Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 1 | package org.onlab.onos.store.service.impl; |
| 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; |
| 13 | import java.util.Set; |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 14 | import java.util.concurrent.CountDownLatch; |
Yuta HIGUCHI | 39da979 | 2014-11-14 02:07:04 -0800 | [diff] [blame] | 15 | import java.util.concurrent.ExecutionException; |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 16 | import java.util.concurrent.ScheduledExecutorService; |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 17 | import java.util.concurrent.TimeUnit; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 18 | |
| 19 | import net.kuujo.copycat.Copycat; |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 20 | import net.kuujo.copycat.CopycatConfig; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 21 | import net.kuujo.copycat.cluster.ClusterConfig; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 22 | import net.kuujo.copycat.cluster.Member; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 23 | import net.kuujo.copycat.cluster.TcpCluster; |
| 24 | import net.kuujo.copycat.cluster.TcpClusterConfig; |
| 25 | import net.kuujo.copycat.cluster.TcpMember; |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 26 | import net.kuujo.copycat.event.EventHandler; |
Madan Jampani | f5d263b | 2014-11-13 10:04:40 -0800 | [diff] [blame] | 27 | import net.kuujo.copycat.event.LeaderElectEvent; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 28 | import net.kuujo.copycat.log.Log; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 29 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 30 | import org.apache.felix.scr.annotations.Activate; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 31 | import org.apache.felix.scr.annotations.Component; |
Yuta HIGUCHI | 657626e | 2014-11-04 20:54:58 -0800 | [diff] [blame] | 32 | import org.apache.felix.scr.annotations.Deactivate; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 33 | import org.apache.felix.scr.annotations.Reference; |
| 34 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 35 | import org.apache.felix.scr.annotations.Service; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 36 | import org.onlab.onos.cluster.ClusterEvent; |
| 37 | import org.onlab.onos.cluster.ClusterEventListener; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 38 | import org.onlab.onos.cluster.ClusterService; |
| 39 | import org.onlab.onos.cluster.ControllerNode; |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 40 | import org.onlab.onos.cluster.DefaultControllerNode; |
| 41 | import org.onlab.onos.cluster.NodeId; |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 42 | import org.onlab.onos.store.cluster.messaging.ClusterCommunicationService; |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 43 | import org.onlab.onos.store.cluster.messaging.ClusterMessage; |
| 44 | import org.onlab.onos.store.cluster.messaging.MessageSubject; |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 45 | import org.onlab.onos.store.service.BatchReadRequest; |
| 46 | import org.onlab.onos.store.service.BatchReadResult; |
| 47 | import org.onlab.onos.store.service.BatchWriteRequest; |
| 48 | import org.onlab.onos.store.service.BatchWriteResult; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 49 | import org.onlab.onos.store.service.DatabaseAdminService; |
| 50 | import org.onlab.onos.store.service.DatabaseException; |
| 51 | import org.onlab.onos.store.service.DatabaseService; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 52 | import org.onlab.onos.store.service.ReadResult; |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 53 | import org.onlab.onos.store.service.ReadStatus; |
| 54 | import org.onlab.onos.store.service.VersionedValue; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 55 | import org.onlab.onos.store.service.WriteResult; |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 56 | import org.onlab.onos.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 | */ |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 66 | @Component(immediate = true) |
| 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 | 13a6f5a | 2014-11-12 10:07:47 -0800 | [diff] [blame] | 85 | // FIXME: point to appropriate path |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 86 | public static final String LOG_FILE_PREFIX = "/tmp/onos-copy-cat-log_"; |
| 87 | |
| 88 | // 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] | 89 | // TODO: Set the path to /opt/onos/config |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 90 | private static final String CONFIG_DIR = "../config"; |
| 91 | |
| 92 | private static final String DEFAULT_MEMBER_FILE = "tablets.json"; |
| 93 | |
| 94 | private static final String DEFAULT_TABLET = "default"; |
| 95 | |
| 96 | // TODO: make this configurable |
| 97 | // initial member configuration file path |
