Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 1 | package org.onlab.onos.store.service.impl; |
| 2 | |
Yuta HIGUCHI | 39ae550 | 2014-11-05 16:42:12 -0800 | [diff] [blame] | 3 | import static org.slf4j.LoggerFactory.getLogger; |
| 4 | |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 5 | import java.io.ByteArrayInputStream; |
| 6 | import java.io.ByteArrayOutputStream; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 7 | import java.util.ArrayList; |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 8 | import java.util.Arrays; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 9 | import java.util.List; |
| 10 | import java.util.Map; |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 11 | import java.util.zip.DeflaterOutputStream; |
| 12 | import java.util.zip.InflaterInputStream; |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 13 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 14 | import net.kuujo.copycat.Command; |
| 15 | import net.kuujo.copycat.Query; |
| 16 | import net.kuujo.copycat.StateMachine; |
| 17 | |
| 18 | import org.onlab.onos.store.serializers.KryoSerializer; |
| 19 | import org.onlab.onos.store.service.ReadRequest; |
| 20 | import org.onlab.onos.store.service.ReadResult; |
Madan Jampani | 37c2e70 | 2014-11-04 18:11:10 -0800 | [diff] [blame] | 21 | import org.onlab.onos.store.service.VersionedValue; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 22 | import org.onlab.onos.store.service.WriteRequest; |
| 23 | import org.onlab.onos.store.service.WriteResult; |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 24 | import org.onlab.onos.store.service.impl.InternalWriteResult.Status; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 25 | import org.onlab.util.KryoNamespace; |
Yuta HIGUCHI | 39ae550 | 2014-11-05 16:42:12 -0800 | [diff] [blame] | 26 | import org.slf4j.Logger; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 27 | |
Yuta HIGUCHI | 1838f88 | 2014-11-05 18:42:00 -0800 | [diff] [blame] | 28 | import com.google.common.collect.ImmutableList; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 29 | import com.google.common.collect.Maps; |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 30 | import com.google.common.io.ByteStreams; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 31 | |
Madan Jampani | 686fa18 | 2014-11-04 23:16:27 -0800 | [diff] [blame] | 32 | /** |
| 33 | * StateMachine whose transitions are coordinated/replicated |
| 34 | * by Raft consensus. |
| 35 | * Each Raft cluster member has a instance of this state machine that is |
| 36 | * independently updated in lock step once there is consensus |
| 37 | * on the next transition. |
| 38 | */ |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 39 | public class DatabaseStateMachine implements StateMachine { |
| 40 | |
Yuta HIGUCHI | 39ae550 | 2014-11-05 16:42:12 -0800 | [diff] [blame] | 41 | private final Logger log = getLogger(getClass()); |
| 42 | |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 43 | // serializer used for snapshot |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 44 | public static final KryoSerializer SERIALIZER = new KryoSerializer() { |
| 45 | @Override |
| 46 | protected void setupKryoPool() { |
| 47 | serializerPool = KryoNamespace.newBuilder() |
| 48 | .register(VersionedValue.class) |
| 49 | .register(State.class) |
Madan Jampani | 9b19a82 | 2014-11-04 21:37:13 -0800 | [diff] [blame] | 50 | .register(ClusterMessagingProtocol.COMMON) |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 51 | .build() |
| 52 | .populate(1); |
| 53 | } |
| 54 | }; |
| 55 | |
| 56 | private State state = new State(); |
| 57 | |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 58 | private boolean compressSnapshot = false; |
| 59 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 60 | @Command |
| 61 | public boolean createTable(String tableName) { |
| 62 | return state.getTables().putIfAbsent(tableName, Maps.newHashMap()) == null; |
| 63 | } |
| 64 | |
| 65 | @Command |
| 66 | public boolean dropTable(String tableName) { |
| 67 | return state.getTables().remove(tableName) != null; |
| 68 | } |
| 69 | |
| 70 | @Command |
| 71 | public boolean dropAllTables() { |
| 72 | state.getTables().clear(); |
| 73 | return true; |
| 74 | } |
| 75 | |
| 76 | @Query |
