Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 1 | package org.onlab.onos.store.service.impl; |
| 2 | |
| 3 | import static org.slf4j.LoggerFactory.getLogger; |
| 4 | |
| 5 | import java.util.ArrayList; |
| 6 | import java.util.Arrays; |
| 7 | import java.util.List; |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 8 | import java.util.concurrent.CountDownLatch; |
| 9 | import java.util.concurrent.TimeUnit; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 10 | |
| 11 | import net.kuujo.copycat.Copycat; |
| 12 | import net.kuujo.copycat.StateMachine; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 13 | import net.kuujo.copycat.cluster.ClusterConfig; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 14 | import net.kuujo.copycat.cluster.TcpCluster; |
| 15 | import net.kuujo.copycat.cluster.TcpClusterConfig; |
| 16 | import net.kuujo.copycat.cluster.TcpMember; |
Yuta HIGUCHI | 0c1c100 | 2014-11-05 13:47:25 -0800 | [diff] [blame] | 17 | import net.kuujo.copycat.log.InMemoryLog; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 18 | import net.kuujo.copycat.log.Log; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 19 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 20 | import org.apache.felix.scr.annotations.Activate; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 21 | import org.apache.felix.scr.annotations.Component; |
Yuta HIGUCHI | 657626e | 2014-11-04 20:54:58 -0800 | [diff] [blame] | 22 | import org.apache.felix.scr.annotations.Deactivate; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 23 | import org.apache.felix.scr.annotations.Reference; |
| 24 | import org.apache.felix.scr.annotations.ReferenceCardinality; |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 25 | import org.apache.felix.scr.annotations.Service; |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 26 | import org.onlab.onos.cluster.ClusterEvent; |
| 27 | import org.onlab.onos.cluster.ClusterEventListener; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 28 | import org.onlab.onos.cluster.ClusterService; |
| 29 | import org.onlab.onos.cluster.ControllerNode; |
| 30 | import org.onlab.onos.store.service.DatabaseAdminService; |
| 31 | import org.onlab.onos.store.service.DatabaseException; |
| 32 | import org.onlab.onos.store.service.DatabaseService; |
| 33 | import org.onlab.onos.store.service.NoSuchTableException; |
| 34 | import org.onlab.onos.store.service.OptimisticLockException; |
| 35 | import org.onlab.onos.store.service.OptionalResult; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 36 | import org.onlab.onos.store.service.ReadRequest; |
| 37 | import org.onlab.onos.store.service.ReadResult; |
| 38 | import org.onlab.onos.store.service.WriteAborted; |
| 39 | import org.onlab.onos.store.service.WriteRequest; |
| 40 | import org.onlab.onos.store.service.WriteResult; |
| 41 | import org.slf4j.Logger; |
| 42 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 43 | /** |
| 44 | * Strongly consistent and durable state management service based on |
| 45 | * Copycat implementation of Raft consensus protocol. |
| 46 | */ |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 47 | @Component(immediate = true) |
| 48 | @Service |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 49 | public class DatabaseManager implements DatabaseService, DatabaseAdminService { |
| 50 | |
| 51 | private final Logger log = getLogger(getClass()); |
| 52 | |
| 53 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Yuta HIGUCHI | 5001ba9 | 2014-11-04 21:33:54 -0800 | [diff] [blame] | 54 | protected ClusterService clusterService; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 55 | |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 56 | @Reference(cardinality = ReferenceCardinality.MANDATORY_UNARY) |
Yuta HIGUCHI | 0c1c100 | 2014-11-05 13:47:25 -0800 | [diff] [blame] | 57 | protected DatabaseProtocolService copycatMessagingProtocol; |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 58 | |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 59 | public static final String LOG_FILE_PREFIX = "/tmp/onos-copy-cat-log"; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 60 | |