| 98 | private String initialMemberConfig = DEFAULT_MEMBER_FILE; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 99 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 100 | public static final MessageSubject RAFT_LEADER_ELECTION_EVENT = |
| 101 | new MessageSubject("raft-leader-election-event"); |
| 102 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 103 | private Copycat copycat; |
| 104 | private DatabaseClient client; |
| 105 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 106 | // guarded by synchronized block |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 107 | private ClusterConfig<TcpMember> clusterConfig; |
| 108 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 109 | private CountDownLatch clusterEventLatch; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 110 | private ClusterEventListener clusterEventListener; |
| 111 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 112 | private Map<String, Set<DefaultControllerNode>> tabletMembers; |
| 113 | |
| 114 | private boolean autoAddMember = false; |
| 115 | |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 116 | private ScheduledExecutorService executor; |
| 117 | |
| 118 | private volatile LeaderElectEvent myLeaderEvent = null; |
| 119 | |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 120 | // TODO make this configurable |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 121 | private int maxLogSizeBytes = 128 * (1024 ^ 2); |
| 122 | |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 123 | // TODO make this configurable |
| 124 | private long electionTimeoutMs = 5000; // CopyCat default: 2000 |
| 125 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 126 | @Activate |
Yuta HIGUCHI | 39da979 | 2014-11-14 02:07:04 -0800 | [diff] [blame] | 127 | public void activate() throws InterruptedException, ExecutionException { |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 128 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 129 | // load tablet configuration |
| 130 | File file = new File(CONFIG_DIR, initialMemberConfig); |
| 131 | log.info("Loading config: {}", file.getAbsolutePath()); |
| 132 | TabletDefinitionStore tabletDef = new TabletDefinitionStore(file); |
| 133 | try { |
| 134 | tabletMembers = tabletDef.read(); |
| 135 | } catch (IOException e) { |
| 136 | log.error("Failed to load tablet config {}", file); |
| 137 | throw new IllegalStateException("Failed to load tablet config", e); |
| 138 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 139 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 140 | // load default tablet configuration and start copycat |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 141 | clusterConfig = new TcpClusterConfig(); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 142 | Set<DefaultControllerNode> defaultMembers = tabletMembers.get(DEFAULT_TABLET); |
| 143 | if (defaultMembers == null || defaultMembers.isEmpty()) { |
| 144 | log.error("No members found in [{}] tablet configuration.", |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 145 | DEFAULT_TABLET); |
| 146 | throw new IllegalStateException("No member found in tablet configuration"); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 147 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 148 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 149 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 150 | final ControllerNode localNode = clusterService.getLocalNode(); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 151 | for (ControllerNode member : defaultMembers) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 152 | final TcpMember tcpMember = new TcpMember(member.ip().toString(), |
| 153 | member.tcpPort()); |
| 154 | if (localNode.equals(member)) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 155 | clusterConfig.setLocalMember(tcpMember); |
| 156 | } else { |
| 157 | clusterConfig.addRemoteMember(tcpMember); |
| 158 | } |
| 159 | } |
| 160 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 161 | if (clusterConfig.getLocalMember() != null) { |
| 162 | |
| 163 | // Wait for a minimum viable Raft cluster to boot up. |
| 164 | waitForClusterQuorum(); |
| 165 | |
| 166 | final TcpCluster cluster; |
| 167 | synchronized (clusterConfig) { |
| 168 | // Create the cluster. |
| 169 | cluster = new TcpCluster(clusterConfig); |
| 170 | } |
| 171 | log.info("Starting cluster: {}", cluster); |
| 172 | |
| 173 | DatabaseEntryExpirationTracker expirationTracker = |
| 174 | new DatabaseEntryExpirationTracker( |
| 175 | clusterConfig.getLocalMember(), |
| 176 | clusterService.getLocalNode(), |
| 177 | clusterCommunicator, |
| 178 | this); |
| 179 | |
| 180 | DatabaseStateMachine stateMachine = new DatabaseStateMachine(); |
| 181 | stateMachine.addEventListener(expirationTracker); |
| 182 | Log consensusLog = new MapDBLog(LOG_FILE_PREFIX + localNode.id(), |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 183 | ClusterMessagingProtocol.DB_SERIALIZER); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 184 | |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 185 | CopycatConfig ccConfig = new CopycatConfig(); |