Yuta HIGUCHI | 1838f88 | 2014-11-05 18:42:00 -0800 | [diff] [blame] | 77 | public List<String> listTables() { |
| 78 | return ImmutableList.copyOf(state.getTables().keySet()); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 79 | } |
| 80 | |
| 81 | @Query |
| 82 | public List<InternalReadResult> read(List<ReadRequest> requests) { |
| 83 | List<InternalReadResult> results = new ArrayList<>(requests.size()); |
| 84 | for (ReadRequest request : requests) { |
| 85 | Map<String, VersionedValue> table = state.getTables().get(request.tableName()); |
| 86 | if (table == null) { |
| 87 | results.add(new InternalReadResult(InternalReadResult.Status.NO_SUCH_TABLE, null)); |
| 88 | continue; |
| 89 | } |
Yuta HIGUCHI | 3bd8cdc | 2014-11-05 19:11:44 -0800 | [diff] [blame] | 90 | VersionedValue value = VersionedValue.copy(table.get(request.key())); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 91 | results.add(new InternalReadResult( |
| 92 | InternalReadResult.Status.OK, |
| 93 | new ReadResult( |
| 94 | request.tableName(), |
| 95 | request.key(), |
| 96 | value))); |
| 97 | } |
| 98 | return results; |
| 99 | } |
| 100 | |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 101 | InternalWriteResult.Status checkIfApplicable(WriteRequest request, |
| 102 | VersionedValue value) { |
| 103 | |
| 104 | switch (request.type()) { |
| 105 | case PUT: |
| 106 | return InternalWriteResult.Status.OK; |
| 107 | |
| 108 | case PUT_IF_ABSENT: |
| 109 | if (value == null) { |
| 110 | return InternalWriteResult.Status.OK; |
| 111 | } |
| 112 | return InternalWriteResult.Status.PREVIOUS_VALUE_MISMATCH; |
| 113 | case PUT_IF_VALUE: |
| 114 | case REMOVE_IF_VALUE: |
| 115 | if (value != null && Arrays.equals(value.value(), request.oldValue())) { |
| 116 | return InternalWriteResult.Status.OK; |
| 117 | } |
| 118 | return InternalWriteResult.Status.PREVIOUS_VALUE_MISMATCH; |
| 119 | case PUT_IF_VERSION: |
| 120 | case REMOVE_IF_VERSION: |
| 121 | if (value != null && request.previousVersion() == value.version()) { |
| 122 | return InternalWriteResult.Status.OK; |
| 123 | } |
| 124 | return InternalWriteResult.Status.PREVIOUS_VERSION_MISMATCH; |
| 125 | case REMOVE: |
| 126 | return InternalWriteResult.Status.OK; |
Yuta HIGUCHI | c6b8f61 | 2014-11-06 19:04:13 -0800 | [diff] [blame] | 127 | default: |
| 128 | break; |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 129 | } |
| 130 | log.error("Should never reach here {}", request); |
| 131 | return InternalWriteResult.Status.ABORTED; |
| 132 | } |
| 133 | |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 134 | @Command |
| 135 | public List<InternalWriteResult> write(List<WriteRequest> requests) { |
Yuta HIGUCHI | bddc81c4 | 2014-11-05 18:53:09 -0800 | [diff] [blame] | 136 | |
| 137 | // applicability check |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 138 | boolean abort = false; |
| 139 | List<InternalWriteResult.Status> validationResults = new ArrayList<>(requests.size()); |
| 140 | for (WriteRequest request : requests) { |
| 141 | Map<String, VersionedValue> table = state.getTables().get(request.tableName()); |
| 142 | if (table == null) { |
| 143 | validationResults.add(InternalWriteResult.Status.NO_SUCH_TABLE); |
| 144 | abort = true; |
| 145 | continue; |
| 146 | } |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 147 | final VersionedValue value = table.get(request.key()); |
| 148 | Status result = checkIfApplicable(request, value); |
| 149 | validationResults.add(result); |
| 150 | if (result != Status.OK) { |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 151 | abort = true; |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 152 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 153 | } |
| 154 | |
| 155 | List<InternalWriteResult> results = new ArrayList<>(requests.size()); |
| 156 | |
| 157 | if (abort) { |
| 158 | for (InternalWriteResult.Status validationResult : validationResults) { |
| 159 | if (validationResult == InternalWriteResult.Status.OK) { |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 160 | // aborted due to applicability check failure on other request |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 161 | results.add(new InternalWriteResult(InternalWriteResult.Status.ABORTED, null)); |
| 162 | } else { |
