Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 1 | package org.onosproject.store.service.impl; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 2 | |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 3 | import static java.util.concurrent.Executors.newSingleThreadScheduledExecutor; |
| 4 | import static org.onlab.util.Tools.namedThreads; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 5 | import static org.slf4j.LoggerFactory.getLogger; |
| 6 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 7 | import java.io.File; |
| 8 | import java.io.IOException; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 9 | import java.util.Collection; |
| 10 | import java.util.Collections; |
| 11 | import java.util.HashSet; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 12 | import java.util.Map; |
Yuta HIGUCHI | f0f2dfc | 2014-11-26 13:59:07 -0800 | [diff] [blame] | 13 | import java.util.Optional; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 14 | import java.util.Set; |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 15 | import java.util.concurrent.CountDownLatch; |
Yuta HIGUCHI | 39da979 | 2014-11-14 02:07:04 -0800 | [diff] [blame] | 16 | import java.util.concurrent.ExecutionException; |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 17 | import java.util.concurrent.ScheduledExecutorService; |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 18 | import java.util.concurrent.TimeUnit; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 19 | |
| 20 | import net.kuujo.copycat.Copycat; |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 21 | import net.kuujo.copycat.CopycatConfig; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 22 | import net.kuujo.copycat.cluster.ClusterConfig; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 23 | import net.kuujo.copycat.cluster.Member; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 24 | import net.kuujo.copycat.cluster.TcpCluster; |
| 25 | import net.kuujo.copycat.cluster.TcpClusterConfig; |
| 26 | import net.kuujo.copycat.cluster.TcpMember; |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 27 | import net.kuujo.copycat.event.EventHandler; |
Madan Jampani | f5d263b | 2014-11-13 10:04:40 -0800 | [diff] [blame] | 28 | import net.kuujo.copycat.event.LeaderElectEvent; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 29 | import net.kuujo.copycat.log.Log; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 30 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 31 | import org.apache.felix.scr.annotations.Activate; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 32 | import org.apache.felix.scr.annotations.Component; |
Yuta HIGUCHI | 657626e | 2014-11-04 20:54:58 -0800 | [diff] [blame] | 33 | import org.apache.felix.scr.annotations.Deactivate; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 34 | import org.apache.felix.scr.annotations.Reference; |
| 35 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 36 | import org.apache.felix.scr.annotations.Service; |
Brian O'Connor | abafb50 | 2014-12-02 22:26:20 -0800 | [diff] [blame] | 37 | import org.onosproject.cluster.ClusterEvent; |
| 38 | import org.onosproject.cluster.ClusterEventListener; |
| 39 | import org.onosproject.cluster.ClusterService; |
| 40 | import org.onosproject.cluster.ControllerNode; |
| 41 | import org.onosproject.cluster.DefaultControllerNode; |
| 42 | import org.onosproject.store.cluster.messaging.ClusterCommunicationService; |
| 43 | import org.onosproject.store.cluster.messaging.ClusterMessage; |
| 44 | import org.onosproject.store.cluster.messaging.MessageSubject; |
| 45 | import org.onosproject.store.service.BatchReadRequest; |
| 46 | import org.onosproject.store.service.BatchReadResult; |
| 47 | import org.onosproject.store.service.BatchWriteRequest; |
| 48 | import org.onosproject.store.service.BatchWriteResult; |
| 49 | import org.onosproject.store.service.DatabaseAdminService; |
| 50 | import org.onosproject.store.service.DatabaseException; |
| 51 | import org.onosproject.store.service.DatabaseService; |
| 52 | import org.onosproject.store.service.ReadResult; |
| 53 | import org.onosproject.store.service.ReadStatus; |
| 54 | import org.onosproject.store.service.VersionedValue; |
| 55 | import org.onosproject.store.service.WriteResult; |
| 56 | import org.onosproject.store.service.WriteStatus; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 57 | import org.onlab.packet.IpAddress; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 58 | import org.slf4j.Logger; |
| 59 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 60 | import com.google.common.collect.ImmutableList; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 61 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 62 | /** |