| 61 | private Copycat copycat; |
| 62 | private DatabaseClient client; |
| 63 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 64 | // guarded by synchronized block |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 65 | private ClusterConfig<TcpMember> clusterConfig; |
| 66 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 67 | private CountDownLatch clusterEventLatch; |
| 68 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 69 | private ClusterEventListener clusterEventListener; |
| 70 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 71 | @Activate |
| 72 | public void activate() { |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 73 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 74 | // TODO: Not every node should be part of the consensus ring. |
Madan Jampani | dfbfa18 | 2014-11-04 22:06:41 -0800 | [diff] [blame] | 75 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 76 | final ControllerNode localNode = clusterService.getLocalNode(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 77 | TcpMember localMember = |
| 78 | new TcpMember( |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 79 | localNode.ip().toString(), |
| 80 | localNode.tcpPort()); |
| 81 | |
| 82 | clusterConfig = new TcpClusterConfig(); |
| 83 | clusterConfig.setLocalMember(localMember); |
| 84 | |
| 85 | List<TcpMember> remoteMembers = new ArrayList<>(clusterService.getNodes().size()); |
| 86 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 87 | clusterEventLatch = new CountDownLatch(1); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 88 | clusterEventListener = new InternalClusterEventListener(); |
| 89 | clusterService.addListener(clusterEventListener); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 90 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 91 | // note: from this point beyond, clusterConfig requires synchronization |
| 92 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 93 | for (ControllerNode node : clusterService.getNodes()) { |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 94 | TcpMember member = new TcpMember(node.ip().toString(), node.tcpPort()); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 95 | if (!member.equals(localMember)) { |
| 96 | remoteMembers.add(member); |
| 97 | } |
| 98 | } |
| 99 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 100 | if (remoteMembers.isEmpty()) { |
| 101 | log.info("This node is the only node in the cluster. " |
| 102 | + "Waiting for others to show up."); |
| 103 | // FIXME: hack trying to relax cases forming multiple consensus rings. |
| 104 | // add seed node configuration to avoid this |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 105 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 106 | // If the node is alone on it's own, wait some time |
| 107 | // hoping other will come up soon |
| 108 | try { |
| 109 | if (!clusterEventLatch.await(120, TimeUnit.SECONDS)) { |
| 110 | log.info("Starting as single node cluster"); |
| 111 | } |
| 112 | } catch (InterruptedException e) { |
| 113 | log.info("Interrupted waiting for others", e); |
| 114 | } |
| 115 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 116 | |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 117 | final TcpCluster cluster; |
| 118 | synchronized (clusterConfig) { |
| 119 | clusterConfig.addRemoteMembers(remoteMembers); |
| 120 | |
| 121 | // Create the cluster. |
| 122 | cluster = new TcpCluster(clusterConfig); |
| 123 | } |
| 124 | log.info("Starting cluster: {}", cluster); |
| 125 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 126 | |
| 127 | StateMachine stateMachine = new DatabaseStateMachine(); |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 128 | // Chronicle + OSGi issue |
Yuta HIGUCHI | 0c1c100 | 2014-11-05 13:47:25 -0800 | [diff] [blame] | 129 | //Log consensusLog = new ChronicleLog(LOG_FILE_PREFIX + "_" + thisNode.id()); |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 130 | //Log consensusLog = new KryoRegisteredInMemoryLog(); |
