358 | leader.creationTime()); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 359 | return newLeadership; |
Ayaka Koshibe | 0d886fc | 2015-04-23 11:53:41 -0700 | [diff] [blame] | 360 | } |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 361 | } catch (Exception e) { |
| 362 | log.debug("Failed to elect leader for {}", path, e); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 363 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 364 | } |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 365 | return null; |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 366 | } |
| 367 | |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 368 | private void electLeaders() { |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 369 | try { |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 370 | candidateMap.entrySet().forEach(entry -> { |
| 371 | String path = entry.getKey(); |
Madan Jampani | cd40e5b | 2015-05-27 17:19:48 -0700 | [diff] [blame] | 372 | Versioned<List<NodeId>> candidates = entry.getValue(); |
| 373 | // for active topics, check if this node can become a leader (if it isn't already) |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 374 | if (activeTopics.contains(path)) { |
| 375 | lockExecutor.submit(() -> { |
Madan Jampani | cd40e5b | 2015-05-27 17:19:48 -0700 | [diff] [blame] | 376 | Leadership leadership = electLeader(path, candidates.value()); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 377 | if (leadership != null) { |
| 378 | CompletableFuture<Leadership> future = pendingFutures.remove(path); |
| 379 | if (future != null) { |
| 380 | future.complete(leadership); |
| 381 | } |
| 382 | } |
| 383 | }); |
| 384 | } |
Madan Jampani | cd40e5b | 2015-05-27 17:19:48 -0700 | [diff] [blame] | 385 | // Raise a CANDIDATES_CHANGED event to force refresh local candidate board |
| 386 | // and also to update local listeners. |
| 387 | // Don't worry about duplicate events as they will be suppressed. |
| 388 | onLeadershipEvent(new LeadershipEvent(LeadershipEvent.Type.CANDIDATES_CHANGED, |
| 389 | new Leadership(path, |
| 390 | candidates.value(), |
| 391 | candidates.version(), |
| 392 | candidates.creationTime()))); |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 393 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 394 | } catch (Exception e) { |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 395 | log.debug("Failure electing leaders", e); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 396 | } |
| 397 | } |
| 398 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 399 | private void onLeadershipEvent(LeadershipEvent leadershipEvent) { |
| 400 | log.trace("Leadership Event: time = {} type = {} event = {}", |
| 401 | leadershipEvent.time(), leadershipEvent.type(), |
| 402 | leadershipEvent); |
| 403 | |
| 404 | Leadership leadershipUpdate = leadershipEvent.subject(); |
| 405 | LeadershipEvent.Type eventType = leadershipEvent.type(); |
| 406 | String topic = leadershipUpdate.topic(); |
| 407 | |
| 408 | AtomicBoolean updateAccepted = new AtomicBoolean(false); |
| 409 | if (eventType.equals(LeadershipEvent.Type.LEADER_ELECTED)) { |
| 410 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
| 411 | if (currentLeadership == null || currentLeadership.epoch() < leadershipUpdate.epoch()) { |
| 412 | updateAccepted.set(true); |
| 413 | return leadershipUpdate; |
| 414 | } |
| 415 | return currentLeadership; |
| 416 | }); |
| 417 | } else if (eventType.equals(LeadershipEvent.Type.LEADER_BOOTED)) { |
| 418 | leaderBoard.compute(topic, (k, currentLeadership) -> { |
| 419 | if (currentLeadership == null || currentLeadership.epoch() <= leadershipUpdate.epoch()) { |
| 420 | updateAccepted.set(true); |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 421 | // FIXME: Removing entries from leaderboard is not safe and should be visited. |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 422 | return null; |
| 423 | } |
| 424 | return currentLeadership; |
| 425 | }); |
| 426 | } else if (eventType.equals(LeadershipEvent.Type.CANDIDATES_CHANGED)) { |
| 427 | candidateBoard.compute(topic, (k, currentInfo) -> { |
| 428 | if (currentInfo == null || currentInfo.epoch() < leadershipUpdate.epoch()) { |
| 429 | updateAccepted.set(true); |
| 430 | return leadershipUpdate; |
| 431 | } |
| 432 | return currentInfo; |
| 433 | }); |
| 434 | } else { |