| 63 | * Strongly consistent and durable state management service based on |
| 64 | * Copycat implementation of Raft consensus protocol. |
| 65 | */ |
Brian O'Connor | 5d55ed4 | 2014-12-01 18:27:47 -0800 | [diff] [blame] | 66 | @Component(immediate = false) |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 67 | @Service |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 68 | public class DatabaseManager implements DatabaseService, DatabaseAdminService { |
| 69 | |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 70 | private static final int RETRY_MS = 500; |
| 71 | |
| 72 | private static final int ACTIVATE_MAX_RETRIES = 100; |
| 73 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 74 | private final Logger log = getLogger(getClass()); |
| 75 | |
| 76 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Yuta HIGUCHI | 5001ba9 | 2014-11-04 21:33:54 -0800 | [diff] [blame] | 77 | protected ClusterService clusterService; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 78 | |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 79 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 80 | protected ClusterCommunicationService clusterCommunicator; |
| 81 | |
| 82 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Yuta HIGUCHI | 0c1c100 | 2014-11-05 13:47:25 -0800 | [diff] [blame] | 83 | protected DatabaseProtocolService copycatMessagingProtocol; |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 84 | |
Yuta HIGUCHI | 43e3a7e | 2014-11-30 23:22:11 -0800 | [diff] [blame] | 85 | public static final String LOG_FILE_PREFIX = "raft/onos-copy-cat-log_"; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 86 | |
| 87 | // Current working dir seems to be /opt/onos/apache-karaf-3.0.2 |
Pavlin Radoslavov | 190f8f9 | 2014-11-11 15:56:14 -0800 | [diff] [blame] | 88 | // TODO: Set the path to /opt/onos/config |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 89 | private static final String CONFIG_DIR = "../config"; |
| 90 | |
| 91 | private static final String DEFAULT_MEMBER_FILE = "tablets.json"; |
| 92 | |
| 93 | private static final String DEFAULT_TABLET = "default"; |
| 94 | |
| 95 | // TODO: make this configurable |
| 96 | // initial member configuration file path |
| 97 | private String initialMemberConfig = DEFAULT_MEMBER_FILE; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 98 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 99 | public static final MessageSubject RAFT_LEADER_ELECTION_EVENT = |
| 100 | new MessageSubject("raft-leader-election-event"); |
| 101 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 102 | private Copycat copycat; |
| 103 | private DatabaseClient client; |
| 104 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 105 | // guarded by synchronized block |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 106 | private ClusterConfig<TcpMember> clusterConfig; |
| 107 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 108 | private CountDownLatch clusterEventLatch; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 109 | private ClusterEventListener clusterEventListener; |
| 110 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 111 | private Map<String, Set<DefaultControllerNode>> tabletMembers; |
| 112 | |
| 113 | private boolean autoAddMember = false; |
| 114 | |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 115 | private ScheduledExecutorService executor; |
| 116 | |
| 117 | private volatile LeaderElectEvent myLeaderEvent = null; |
| 118 | |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 119 | // TODO make this configurable |
Madan Jampani | 4bb70c6 | 2014-11-25 23:47:12 -0800 | [diff] [blame] | 120 | private int maxLogSizeBytes = 128 * (1024 * 1024); |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 121 | |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 122 | // TODO make this configurable |
| 123 | private long electionTimeoutMs = 5000; // CopyCat default: 2000 |
| 124 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 125 | @Activate |
Yuta HIGUCHI | 39da979 | 2014-11-14 02:07:04 -0800 | [diff] [blame] | 126 | public void activate() throws InterruptedException, ExecutionException { |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 127 | |
Yuta HIGUCHI | 43e3a7e | 2014-11-30 23:22:11 -0800 | [diff] [blame] | 128 | // KARAF_DATA |
| 129 | // http://karaf.apache.org/manual/latest/users-guide/start-stop.html |
| 130 | final String dataDir = System.getProperty("karaf.data", "./data"); |
| 131 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 132 | // load tablet configuration |
| 133 | File file = new File(CONFIG_DIR, initialMemberConfig); |
| 134 | log.info("Loading config: {}", file.getAbsolutePath()); |
| 135 | TabletDefinitionStore tabletDef = new TabletDefinitionStore(file); |