Madan Jampani | 2ee2000 | 2014-11-06 20:06:12 -0800 | [diff] [blame] | 131 | Log consensusLog = new MapDBLog(LOG_FILE_PREFIX + localNode.id(), |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 132 | ClusterMessagingProtocol.SERIALIZER); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 133 | |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 134 | copycat = new Copycat(stateMachine, consensusLog, cluster, copycatMessagingProtocol); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 135 | copycat.start(); |
| 136 | |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 137 | client = new DatabaseClient(copycatMessagingProtocol.createClient(localMember)); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 138 | |
| 139 | log.info("Started."); |
| 140 | } |
| 141 | |
Yuta HIGUCHI | 657626e | 2014-11-04 20:54:58 -0800 | [diff] [blame] | 142 | @Deactivate |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 143 | public void deactivate() { |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 144 | clusterService.removeListener(clusterEventListener); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 145 | copycat.stop(); |
Yuta HIGUCHI | 657626e | 2014-11-04 20:54:58 -0800 | [diff] [blame] | 146 | log.info("Stopped."); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 147 | } |
| 148 | |
| 149 | @Override |
| 150 | public boolean createTable(String name) { |
| 151 | return client.createTable(name); |
| 152 | } |
| 153 | |
| 154 | @Override |
| 155 | public void dropTable(String name) { |
| 156 | client.dropTable(name); |
| 157 | } |
| 158 | |
| 159 | @Override |
| 160 | public void dropAllTables() { |
| 161 | client.dropAllTables(); |
| 162 | } |
| 163 | |
| 164 | @Override |
| 165 | public List<String> listTables() { |
| 166 | return client.listTables(); |
| 167 | } |
| 168 | |
| 169 | @Override |
| 170 | public ReadResult read(ReadRequest request) { |
| 171 | return batchRead(Arrays.asList(request)).get(0).get(); |
| 172 | } |
| 173 | |
| 174 | @Override |
| 175 | public List<OptionalResult<ReadResult, DatabaseException>> batchRead( |
| 176 | List<ReadRequest> batch) { |
| 177 | List<OptionalResult<ReadResult, DatabaseException>> readResults = new ArrayList<>(batch.size()); |
| 178 | for (InternalReadResult internalReadResult : client.batchRead(batch)) { |
| 179 | if (internalReadResult.status() == InternalReadResult.Status.NO_SUCH_TABLE) { |
| 180 | readResults.add(new DatabaseOperationResult<ReadResult, DatabaseException>( |
| 181 | new NoSuchTableException())); |
| 182 | } else { |
| 183 | readResults.add(new DatabaseOperationResult<ReadResult, DatabaseException>( |
| 184 | internalReadResult.result())); |
| 185 | } |
| 186 | } |
| 187 | return readResults; |
| 188 | } |
| 189 | |
| 190 | @Override |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 191 | public OptionalResult<WriteResult, DatabaseException> writeNothrow(WriteRequest request) { |
| 192 | return batchWrite(Arrays.asList(request)).get(0); |
| 193 | } |
| 194 | |
| 195 | @Override |
Yuta HIGUCHI | c6b8f61 | 2014-11-06 19:04:13 -0800 | [diff] [blame] | 196 | public WriteResult write(WriteRequest request) { |
| 197 | // throws OptimisticLockException, PreconditionFailedException { |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 198 | return writeNothrow(request).get(); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 199 | } |
| 200 | |
| 201 | @Override |
| 202 | public List<OptionalResult<WriteResult, DatabaseException>> batchWrite( |
| 203 | List<WriteRequest> batch) { |
| 204 | List<OptionalResult<WriteResult, DatabaseException>> writeResults = new ArrayList<>(batch.size()); |
| 205 | for (InternalWriteResult internalWriteResult : client.batchWrite(batch)) { |
| 206 | if (internalWriteResult.status() == InternalWriteResult.Status.NO_SUCH_TABLE) { |
| 207 | writeResults.add(new DatabaseOperationResult<WriteResult, DatabaseException>( |
| 208 | new NoSuchTableException())); |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 209 | } else if (internalWriteResult.status() == InternalWriteResult.Status.PREVIOUS_VERSION_MISMATCH) { |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 210 | writeResults.add(new DatabaseOperationResult<WriteResult, DatabaseException>( |