| 186 | ccConfig.setMaxLogSize(maxLogSizeBytes); |
Yuta HIGUCHI | d88400b | 2014-11-25 12:13:30 -0800 | [diff] [blame] | 187 | ccConfig.setElectionTimeout(electionTimeoutMs); |
Yuta HIGUCHI | 4e45081 | 2014-11-23 01:53:28 -0800 | [diff] [blame] | 188 | |
| 189 | copycat = new Copycat(stateMachine, consensusLog, cluster, copycatMessagingProtocol, ccConfig); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 190 | copycat.event(LeaderElectEvent.class).registerHandler(new RaftLeaderElectionMonitor()); |
| 191 | copycat.event(LeaderElectEvent.class).registerHandler(expirationTracker); |
| 192 | } |
| 193 | |
| 194 | client = new DatabaseClient(copycatMessagingProtocol); |
| 195 | clusterCommunicator.addSubscriber(RAFT_LEADER_ELECTION_EVENT, client); |
| 196 | |
| 197 | // Starts copycat if this node is a participant |
| 198 | // of the Raft cluster. |
| 199 | if (copycat != null) { |
| 200 | copycat.start().get(); |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 201 | |
| 202 | executor = |
| 203 | newSingleThreadScheduledExecutor(namedThreads("db-heartbeat-%d")); |
| 204 | executor.scheduleWithFixedDelay(new LeaderAdvertiser(), 5, 2, TimeUnit.SECONDS); |
| 205 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 206 | } |
| 207 | |
| 208 | client.waitForLeader(); |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 209 | |
| 210 | // Try and list the tables to verify database manager is |
| 211 | // in a state where it can serve requests. |
| 212 | tryTableListing(); |
| 213 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 214 | log.info("Started."); |
| 215 | } |
| 216 | |
| 217 | @Deactivate |
| 218 | public void deactivate() { |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 219 | if (executor != null) { |
| 220 | executor.shutdownNow(); |
| 221 | } |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 222 | clusterService.removeListener(clusterEventListener); |
| 223 | // TODO: ClusterCommunicationService must support more than one |
| 224 | // handler per message subject. |
| 225 | clusterCommunicator.removeSubscriber(RAFT_LEADER_ELECTION_EVENT); |
| 226 | if (copycat != null) { |
| 227 | copycat.stop(); |
| 228 | } |
| 229 | log.info("Stopped."); |
| 230 | } |
| 231 | |
| 232 | private void waitForClusterQuorum() { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 233 | // note: from this point beyond, clusterConfig requires synchronization |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 234 | clusterEventLatch = new CountDownLatch(1); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 235 | clusterEventListener = new InternalClusterEventListener(); |
| 236 | clusterService.addListener(clusterEventListener); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 237 | |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 238 | final int raftClusterSize = clusterConfig.getMembers().size(); |
| 239 | final int raftClusterQuorumSize = (int) (Math.floor(raftClusterSize / 2)) + 1; |
| 240 | if (clusterService.getNodes().size() < raftClusterQuorumSize) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 241 | // current cluster size smaller then expected |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 242 | try { |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 243 | final int waitTimeSec = 120; |
| 244 | log.info("Waiting for a maximum of {}s for raft cluster quorum to boot up...", waitTimeSec); |
| 245 | if (!clusterEventLatch.await(waitTimeSec, TimeUnit.SECONDS)) { |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 246 | log.info("Starting with {}/{} nodes cluster", |
| 247 | clusterService.getNodes().size(), |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 248 | raftClusterSize); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 249 | } |
| 250 | } catch (InterruptedException e) { |
Pavlin Radoslavov | af529d8 | 2014-11-22 21:03:19 -0800 | [diff] [blame] | 251 | log.info("Interrupted waiting for raft quorum.", e); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 252 | } |
| 253 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 254 | } |
| 255 | |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 256 | private void tryTableListing() throws InterruptedException { |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 257 | int retries = 0; |
| 258 | do { |
| 259 | try { |
| 260 | listTables(); |
| 261 | return; |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 262 | } catch (DatabaseException.Timeout e) { |
| 263 | log.debug("Failed to listTables. Will retry...", e); |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 264 | } catch (DatabaseException e) { |
Yuta HIGUCHI | 2805222 | 2014-11-20 16:45:32 -0800 | [diff] [blame] | 265 | log.debug("Failed to listTables. Will retry later...", e); |
| 266 | Thread.sleep(RETRY_MS); |
Yuta HIGUCHI | 0136562 | 2014-11-25 17:35:15 -0800 | [diff] [blame^] | 267 | } |
| 268 | if (retries == ACTIVATE_MAX_RETRIES) { |
| 269 | log.error("Failed to listTables after multiple attempts. Giving up."); |