| 136 | try { |
| 137 | tabletMembers = tabletDef.read(); |
| 138 | } catch (IOException e) { |
| 139 | log.error("Failed to load tablet config {}", file); |
| 140 | throw new IllegalStateException("Failed to load tablet config", e); |
| 141 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 142 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 143 | // load default tablet configuration and start copycat |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 144 | clusterConfig = new TcpClusterConfig(); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 145 | Set<DefaultControllerNode> defaultMembers = tabletMembers.get(DEFAULT_TABLET); |
| 146 | if (defaultMembers == null || defaultMembers.isEmpty()) { |
| 147 | log.error("No members found in [{}] tablet configuration.", |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 148 | DEFAULT_TABLET); |
| 149 | throw new IllegalStateException("No member found in tablet configuration"); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 150 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 151 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 152 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 153 | final ControllerNode localNode = clusterService.getLocalNode(); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 154 | for (ControllerNode member : defaultMembers) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 155 | final TcpMember tcpMember = new TcpMember(member.ip().toString(), |
| 156 | member.tcpPort()); |
| 157 | if (localNode.equals(member)) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 158 | clusterConfig.setLocalMember(tcpMember); |
| 159 | } else { |
| 160 | clusterConfig.addRemoteMember(tcpMember); |
| 161 | } |
| 162 | } |
| 163 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 164 | if (clusterConfig.getLocalMember() != null) { |
| 165 | |
| 166 | // Wait for a minimum viable Raft cluster to boot up. |
| 167 | waitForClusterQuorum(); |
| 168 | |
| 169 | final TcpCluster cluster; |
| 170 | synchronized (clusterConfig) { |
| 171 | // Create the cluster. |
| 172 | cluster = new TcpCluster(clusterConfig); |
| 173 | } |
| 174 | log.info("Starting cluster: {}", cluster); |
| 175 | |
| 176 | DatabaseEntryExpirationTracker expirationTracker = |
| 177 | new DatabaseEntryExpirationTracker( |
| 178 | clusterConfig.getLocalMember(), |
| 179 | clusterService.getLocalNode(), |
| 180 | clusterCommunicator, |
| 181 | this); |
| 182 | |
| 183 | DatabaseStateMachine stateMachine = new DatabaseStateMachine(); |
| 184 | stateMachine.addEventListener(expirationTracker); |
Yuta HIGUCHI | 43e3a7e | 2014-11-30 23:22:11 -0800 | [diff] [blame] | 185 | Log consensusLog = new MapDBLog(dataDir + "/" + LOG_FILE_PREFIX + localNode.id(), |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 186 | ClusterMessagingProtocol.DB_SERIALIZER); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 187 | |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 188 | CopycatConfig ccConfig = new CopycatConfig(); |
| 189 | ccConfig.setMaxLogSize(maxLogSizeBytes); |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 190 | ccConfig.setElectionTimeout(electionTimeoutMs); |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 191 | |
| 192 | copycat = new Copycat(stateMachine, consensusLog, cluster, copycatMessagingProtocol, ccConfig); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 193 | copycat.event(LeaderElectEvent.class).registerHandler(new RaftLeaderElectionMonitor()); |
| 194 | copycat.event(LeaderElectEvent.class).registerHandler(expirationTracker); |
| 195 | } |
| 196 | |
| 197 | client = new DatabaseClient(copycatMessagingProtocol); |
| 198 | clusterCommunicator.addSubscriber(RAFT_LEADER_ELECTION_EVENT, client); |
| 199 | |
| 200 | // Starts copycat if this node is a participant |
| 201 | // of the Raft cluster. |
| 202 | if (copycat != null) { |
| 203 | copycat.start().get(); |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 204 | |
| 205 | executor = |
| 206 | newSingleThreadScheduledExecutor(namedThreads("db-heartbeat-%d")); |
| 207 | executor.scheduleWithFixedDelay(new LeaderAdvertiser(), 5, 2, TimeUnit.SECONDS); |
| 208 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 209 | } |
| 210 | |
| 211 | client.waitForLeader(); |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 212 | |
| 213 | // Try and list the tables to verify database manager is |
| 214 | // in a state where it can serve requests. |
| 215 | tryTableListing(); |
| 216 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 217 | log.info("Started."); |
| 218 | } |
| 219 | |
| 220 | @Deactivate |
| 221 | public void deactivate() { |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 222 | if (executor != null) { |