| 163 | results.add(new InternalWriteResult(validationResult, null)); |
| 164 | } |
| 165 | } |
| 166 | return results; |
| 167 | } |
| 168 | |
Yuta HIGUCHI | bddc81c4 | 2014-11-05 18:53:09 -0800 | [diff] [blame] | 169 | // apply changes |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 170 | for (WriteRequest request : requests) { |
| 171 | Map<String, VersionedValue> table = state.getTables().get(request.tableName()); |
Yuta HIGUCHI | bddc81c4 | 2014-11-05 18:53:09 -0800 | [diff] [blame] | 172 | // FIXME: If this method could be called by multiple thread, |
| 173 | // synchronization scope is wrong. |
| 174 | // Whole function including applicability check needs to be protected. |
| 175 | // Confirm copycat's thread safety requirement for StateMachine |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 176 | // TODO: If we need isolation, we need to block reads also |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 177 | synchronized (table) { |
Yuta HIGUCHI | 361664e | 2014-11-06 17:28:47 -0800 | [diff] [blame] | 178 | switch (request.type()) { |
| 179 | case PUT: |
| 180 | case PUT_IF_ABSENT: |
| 181 | case PUT_IF_VALUE: |
| 182 | case PUT_IF_VERSION: |
| 183 | VersionedValue newValue = new VersionedValue(request.newValue(), state.nextVersion()); |
| 184 | VersionedValue previousValue = table.put(request.key(), newValue); |
| 185 | WriteResult putResult = new WriteResult(request.tableName(), request.key(), previousValue); |
| 186 | results.add(InternalWriteResult.ok(putResult)); |
| 187 | break; |
| 188 | |
| 189 | case REMOVE: |
| 190 | case REMOVE_IF_VALUE: |
| 191 | case REMOVE_IF_VERSION: |
| 192 | VersionedValue removedValue = table.remove(request.key()); |
| 193 | WriteResult removeResult = new WriteResult(request.tableName(), request.key(), removedValue); |
| 194 | results.add(InternalWriteResult.ok(removeResult)); |
| 195 | break; |
| 196 | |
| 197 | default: |
| 198 | log.error("Invalid WriteRequest type {}", request.type()); |
| 199 | break; |
| 200 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 201 | } |
| 202 | } |
| 203 | return results; |
| 204 | } |
| 205 | |
| 206 | public class State { |
| 207 | |
| 208 | private final Map<String, Map<String, VersionedValue>> tables = |
| 209 | Maps.newHashMap(); |
| 210 | private long versionCounter = 1; |
| 211 | |
| 212 | Map<String, Map<String, VersionedValue>> getTables() { |
| 213 | return tables; |
| 214 | } |
| 215 | |
| 216 | long nextVersion() { |
| 217 | return versionCounter++; |
| 218 | } |
| 219 | } |
| 220 | |
| 221 | @Override |
| 222 | public byte[] takeSnapshot() { |
| 223 | try { |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 224 | if (compressSnapshot) { |
| 225 | byte[] input = SERIALIZER.encode(state); |
| 226 | ByteArrayOutputStream comp = new ByteArrayOutputStream(input.length); |
| 227 | DeflaterOutputStream compressor = new DeflaterOutputStream(comp); |
| 228 | compressor.write(input, 0, input.length); |
| 229 | compressor.close(); |
| 230 | return comp.toByteArray(); |
| 231 | } else { |
| 232 | return SERIALIZER.encode(state); |
| 233 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 234 | } catch (Exception e) { |
Madan Jampani | 778f7ad | 2014-11-05 22:46:15 -0800 | [diff] [blame] | 235 | log.error("Failed to take snapshot", e); |
| 236 | throw new SnapshotException(e); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 237 | } |
| 238 | } |
| 239 | |
| 240 | @Override |
| 241 | public void installSnapshot(byte[] data) { |
| 242 | try { |
Yuta HIGUCHI | a7680a3 | 2014-11-06 22:17:37 -0800 | [diff] [blame] | 243 | if (compressSnapshot) { |
| 244 | ByteArrayInputStream in = new ByteArrayInputStream(data); |
| 245 | InflaterInputStream decompressor = new InflaterInputStream(in); |
| 246 | ByteStreams.toByteArray(decompressor); |
| 247 | this.state = SERIALIZER.decode(ByteStreams.toByteArray(decompressor)); |
| 248 | } else { |
| 249 | this.state = SERIALIZER.decode(data); |
| 250 | } |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 251 | } catch (Exception e) { |
Madan Jampani | 778f7ad | 2014-11-05 22:46:15 -0800 | [diff] [blame] | 252 | log.error("Failed to install from snapshot", e); |
| 253 | throw new SnapshotException(e); |
Madan Jampani | 08822c4 | 2014-11-04 17:17:46 -0800 | [diff] [blame] | 254 | } |
| 255 | } |
| 256 | } |