| 270 | // Exiting hoping things will be fixed by the time |
| 271 | // others start using the service |
| 272 | return; |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 273 | } |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 274 | retries++; |
Madan Jampani | 71582ed | 2014-11-18 10:06:01 -0800 | [diff] [blame] | 275 | } while (true); |
| 276 | } |
| 277 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 278 | @Override |
| 279 | public boolean createTable(String name) { |
| 280 | return client.createTable(name); |
| 281 | } |
| 282 | |
| 283 | @Override |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 284 | public boolean createTable(String name, int ttlMillis) { |
| 285 | return client.createTable(name, ttlMillis); |
| 286 | } |
| 287 | |
| 288 | @Override |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 289 | public void dropTable(String name) { |
| 290 | client.dropTable(name); |
| 291 | } |
| 292 | |
| 293 | @Override |
| 294 | public void dropAllTables() { |
| 295 | client.dropAllTables(); |
| 296 | } |
| 297 | |
| 298 | @Override |
Madan Jampani | f5d263b | 2014-11-13 10:04:40 -0800 | [diff] [blame] | 299 | public Set<String> listTables() { |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 300 | return client.listTables(); |
| 301 | } |
| 302 | |
| 303 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 304 | public VersionedValue get(String tableName, String key) { |
| 305 | BatchReadRequest batchRequest = new BatchReadRequest.Builder().get(tableName, key).build(); |
| 306 | ReadResult readResult = batchRead(batchRequest).getAsList().get(0); |
| 307 | if (readResult.status().equals(ReadStatus.OK)) { |
| 308 | return readResult.value(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 309 | } |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 310 | throw new DatabaseException("get failed due to status: " + readResult.status()); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 311 | } |
| 312 | |
| 313 | @Override |
Yuta HIGUCHI | 841c0b6 | 2014-11-13 20:27:14 -0800 | [diff] [blame] | 314 | public Map<String, VersionedValue> getAll(String tableName) { |
| 315 | return client.getAll(tableName); |
| 316 | } |
| 317 | |
| 318 | |
| 319 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 320 | public BatchReadResult batchRead(BatchReadRequest batchRequest) { |
| 321 | return new BatchReadResult(client.batchRead(batchRequest)); |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 322 | } |
| 323 | |
| 324 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 325 | public BatchWriteResult batchWrite(BatchWriteRequest batchRequest) { |
| 326 | return new BatchWriteResult(client.batchWrite(batchRequest)); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 327 | } |
| 328 | |
| 329 | @Override |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 330 | public VersionedValue put(String tableName, String key, byte[] value) { |
| 331 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder().put(tableName, key, value).build(); |
| 332 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 333 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 334 | return writeResult.previousValue(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 335 | } |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 336 | throw new DatabaseException("put failed due to status: " + writeResult.status()); |
| 337 | } |
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 | @Override |
| 340 | public boolean putIfAbsent(String tableName, String key, byte[] value) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 341 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 342 | .putIfAbsent(tableName, key, value).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 343 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 344 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 345 | return true; |
| 346 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 347 | return false; |
| 348 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 349 | throw new DatabaseException("putIfAbsent failed due to status: " |
| 350 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 351 | } |
| 352 | |
| 353 | @Override |
| 354 | public boolean putIfVersionMatches(String tableName, String key, |
| 355 | byte[] value, long version) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 356 | BatchWriteRequest batchRequest = |
| 357 | new BatchWriteRequest.Builder() |
| 358 | .putIfVersionMatches(tableName, key, value, version).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 359 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 360 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 361 | return true; |
| 362 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 363 | return false; |
| 364 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 365 | throw new DatabaseException("putIfVersionMatches failed due to status: " |
| 366 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 367 | } |
| 368 | |
| 369 | @Override |
| 370 | public boolean putIfValueMatches(String tableName, String key, |