| 223 | executor.shutdownNow(); |
| 224 | } |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 225 | clusterService.removeListener(clusterEventListener); |
| 226 | // TODO: ClusterCommunicationService must support more than one |
| 227 | // handler per message subject. |
| 228 | clusterCommunicator.removeSubscriber(RAFT_LEADER_ELECTION_EVENT); |
| 229 | if (copycat != null) { |
| 230 | copycat.stop(); |
| 231 | } |
| 232 | log.info("Stopped."); |
| 233 | } |
| 234 | |
| 235 | private void waitForClusterQuorum() { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 236 | // note: from this point beyond, clusterConfig requires synchronization |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 237 | clusterEventLatch = new CountDownLatch(1); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 238 | clusterEventListener = new InternalClusterEventListener(); |
| 239 | clusterService.addListener(clusterEventListener); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 240 | |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 241 | final int raftClusterSize = clusterConfig.getMembers().size(); |
| 242 | final int raftClusterQuorumSize = (int) (Math.floor(raftClusterSize / 2)) + 1; |
| 243 | if (clusterService.getNodes().size() < raftClusterQuorumSize) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 244 | // current cluster size smaller then expected |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 245 | try { |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 246 | final int waitTimeSec = 120; |
| 247 | log.info("Waiting for a maximum of {}s for raft cluster quorum to boot up...", waitTimeSec); |
| 248 | if (!clusterEventLatch.await(waitTimeSec, TimeUnit.SECONDS)) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 249 | log.info("Starting with {}/{} nodes cluster", |
| 250 | clusterService.getNodes().size(), |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 251 | raftClusterSize); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 252 | } |
| 253 | } catch (InterruptedException e) { |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 254 | log.info("Interrupted waiting for raft quorum.", e); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 255 | } |
| 256 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 257 | } |
| 258 | |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 259 | private void tryTableListing() throws InterruptedException { |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 260 | int retries = 0; |
| 261 | do { |
| 262 | try { |
| 263 | listTables(); |
| 264 | return; |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 265 | } catch (DatabaseException.Timeout e) { |
| 266 | log.debug("Failed to listTables. Will retry...", e); |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 267 | } catch (DatabaseException e) { |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 268 | log.debug("Failed to listTables. Will retry later...", e); |
| 269 | Thread.sleep(RETRY_MS); |
Yuta HIGUCHI | 0136562 | 2014-11-25 17:35:15 -0800 | [diff] [blame] | 270 | } |
| 271 | if (retries == ACTIVATE_MAX_RETRIES) { |
| 272 | log.error("Failed to listTables after multiple attempts. Giving up."); |
| 273 | // Exiting hoping things will be fixed by the time |
| 274 | // others start using the service |
| 275 | return; |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 276 | } |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 277 | retries++; |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 278 | } while (true); |
| 279 | } |
| 280 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 281 | @Override |
| 282 | public boolean createTable(String name) { |
| 283 | return client.createTable(name); |
| 284 | } |
| 285 | |
| 286 | @Override |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 287 | public boolean createTable(String name, int ttlMillis) { |
| 288 | return client.createTable(name, ttlMillis); |
| 289 | } |
| 290 | |
| 291 | @Override |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 292 | public void dropTable(String name) { |
| 293 | client.dropTable(name); |
| 294 | } |
| 295 | |
| 296 | @Override |
| 297 | public void dropAllTables() { |
| 298 | client.dropAllTables(); |
| 299 | } |
| 300 | |
| 301 | @Override |
Madan Jampani | f5d263b | 2014-11-13 10:04:40 -0800 | [diff] [blame] | 302 | public Set<String> listTables() { |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 303 | return client.listTables(); |
| 304 | } |
| 305 | |
| 306 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 307 | public VersionedValue get(String tableName, String key) { |
| 308 | BatchReadRequest batchRequest = new BatchReadRequest.Builder().get(tableName, key).build(); |
