2 * Copyright (c) 2016 Brocade Communications Systems, Inc. and others. All rights reserved.
4 * This program and the accompanying materials are made available under the
5 * terms of the Eclipse Public License v1.0 which accompanies this distribution,
6 * and is available at http://www.eclipse.org/legal/epl-v10.html
8 package org.opendaylight.controller.cluster.raft;
10 import static org.junit.Assert.assertEquals;
11 import com.google.common.base.Optional;
12 import com.google.common.collect.ImmutableMap;
13 import com.google.common.collect.Sets;
14 import java.util.Arrays;
15 import java.util.concurrent.TimeUnit;
16 import org.junit.Test;
17 import org.opendaylight.controller.cluster.notifications.LeaderStateChanged;
18 import org.opendaylight.controller.cluster.raft.AbstractRaftActorIntegrationTest.TestRaftActor.Builder;
19 import org.opendaylight.controller.cluster.raft.base.messages.ApplyState;
20 import org.opendaylight.controller.cluster.raft.base.messages.ElectionTimeout;
21 import org.opendaylight.controller.cluster.raft.base.messages.SnapshotComplete;
22 import org.opendaylight.controller.cluster.raft.messages.AppendEntries;
23 import org.opendaylight.controller.cluster.raft.persisted.ServerConfigurationPayload;
24 import org.opendaylight.controller.cluster.raft.persisted.ServerInfo;
25 import org.opendaylight.controller.cluster.raft.persisted.UpdateElectionTerm;
26 import org.opendaylight.controller.cluster.raft.policy.DisableElectionsRaftPolicy;
27 import org.opendaylight.controller.cluster.raft.utils.InMemoryJournal;
28 import org.opendaylight.controller.cluster.raft.utils.MessageCollectorActor;
29 import akka.actor.ActorRef;
30 import akka.dispatch.Dispatchers;
31 import scala.concurrent.duration.FiniteDuration;
34 * Integration test for various scenarios involving non-voting followers.
36 * @author Thomas Pantelis
38 public class NonVotingFollowerIntegrationTest extends AbstractRaftActorIntegrationTest {
39 private TestRaftActor followerInstance;
40 private TestRaftActor leaderInstance;
41 private final Builder follower1Builder = TestRaftActor.newBuilder();
44 * Tests non-voting follower re-sync after the non-persistent leader restarts with an empty log. In this
45 * case the follower's log will be ahead of the leader's log as the follower retains the previous
46 * data in memory. The leader must force an install snapshot to re-sync the follower's state.
49 public void testFollowerResyncWithEmptyLeaderLogAfterNonPersistentLeaderRestart() {
50 testLog.info("testFollowerResyncWithEmptyLeaderLogAfterNonPersistentLeaderRestart starting");
52 setupLeaderAndNonVotingFollower();
54 // Add log entries and verify they are committed and applied by both nodes.
56 expSnapshotState.add(sendPayloadData(leaderActor, "zero"));
57 expSnapshotState.add(sendPayloadData(leaderActor, "one"));
58 expSnapshotState.add(sendPayloadData(leaderActor, "two"));
60 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, 3);
61 MessageCollectorActor.expectMatching(follower1CollectorActor, ApplyState.class, 3);
63 assertEquals("Leader journal lastIndex", 2, leaderContext.getReplicatedLog().lastIndex());
64 assertEquals("Leader commit index", 2, leaderContext.getCommitIndex());
65 assertEquals("Follower journal lastIndex", 2, follower1Context.getReplicatedLog().lastIndex());
66 assertEquals("Follower commit index", 2, follower1Context.getCommitIndex());
67 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
69 // Persisted journal should only contain the ServerConfigurationPayload and 2 UpdateElectionTerm entries.
70 assertEquals("Leader persisted journal size", 3, InMemoryJournal.get(leaderId).size());
74 killActor(leaderActor);
75 MessageCollectorActor.clearMessages(follower1CollectorActor);
77 createNewLeaderActor();
79 //follower1Actor.underlyingActor().startDropMessages(AppendEntries.class);
82 assertEquals("Leader term", currentTerm, leaderContext.getTermInformation().getCurrentTerm());
83 assertEquals("Leader journal lastIndex", -1, leaderContext.getReplicatedLog().lastIndex());
84 assertEquals("Leader commit index", -1, leaderContext.getCommitIndex());
86 // After restart, the leader's log and the follower's log will be ahead so the leader should force an
87 // install snapshot to re-sync the follower's log and state.