| 371 | byte[] oldValue, byte[] newValue) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 372 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 373 | .putIfValueMatches(tableName, key, oldValue, newValue).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 374 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 375 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 376 | return true; |
| 377 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 378 | return false; |
| 379 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 380 | throw new DatabaseException("putIfValueMatches failed due to status: " |
| 381 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 382 | } |
| 383 | |
| 384 | @Override |
| 385 | public VersionedValue remove(String tableName, String key) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 386 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 387 | .remove(tableName, key).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 388 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 389 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 390 | return writeResult.previousValue(); |
| 391 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 392 | throw new DatabaseException("remove failed due to status: " |
| 393 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 394 | } |
| 395 | |
| 396 | @Override |
| 397 | public boolean removeIfVersionMatches(String tableName, String key, |
| 398 | long version) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 399 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 400 | .removeIfVersionMatches(tableName, key, version).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 401 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 402 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 403 | return true; |
| 404 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 405 | return false; |
| 406 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 407 | throw new DatabaseException("removeIfVersionMatches failed due to status: " |
| 408 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 409 | } |
| 410 | |
| 411 | @Override |
| 412 | public boolean removeIfValueMatches(String tableName, String key, |
| 413 | byte[] value) { |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 414 | BatchWriteRequest batchRequest = new BatchWriteRequest.Builder() |
| 415 | .removeIfValueMatches(tableName, key, value).build(); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 416 | WriteResult writeResult = batchWrite(batchRequest).getAsList().get(0); |
| 417 | if (writeResult.status().equals(WriteStatus.OK)) { |
| 418 | return true; |
| 419 | } else if (writeResult.status().equals(WriteStatus.PRECONDITION_VIOLATION)) { |
| 420 | return false; |
| 421 | } |
Madan Jampani | 44e6a54 | 2014-11-12 01:06:51 -0800 | [diff] [blame] | 422 | throw new DatabaseException("removeIfValueMatches failed due to status: " |
| 423 | + writeResult.status()); |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 424 | } |
| 425 | |
| 426 | @Override |
| 427 | public void addMember(final ControllerNode node) { |
| 428 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
| 429 | node.tcpPort()); |
| 430 | log.info("{} was added to the cluster", tcpMember); |
| 431 | synchronized (clusterConfig) { |
| 432 | clusterConfig.addRemoteMember(tcpMember); |
| 433 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 434 | } |
| 435 | |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 436 | private final class LeaderAdvertiser implements Runnable { |
| 437 | |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 438 | @Override |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 439 | public void run() { |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 440 | try { |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 441 | LeaderElectEvent event = myLeaderEvent; |
| 442 | if (event != null) { |
| 443 | log.trace("Broadcasting RAFT_LEADER_ELECTION_EVENT: {}", event); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 444 | // This node just became the leader. |
| 445 | clusterCommunicator.broadcastIncludeSelf( |
| 446 | new ClusterMessage( |
| 447 | clusterService.getLocalNode().id(), |
| 448 | RAFT_LEADER_ELECTION_EVENT, |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 449 | ClusterMessagingProtocol.DB_SERIALIZER.encode(event))); |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 450 | } |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 451 | } catch (Exception e) { |
| 452 | log.debug("LeaderAdvertiser failed with exception", e); |
| 453 | } |
| 454 | } |
| 455 | |
| 456 | } |
| 457 | |
| 458 | private final class RaftLeaderElectionMonitor implements EventHandler<LeaderElectEvent> { |
| 459 | @Override |
| 460 | public void handle(LeaderElectEvent event) { |
| 461 | try { |
| 462 | log.debug("Received LeaderElectEvent: {}", event); |
| 463 | if (clusterConfig.getLocalMember() != null && event.leader().equals(clusterConfig.getLocalMember())) { |