| 309 | ReadResult readResult = batchRead(batchRequest).getAsList().get(0); |
| 310 | if (readResult.status().equals(ReadStatus.OK)) { |
| 311 | return readResult.value(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 312 | } |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 313 | throw new DatabaseException("get failed due to status: " + readResult.status()); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 314 | } |
| 315 | |
| 316 | @Override |
Yuta HIGUCHI | 841c0b6 | 2014-11-13 20:27:14 -0800 | [diff] [blame] | 317 | public Map<String, VersionedValue> getAll(String tableName) { |
| 318 | return client.getAll(tableName); |
| 319 | } |
| 320 | |
| 321 | |
| 322 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 323 | public BatchReadResult batchRead(BatchReadRequest batchRequest) { |
| 324 | return new BatchReadResult(client.batchRead(batchRequest)); |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 325 | } |
| 326 | |
| 327 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 328 | public BatchWriteResult batchWrite(BatchWriteRequest batchRequest) { |
| 329 | return new BatchWriteResult(client.batchWrite(batchRequest)); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 330 | } |
| 331 | |
| 332 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 333 | public VersionedValue put(String tableName, String key, byte[] value) { |
| 334 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder().put(tableName, key, value).build(); |
| 335 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 336 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 337 | return writeResult.previousValue(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 338 | } |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 339 | throw new DatabaseException("put failed due to status: " + writeResult.status()); |
| 340 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 341 | |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 342 | @Override |
| 343 | public boolean putIfAbsent(String tableName, String key, byte[] value) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 344 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 345 | .putIfAbsent(tableName, key, value).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 346 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 347 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 348 | return true; |
| 349 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 350 | return false; |
| 351 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 352 | throw new DatabaseException("putIfAbsent failed due to status: " |
| 353 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 354 | } |
| 355 | |
| 356 | @Override |
| 357 | public boolean putIfVersionMatches(String tableName, String key, |
| 358 | byte[] value, long version) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 359 | BatchWriteRequest batchRequest = |
| 360 | new BatchWriteRequest.Builder() |
| 361 | .putIfVersionMatches(tableName, key, value, version).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 362 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 363 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 364 | return true; |
| 365 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 366 | return false; |
| 367 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 368 | throw new DatabaseException("putIfVersionMatches failed due to status: " |
| 369 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 370 | } |
| 371 | |
| 372 | @Override |
| 373 | public boolean putIfValueMatches(String tableName, String key, |
| 374 | byte[] oldValue, byte[] newValue) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 375 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 376 | .putIfValueMatches(tableName, key, oldValue, newValue).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 377 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 378 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 379 | return true; |
| 380 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 381 | return false; |
| 382 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 383 | throw new DatabaseException("putIfValueMatches failed due to status: " |
| 384 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 385 | } |
| 386 | |
| 387 | @Override |
| 388 | public VersionedValue remove(String tableName, String key) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 389 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 390 | .remove(tableName, key).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 391 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 392 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 393 | return writeResult.previousValue(); |