| 435 | throw new IllegalStateException("Unknown event type."); |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 436 | } |
Ayaka Koshibe | fd26a30 | 2015-04-13 13:59:54 -0700 | [diff] [blame] | 437 | |
Madan Jampani | d46e18f | 2015-05-04 23:19:33 -0700 | [diff] [blame] | 438 | if (updateAccepted.get()) { |
| 439 | eventDispatcher.post(leadershipEvent); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 440 | } |
| 441 | } |
| 442 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 443 | private void rerunForLeadership(String path, CompletableFuture<Leadership> future) { |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 444 | lockExecutor.schedule( |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 445 | () -> doRunForLeadership(path, future), |
Madan Jampani | 8d37efc | 2015-06-01 17:25:48 -0700 | [diff] [blame] | 446 | RandomUtils.nextInt(WAIT_BEFORE_RETRY_MILLIS), |
| 447 | TimeUnit.MILLISECONDS); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 448 | } |
| 449 | |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 450 | private void retryWithdraw(String path, CompletableFuture<Void> future) { |
Madan Jampani | 7d243db | 2015-05-27 13:16:02 -0700 | [diff] [blame] | 451 | lockExecutor.schedule( |
Madan Jampani | de003d9 | 2015-05-11 17:14:20 -0700 | [diff] [blame] | 452 | () -> doWithdraw(path, future), |
Madan Jampani | 8d37efc | 2015-06-01 17:25:48 -0700 | [diff] [blame] | 453 | RandomUtils.nextInt(WAIT_BEFORE_RETRY_MILLIS), |
| 454 | TimeUnit.MILLISECONDS); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 455 | } |
| 456 | |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 457 | private void scheduleStaleLeadershipPurge(int afterDelaySec) { |
| 458 | if (staleLeadershipPurgeScheduled.compareAndSet(false, true)) { |
| 459 | staleLeadershipPurgeExecutor.schedule( |
| 460 | this::purgeStaleLeadership, |
| 461 | afterDelaySec, |
| 462 | TimeUnit.SECONDS); |
| 463 | } |
| 464 | } |
| 465 | |
| 466 | /** |
| 467 | * Purges locks held by inactive nodes and evicts inactive nodes from candidacy. |
| 468 | */ |
| 469 | private void purgeStaleLeadership() { |
| 470 | AtomicBoolean rerunPurge = new AtomicBoolean(false); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 471 | try { |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 472 | staleLeadershipPurgeScheduled.set(false); |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 473 | leaderMap.entrySet() |
| 474 | .stream() |
Ayaka Koshibe | 4a3c239 | 2015-04-22 18:03:13 -0700 | [diff] [blame] | 475 | .filter(e -> clusterService.getState(e.getValue().value()) == INACTIVE) |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 476 | .forEach(entry -> { |
| 477 | String path = entry.getKey(); |
| 478 | NodeId nodeId = entry.getValue().value(); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 479 | try { |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 480 | leaderMap.computeIf(path, nodeId::equals, (topic, leader) -> null); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 481 | } catch (Exception e) { |
Madan Jampani | ce8392b | 2015-05-20 11:13:47 -0700 | [diff] [blame] | 482 | log.debug("Failed to purge stale lock held by {} for {}", nodeId, path, e); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 483 | rerunPurge.set(true); |
| 484 | } |
| 485 | }); |
| 486 | |
| 487 | candidateMap.entrySet() |
| 488 | .forEach(entry -> { |
| 489 | String path = entry.getKey(); |
| 490 | Versioned<List<NodeId>> candidates = entry.getValue(); |
| 491 | List<NodeId> candidatesList = candidates != null |
| 492 | ? candidates.value() : Collections.emptyList(); |
| 493 | List<NodeId> activeCandidatesList = |
| 494 | candidatesList.stream() |
| 495 | .filter(n -> clusterService.getState(n) == ACTIVE) |
| 496 | .filter(n -> !localNodeId.equals(n) || activeTopics.contains(path)) |
| 497 | .collect(Collectors.toList()); |
| 498 | if (activeCandidatesList.size() < candidatesList.size()) { |
| 499 | Set<NodeId> removedCandidates = |
| 500 | Sets.difference(Sets.newHashSet(candidatesList), |
| 501 | Sets.newHashSet(activeCandidatesList)); |
| 502 | try { |
Madan Jampani | d09441b | 2015-06-09 14:50:55 -0700 | [diff] [blame] | 503 | candidateMap.computeIf(path, |
| 504 | c -> c.stream() |
| 505 | .filter(n -> clusterService.getState(n) == INACTIVE) |
| 506 | .count() > 0, |
| 507 | (topic, c) -> c.stream() |