89 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, SnapshotComplete.class);
91 assertEquals("Follower term", currentTerm, follower1Context.getTermInformation().getCurrentTerm());
92 assertEquals("Follower journal lastIndex", -1, follower1Context.getReplicatedLog().lastIndex());
93 assertEquals("Follower commit index", -1, follower1Context.getCommitIndex());
95 expSnapshotState.add(sendPayloadData(leaderActor, "zero-1"));
97 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, ApplyState.class);
99 assertEquals("Follower journal lastIndex", 0, follower1Context.getReplicatedLog().lastIndex());
100 assertEquals("Follower commit index", 0, follower1Context.getCommitIndex());
101 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
103 testLog.info("testFollowerResyncWithEmptyLeaderLogAfterNonPersistentLeaderRestart ending");
107 * Tests non-voting follower re-sync after the non-persistent leader restarts and commits new log
108 * entries prior to re-connecting to the follower. The leader's last index will still be less than the
109 * follower's last index corresponding to the previous data retained in memory. So the follower's log
110 * will be ahead of the leader's log and the leader must force an install snapshot to re-sync the
114 public void testFollowerResyncWithLessLeaderLogEntriesAfterNonPersistentLeaderRestart() {
115 testLog.info("testFollowerResyncWithLessLeaderLogEntriesAfterNonPersistentLeaderRestart starting");
117 setupLeaderAndNonVotingFollower();
119 // Add log entries and verify they are committed and applied by both nodes.
121 expSnapshotState.add(sendPayloadData(leaderActor, "zero"));
122 expSnapshotState.add(sendPayloadData(leaderActor, "one"));
123 expSnapshotState.add(sendPayloadData(leaderActor, "two"));
125 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, 3);
126 MessageCollectorActor.expectMatching(follower1CollectorActor, ApplyState.class, 3);
128 assertEquals("Leader journal lastIndex", 2, leaderContext.getReplicatedLog().lastIndex());
129 assertEquals("Leader commit index", 2, leaderContext.getCommitIndex());
130 assertEquals("Follower journal lastIndex", 2, follower1Context.getReplicatedLog().lastIndex());
131 assertEquals("Follower commit index", 2, follower1Context.getCommitIndex());
132 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
134 // Restart the leader
136 killActor(leaderActor);
137 MessageCollectorActor.clearMessages(follower1CollectorActor);
139 // Temporarily drop AppendEntries to simulate a disconnect when the leader restarts.
140 followerInstance.startDropMessages(AppendEntries.class);
142 createNewLeaderActor();
145 assertEquals("Leader term", currentTerm, leaderContext.getTermInformation().getCurrentTerm());
146 assertEquals("Leader journal lastIndex", -1, leaderContext.getReplicatedLog().lastIndex());
147 assertEquals("Leader commit index", -1, leaderContext.getCommitIndex());
149 // Add new log entries to the leader - one less than the prior log entries
151 expSnapshotState.add(sendPayloadData(leaderActor, "zero-1"));
152 expSnapshotState.add(sendPayloadData(leaderActor, "one-1"));
154 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, 2);
155 assertEquals("Leader journal lastIndex", 1, leaderContext.getReplicatedLog().lastIndex());
156 assertEquals("Leader commit index", 1, leaderContext.getCommitIndex());
158 // Re-enable AppendEntries to the follower. The leaders previous index will be present in the
159 // follower's but the terms won't match and the follower's log will be ahead of the leader's log
160 // The leader should force an install snapshot to re-sync the entire follower's log and state.
162 followerInstance.stopDropMessages(AppendEntries.class);
163 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, SnapshotComplete.class);
165 assertEquals("Follower term", currentTerm, follower1Context.getTermInformation().getCurrentTerm());
166 assertEquals("Follower journal lastIndex", 1, follower1Context.getReplicatedLog().lastIndex());
167 assertEquals("Follower journal lastTerm", currentTerm, follower1Context.getReplicatedLog().lastTerm());
168 assertEquals("Follower commit index", 1, follower1Context.getCommitIndex());
169 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
171 testLog.info("testFollowerResyncWithLessLeaderLogEntriesAfterNonPersistentLeaderRestart ending");
175 * Tests non-voting follower re-sync after the non-persistent leader restarts and commits new log
176 * entries prior to re-connecting to the follower. The leader's last index will be 1 greater than the
177 * follower's last index corresponding to the previous data retained in memory. So the follower's log
178 * will be behind the leader's log but the leader's log entries will have a higher term. In this case the
179 * leader should force an install snapshot to re-sync the follower's state.
182 public void testFollowerResyncWithOneMoreLeaderLogEntryAfterNonPersistentLeaderRestart() {
183 testLog.info("testFollowerResyncWithOneMoreLeaderLogEntryAfterNonPersistentLeaderRestart starting");
185 setupLeaderAndNonVotingFollower();
187 // Add log entries and verify they are committed and applied by both nodes.