| 394 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 395 | throw new DatabaseException("remove failed due to status: " |
| 396 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 397 | } |
| 398 | |
| 399 | @Override |
| 400 | public boolean removeIfVersionMatches(String tableName, String key, |
| 401 | long version) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 402 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 403 | .removeIfVersionMatches(tableName, key, version).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 404 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 405 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 406 | return true; |
| 407 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 408 | return false; |
| 409 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 410 | throw new DatabaseException("removeIfVersionMatches failed due to status: " |
| 411 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 412 | } |
| 413 | |
| 414 | @Override |
| 415 | public boolean removeIfValueMatches(String tableName, String key, |
| 416 | byte[] value) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 417 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 418 | .removeIfValueMatches(tableName, key, value).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 419 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 420 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 421 | return true; |
| 422 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 423 | return false; |
| 424 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 425 | throw new DatabaseException("removeIfValueMatches failed due to status: " |
| 426 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 427 | } |
| 428 | |
| 429 | @Override |
| 430 | public void addMember(final ControllerNode node) { |
| 431 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
| 432 | node.tcpPort()); |
| 433 | log.info("{} was added to the cluster", tcpMember); |
| 434 | synchronized (clusterConfig) { |
| 435 | clusterConfig.addRemoteMember(tcpMember); |
| 436 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 437 | } |
| 438 | |
Yuta HIGUCHI | f0f2dfc | 2014-11-26 13:59:07 -0800 | [diff] [blame] | 439 | @Override |
| 440 | public Optional<ControllerNode> leader() { |
| 441 | if (copycat != null) { |
| 442 | if (copycat.isLeader()) { |
| 443 | return Optional.of(clusterService.getLocalNode()); |
| 444 | } |
| 445 | Member leader = copycat.cluster().remoteMember(copycat.leader()); |
| 446 | return Optional.ofNullable(getNodeIdFromMember(leader)); |
| 447 | } |
| 448 | return Optional.ofNullable(getNodeIdFromMember(client.getCurrentLeader())); |
| 449 | } |
| 450 | |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 451 | private final class LeaderAdvertiser implements Runnable { |
| 452 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 453 | @Override |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 454 | public void run() { |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 455 | try { |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 456 | LeaderElectEvent event = myLeaderEvent; |
| 457 | if (event != null) { |
| 458 | log.trace("Broadcasting RAFT_LEADER_ELECTION_EVENT: {}", event); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 459 | // This node just became the leader. |
| 460 | clusterCommunicator.broadcastIncludeSelf( |
| 461 | new ClusterMessage( |
| 462 | clusterService.getLocalNode().id(), |
| 463 | RAFT_LEADER_ELECTION_EVENT, |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 464 | ClusterMessagingProtocol.DB_SERIALIZER.encode(event))); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 465 | } |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 466 | } catch (Exception e) { |
| 467 | log.debug("LeaderAdvertiser failed with exception", e); |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | } |
| 472 | |
| 473 | private final class RaftLeaderElectionMonitor implements EventHandler<LeaderElectEvent> { |
| 474 | @Override |
| 475 | public void handle(LeaderElectEvent event) { |
| 476 | try { |
| 477 | log.debug("Received LeaderElectEvent: {}", event); |
| 478 | if (clusterConfig.getLocalMember() != null && event.leader().equals(clusterConfig.getLocalMember())) { |
| 479 | log.debug("Broadcasting RAFT_LEADER_ELECTION_EVENT"); |
| 480 | myLeaderEvent = event; |
| 481 | // This node just became the leader. |
| 482 | clusterCommunicator.broadcastIncludeSelf( |
| 483 | new ClusterMessage( |
| 484 | clusterService.getLocalNode().id(), |
| 485 | RAFT_LEADER_ELECTION_EVENT, |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 486 | ClusterMessagingProtocol.DB_SERIALIZER.encode(event))); |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 487 | } else { |