| 211 | new OptimisticLockException())); |
| 212 | } else if (internalWriteResult.status() == InternalWriteResult.Status.PREVIOUS_VALUE_MISMATCH) { |
| 213 | // TODO: throw a different exception? |
| 214 | writeResults.add(new DatabaseOperationResult<WriteResult, DatabaseException>( |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 215 | new OptimisticLockException())); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 216 | } else if (internalWriteResult.status() == InternalWriteResult.Status.ABORTED) { |
| 217 | writeResults.add(new DatabaseOperationResult<WriteResult, DatabaseException>( |
| 218 | new WriteAborted())); |
| 219 | } else { |
| 220 | writeResults.add(new DatabaseOperationResult<WriteResult, DatabaseException>( |
| 221 | internalWriteResult.result())); |
| 222 | } |
| 223 | } |
| 224 | return writeResults; |
| 225 | |
| 226 | } |
| 227 | |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 228 | private final class InternalClusterEventListener |
| 229 | implements ClusterEventListener { |
| 230 | |
| 231 | @Override |
| 232 | public void event(ClusterEvent event) { |
| 233 | // TODO: Not every node should be part of the consensus ring. |
| 234 | |
| 235 | final ControllerNode node = event.subject(); |
| 236 | final TcpMember tcpMember = new TcpMember(node.ip().toString(), |
| 237 | node.tcpPort()); |
| 238 | |
| 239 | log.trace("{}", event); |
| 240 | switch (event.type()) { |
| 241 | case INSTANCE_ACTIVATED: |
| 242 | case INSTANCE_ADDED: |
| 243 | log.info("{} was added to the cluster", tcpMember); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 244 | synchronized (clusterConfig) { |
| 245 | clusterConfig.addRemoteMember(tcpMember); |
| 246 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 247 | break; |
| 248 | case INSTANCE_DEACTIVATED: |
| 249 | case INSTANCE_REMOVED: |
| 250 | log.info("{} was removed from the cluster", tcpMember); |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 251 | synchronized (clusterConfig) { |
| 252 | clusterConfig.removeRemoteMember(tcpMember); |
| 253 | } |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 254 | break; |
| 255 | default: |
| 256 | break; |
| 257 | } |
Yuta HIGUCHI | 79a1e5e | 2014-11-05 17:42:01 -0800 | [diff] [blame] | 258 | if (copycat != null) { |
| 259 | log.debug("Current cluster: {}", copycat.cluster()); |
| 260 | } |
| 261 | clusterEventLatch.countDown(); |
Yuta HIGUCHI | 5027b6b | 2014-11-05 16:23:26 -0800 | [diff] [blame] | 262 | } |
| 263 | |
| 264 | } |
| 265 | |
| 266 | public static final class KryoRegisteredInMemoryLog extends InMemoryLog { |
| 267 | public KryoRegisteredInMemoryLog() { |
| 268 | super(); |
| 269 | // required to deserialize object across bundles |
| 270 | super.kryo.register(TcpMember.class, new TcpMemberSerializer()); |
| 271 | super.kryo.register(TcpClusterConfig.class, new TcpClusterConfigSerializer()); |
| 272 | } |
| 273 | } |
| 274 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 275 | private class DatabaseOperationResult<R, E extends DatabaseException> implements OptionalResult<R, E> { |
| 276 | |
| 277 | private final R result; |
| 278 | private final DatabaseException exception; |
| 279 | |
| 280 | public DatabaseOperationResult(R result) { |
| 281 | this.result = result; |
| 282 | this.exception = null; |
| 283 | } |
| 284 | |
| 285 | public DatabaseOperationResult(DatabaseException exception) { |
| 286 | this.result = null; |
| 287 | this.exception = exception; |
| 288 | } |
| 289 | |
| 290 | @Override |
| 291 | public R get() { |
| 292 | if (result != null) { |
| 293 | return result; |
| 294 | } |
| 295 | throw exception; |
| 296 | } |
| 297 | |
| 298 | @Override |
| 299 | public boolean hasValidResult() { |
| 300 | return result != null; |
| 301 | } |
| 302 | |
| 303 | @Override |
| 304 | public String toString() { |
| 305 | if (result != null) { |
| 306 | return result.toString(); |
| 307 | } else { |
| 308 | return exception.toString(); |
| 309 | } |
| 310 | } |
| 311 | } |
| 312 | } |