| 464 | log.debug("Broadcasting RAFT_LEADER_ELECTION_EVENT"); |
| 465 | myLeaderEvent = event; |
| 466 | // This node just became the leader. |
| 467 | clusterCommunicator.broadcastIncludeSelf( |
| 468 | new ClusterMessage( |
| 469 | clusterService.getLocalNode().id(), |
| 470 | RAFT_LEADER_ELECTION_EVENT, |
Yuta HIGUCHI | 91768e3 | 2014-11-22 05:06:35 -0800 | [diff] [blame] | 471 | ClusterMessagingProtocol.DB_SERIALIZER.encode(event))); |
Yuta HIGUCHI | f8535f3 | 2014-11-20 11:11:01 -0800 | [diff] [blame] | 472 | } else { |
| 473 | if (myLeaderEvent != null) { |
| 474 | log.debug("This node is no longer the Leader"); |
| 475 | } |
| 476 | myLeaderEvent = null; |
| 477 | } |
Madan Jampani | 5ce3025 | 2014-11-17 20:53:17 -0800 | [diff] [blame] | 478 | } catch (IOException e) { |
| 479 | log.error("Failed to broadcast raft leadership change event", e); |
| 480 | } |
| 481 | } |
| 482 | } |
| 483 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 484 | private final class InternalClusterEventListener |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 485 | implements ClusterEventListener { |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 486 | |
| 487 | @Override |
| 488 | public void event(ClusterEvent event) { |
| 489 | // TODO: Not every node should be part of the consensus ring. |
| 490 | |
| 491 | final ControllerNode node = event.subject(); |
| 492 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 493 | node.tcpPort()); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 494 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 495 | switch (event.type()) { |
| 496 | case INSTANCE_ACTIVATED: |
| 497 | case INSTANCE_ADDED: |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 498 | if (autoAddMember) { |
| 499 | synchronized (clusterConfig) { |
| 500 | if (!clusterConfig.getMembers().contains(tcpMember)) { |
| 501 | log.info("{} was automatically added to the cluster", tcpMember); |
| 502 | clusterConfig.addRemoteMember(tcpMember); |
| 503 | } |
| 504 | } |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 505 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 506 | break; |
| 507 | case INSTANCE_DEACTIVATED: |
| 508 | case INSTANCE_REMOVED: |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 509 | if (autoAddMember) { |
| 510 | Set<DefaultControllerNode> members |
Madan Jampani | 12390c1 | 2014-11-12 00:35:56 -0800 | [diff] [blame] | 511 | = tabletMembers.getOrDefault(DEFAULT_TABLET, |
| 512 | Collections.emptySet()); |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 513 | // remove only if not the initial members |
| 514 | if (!members.contains(node)) { |
| 515 | synchronized (clusterConfig) { |
| 516 | if (clusterConfig.getMembers().contains(tcpMember)) { |
| 517 | log.info("{} was automatically removed from the cluster", tcpMember); |
| 518 | clusterConfig.removeRemoteMember(tcpMember); |
| 519 | } |
| 520 | } |
| 521 | } |
| 522 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 523 | break; |
| 524 | default: |
| 525 | break; |
| 526 | } |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 527 | if (copycat != null) { |
| 528 | log.debug("Current cluster: {}", copycat.cluster()); |
| 529 | } |
| 530 | clusterEventLatch.countDown(); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 531 | } |
| 532 | |
| 533 | } |
| 534 | |
Yuta HIGUCHI | 60731cb | 2014-11-11 01:34:46 -0800 | [diff] [blame] | 535 | @Override |
| 536 | public void removeMember(final ControllerNode node) { |
| 537 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
| 538 | node.tcpPort()); |
| 539 | log.info("{} was removed from the cluster", tcpMember); |
| 540 | synchronized (clusterConfig) { |
| 541 | clusterConfig.removeRemoteMember(tcpMember); |
| 542 | } |
| 543 | } |
| 544 | |
| 545 | @Override |
| 546 | public Collection<ControllerNode> listMembers() { |
| 547 | if (copycat == null) { |
| 548 | return ImmutableList.of(); |
| 549 | } |
| 550 | Set<ControllerNode> members = new HashSet<>(); |
| 551 | for (Member member : copycat.cluster().members()) { |
| 552 | if (member instanceof TcpMember) { |
| 553 | final TcpMember tcpMember = (TcpMember) member; |
| 554 | // TODO assuming tcpMember#host to be IP address, |
| 555 | // but if not lookup DNS, etc. first |
| 556 | IpAddress ip = IpAddress.valueOf(tcpMember.host()); |
| 557 | int tcpPort = tcpMember.port(); |
| 558 | NodeId id = getNodeIdFromIp(ip, tcpPort); |
| 559 | if (id == null) { |
| 560 | log.info("No NodeId found for {}:{}", ip, tcpPort); |
| 561 | continue; |
| 562 | } |
| 563 | members.add(new DefaultControllerNode(id, ip, tcpPort)); |
| 564 | } |
| 565 | } |
| 566 | return members; |
| 567 | } |
| 568 | |
| 569 | private NodeId getNodeIdFromIp(IpAddress ip, int tcpPort) { |
| 570 | for (ControllerNode node : clusterService.getNodes()) { |
| 571 | if (node.ip().equals(ip) && |
| 572 | node.tcpPort() == tcpPort) { |
| 573 | return node.id(); |
| 574 | } |
| 575 | } |
| 576 | return null; |
| 577 | } |
Madan Jampani | def2c65 | 2014-11-12 13:50:10 -0800 | [diff] [blame] | 578 | } |