| 488 | if (myLeaderEvent != null) { |
| 489 | log.debug("This node is no longer the Leader"); |
| 490 | } |
| 491 | myLeaderEvent = null; |
| 492 | } |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 493 | } catch (IOException e) { |
| 494 | log.error("Failed to broadcast raft leadership change event", e); |
| 495 | } |
| 496 | } |
| 497 | } |
| 498 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 499 | private final class InternalClusterEventListener |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 500 | implements ClusterEventListener { |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 501 | |
| 502 | @Override |
| 503 | public void event(ClusterEvent event) { |
| 504 | // TODO: Not every node should be part of the consensus ring. |
| 505 | |
| 506 | final ControllerNode node = event.subject(); |
| 507 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 508 | node.tcpPort()); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 509 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 510 | switch (event.type()) { |
| 511 | case INSTANCE_ACTIVATED: |
| 512 | case INSTANCE_ADDED: |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 513 | if (autoAddMember) { |
| 514 | synchronized (clusterConfig) { |
| 515 | if (!clusterConfig.getMembers().contains(tcpMember)) { |
| 516 | log.info("{} was automatically added to the cluster", tcpMember); |
| 517 | clusterConfig.addRemoteMember(tcpMember); |
| 518 | } |
| 519 | } |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 520 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 521 | break; |
| 522 | case INSTANCE_DEACTIVATED: |
| 523 | case INSTANCE_REMOVED: |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 524 | if (autoAddMember) { |
| 525 | Set<DefaultControllerNode> members |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 526 | = tabletMembers.getOrDefault(DEFAULT_TABLET, |
| 527 | Collections.emptySet()); |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 528 | // remove only if not the initial members |
| 529 | if (!members.contains(node)) { |
| 530 | synchronized (clusterConfig) { |
| 531 | if (clusterConfig.getMembers().contains(tcpMember)) { |
| 532 | log.info("{} was automatically removed from the cluster", tcpMember); |
| 533 | clusterConfig.removeRemoteMember(tcpMember); |
| 534 | } |
| 535 | } |
| 536 | } |
| 537 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 538 | break; |
| 539 | default: |
| 540 | break; |
| 541 | } |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 542 | if (copycat != null) { |
| 543 | log.debug("Current cluster: {}", copycat.cluster()); |
| 544 | } |
| 545 | clusterEventLatch.countDown(); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 546 | } |
| 547 | |
| 548 | } |
| 549 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 550 | @Override |
| 551 | public void removeMember(final ControllerNode node) { |
| 552 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
| 553 | node.tcpPort()); |
| 554 | log.info("{} was removed from the cluster", tcpMember); |
| 555 | synchronized (clusterConfig) { |
| 556 | clusterConfig.removeRemoteMember(tcpMember); |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | @Override |
| 561 | public Collection<ControllerNode> listMembers() { |
| 562 | if (copycat == null) { |
| 563 | return ImmutableList.of(); |
| 564 | } |
| 565 | Set<ControllerNode> members = new HashSet<>(); |
| 566 | for (Member member : copycat.cluster().members()) { |
Yuta HIGUCHI | f0f2dfc | 2014-11-26 13:59:07 -0800 | [diff] [blame] | 567 | ControllerNode node = getNodeIdFromMember(member); |
| 568 | if (node == null) { |
| 569 | log.info("No Node found for {}", member); |
| 570 | continue; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 571 | } |
Yuta HIGUCHI | f0f2dfc | 2014-11-26 13:59:07 -0800 | [diff] [blame] | 572 | members.add(node); |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 573 | } |
| 574 | return members; |
| 575 | } |
| 576 | |
Yuta HIGUCHI | f0f2dfc | 2014-11-26 13:59:07 -0800 | [diff] [blame] | 577 | private ControllerNode getNodeIdFromMember(Member member) { |
| 578 | if (member instanceof TcpMember) { |
| 579 | final TcpMember tcpMember = (TcpMember) member; |
| 580 | // TODO assuming tcpMember#host to be IP address, |
| 581 | // but if not lookup DNS, etc. first |
| 582 | IpAddress ip = IpAddress.valueOf(tcpMember.host()); |
| 583 | int tcpPort = tcpMember.port(); |
| 584 | for (ControllerNode node : clusterService.getNodes()) { |
| 585 | if (node.ip().equals(ip) && |
| 586 | node.tcpPort() == tcpPort) { |
| 587 | return node; |
| 588 | } |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 589 | } |
| 590 | } |
| 591 | return null; |
| 592 | } |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 593 | } |