| 508 | .filter(n -> clusterService.getState(n) == ACTIVE) |
| 509 | .filter(n -> !localNodeId.equals(n) || |
| 510 | activeTopics.contains(path)) |
| 511 | .collect(Collectors.toList())); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 512 | } catch (Exception e) { |
Madan Jampani | ce8392b | 2015-05-20 11:13:47 -0700 | [diff] [blame] | 513 | log.debug("Failed to evict inactive candidates {} from " |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 514 | + "candidate list for {}", removedCandidates, path, e); |
| 515 | rerunPurge.set(true); |
| 516 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 517 | } |
Ayaka Koshibe | c19b8b8 | 2015-04-08 15:18:24 -0700 | [diff] [blame] | 518 | }); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 519 | } catch (Exception e) { |
Madan Jampani | ce8392b | 2015-05-20 11:13:47 -0700 | [diff] [blame] | 520 | log.debug("Failure purging state leadership.", e); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 521 | rerunPurge.set(true); |
| 522 | } |
| 523 | |
| 524 | if (rerunPurge.get()) { |
Madan Jampani | ce8392b | 2015-05-20 11:13:47 -0700 | [diff] [blame] | 525 | log.debug("Rescheduling stale leadership purge due to errors encountered in previous run"); |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 526 | scheduleStaleLeadershipPurge(DELAY_BETWEEN_STALE_LEADERSHIP_PURGE_ATTEMPTS_SEC); |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 527 | } |
| 528 | } |
| 529 | |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 530 | private void refreshLeaderBoard() { |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 531 | try { |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 532 | Map<String, Leadership> newLeaderBoard = Maps.newHashMap(); |
| 533 | leaderMap.entrySet().forEach(entry -> { |
| 534 | String path = entry.getKey(); |
| 535 | Versioned<NodeId> leader = entry.getValue(); |
| 536 | Leadership leadership = new Leadership(path, |
| 537 | leader.value(), |
| 538 | leader.version(), |
| 539 | leader.creationTime()); |
| 540 | newLeaderBoard.put(path, leadership); |
| 541 | }); |
| 542 | |
| 543 | // first take snapshot of current leader board. |
| 544 | Map<String, Leadership> currentLeaderBoard = ImmutableMap.copyOf(leaderBoard); |
| 545 | |
| 546 | MapDifference<String, Leadership> diff = Maps.difference(currentLeaderBoard, newLeaderBoard); |
| 547 | |
| 548 | // evict stale leaders |
| 549 | diff.entriesOnlyOnLeft().forEach((path, leadership) -> { |
| 550 | log.debug("Evicting {} from leaderboard. It is no longer active leader.", leadership); |
| 551 | onLeadershipEvent(new LeadershipEvent(LeadershipEvent.Type.LEADER_BOOTED, leadership)); |
| 552 | }); |
| 553 | |
| 554 | // add missing leaders |
| 555 | diff.entriesOnlyOnRight().forEach((path, leadership) -> { |
| 556 | log.debug("Adding {} to leaderboard. It is now the active leader.", leadership); |
| 557 | onLeadershipEvent(new LeadershipEvent(LeadershipEvent.Type.LEADER_ELECTED, leadership)); |
| 558 | }); |
| 559 | |
| 560 | // add updated leaders |
| 561 | diff.entriesDiffering().forEach((path, difference) -> { |
| 562 | Leadership current = difference.leftValue(); |
| 563 | Leadership updated = difference.rightValue(); |
| 564 | if (current.epoch() < updated.epoch()) { |
| 565 | log.debug("Updated {} in leaderboard.", updated); |
| 566 | onLeadershipEvent(new LeadershipEvent(LeadershipEvent.Type.LEADER_ELECTED, updated)); |
Madan Jampani | 096da0e | 2015-06-04 14:30:19 -0700 | [diff] [blame] | 567 | } |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 568 | }); |
| 569 | } catch (Exception e) { |
Madan Jampani | 6f38daf | 2015-06-04 15:26:38 -0700 | [diff] [blame] | 570 | log.debug("Failed to refresh leader board", e); |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 571 | } |
Madan Jampani | d14166a | 2015-02-24 17:37:51 -0800 | [diff] [blame] | 572 | } |
Madan Jampani | cc58675 | 2015-05-18 16:19:27 -0700 | [diff] [blame] | 573 | |
| 574 | private class InternalClusterEventListener implements ClusterEventListener { |
| 575 | |
| 576 | @Override |
| 577 | public void event(ClusterEvent event) { |
| 578 | if (event.type() == Type.INSTANCE_DEACTIVATED || event.type() == Type.INSTANCE_REMOVED) { |
| 579 | scheduleStaleLeadershipPurge(0); |
| 580 | } |
| 581 | } |
| 582 | } |
Madan Jampani | a14047d | 2015-02-25 12:23:02 -0800 | [diff] [blame] | 583 | } |