189 expSnapshotState.add(sendPayloadData(leaderActor, "zero"));
190 expSnapshotState.add(sendPayloadData(leaderActor, "one"));
192 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, 2);
193 MessageCollectorActor.expectMatching(follower1CollectorActor, ApplyState.class, 2);
195 assertEquals("Leader journal lastIndex", 1, leaderContext.getReplicatedLog().lastIndex());
196 assertEquals("Leader commit index", 1, leaderContext.getCommitIndex());
197 assertEquals("Follower journal lastIndex", 1, follower1Context.getReplicatedLog().lastIndex());
198 assertEquals("Follower commit index", 1, follower1Context.getCommitIndex());
199 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
201 // Restart the leader
203 killActor(leaderActor);
204 MessageCollectorActor.clearMessages(follower1CollectorActor);
206 // Temporarily drop AppendEntries to simulate a disconnect when the leader restarts.
207 followerInstance.startDropMessages(AppendEntries.class);
209 createNewLeaderActor();
212 assertEquals("Leader term", currentTerm, leaderContext.getTermInformation().getCurrentTerm());
213 assertEquals("Leader journal lastIndex", -1, leaderContext.getReplicatedLog().lastIndex());
214 assertEquals("Leader commit index", -1, leaderContext.getCommitIndex());
216 // Add new log entries to the leader - one more than the prior log entries
218 expSnapshotState.add(sendPayloadData(leaderActor, "zero-1"));
219 expSnapshotState.add(sendPayloadData(leaderActor, "one-1"));
220 expSnapshotState.add(sendPayloadData(leaderActor, "two-1"));
222 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, 3);
223 assertEquals("Leader journal lastIndex", 2, leaderContext.getReplicatedLog().lastIndex());
224 assertEquals("Leader commit index", 2, leaderContext.getCommitIndex());
225 assertEquals("Leader replicatedToAllIndex", -1, leaderInstance.getCurrentBehavior().getReplicatedToAllIndex());
227 // Re-enable AppendEntries to the follower. The follower's log will be out of sync and it should
228 // should force the leader to install snapshot to re-sync the entire follower's log and state.
230 followerInstance.stopDropMessages(AppendEntries.class);
231 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, SnapshotComplete.class);
233 assertEquals("Follower term", currentTerm, follower1Context.getTermInformation().getCurrentTerm());
234 assertEquals("Follower journal lastIndex", 2, follower1Context.getReplicatedLog().lastIndex());
235 assertEquals("Follower journal lastTerm", currentTerm, follower1Context.getReplicatedLog().lastTerm());
236 assertEquals("Follower commit index", 2, follower1Context.getCommitIndex());
237 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
239 testLog.info("testFollowerResyncWithOneMoreLeaderLogEntryAfterNonPersistentLeaderRestart ending");
243 * Tests non-voting follower re-sync after the non-persistent leader restarts and commits new log
244 * entries prior to re-connecting to the follower. The leader's last index will be greater than the
245 * follower's last index corresponding to the previous data retained in memory. So the follower's log
246 * will be behind the leader's log but the leader's log entries will have a higher term. It also adds a
247 * "down" peer on restart so the leader doesn't trim its log as it's trying to resync the follower.
248 * Eventually the follower should force the leader to install snapshot to re-sync its state.
251 public void testFollowerResyncWithMoreLeaderLogEntriesAndDownPeerAfterNonPersistentLeaderRestart() {
252 testLog.info("testFollowerResyncWithMoreLeaderLogEntriesAndDownPeerAfterNonPersistentLeaderRestart starting");
254 setupLeaderAndNonVotingFollower();
256 // Add log entries and verify they are committed and applied by both nodes.
258 expSnapshotState.add(sendPayloadData(leaderActor, "zero"));
259 expSnapshotState.add(sendPayloadData(leaderActor, "one"));
260 expSnapshotState.add(sendPayloadData(leaderActor, "two"));
262 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, expSnapshotState.size());
263 MessageCollectorActor.expectMatching(follower1CollectorActor, ApplyState.class, expSnapshotState.size());
266 assertEquals("Leader journal lastIndex", lastIndex, leaderContext.getReplicatedLog().lastIndex());
267 assertEquals("Leader commit index", lastIndex, leaderContext.getCommitIndex());
268 assertEquals("Follower journal lastIndex", lastIndex, follower1Context.getReplicatedLog().lastIndex());
269 assertEquals("Follower commit index", lastIndex, follower1Context.getCommitIndex());
270 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
272 MessageCollectorActor.clearMessages(follower1CollectorActor);
273 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, AppendEntries.class);
274 assertEquals("Follower snapshot index", lastIndex - 1, follower1Context.getReplicatedLog().getSnapshotIndex());
275 assertEquals("Follower journal size", 1, leaderContext.getReplicatedLog().size());
277 // Restart the leader
279 killActor(leaderActor);
280 MessageCollectorActor.clearMessages(follower1CollectorActor);
282 // Temporarily drop AppendEntries to simulate a disconnect when the leader restarts.
283 followerInstance.startDropMessages(AppendEntries.class);
285 // Add a "down" peer so the leader doesn't trim its log as it's trying to resync the follower. The
286 // leader will keep decrementing the follower's nextIndex to try to find a matching index. Since
287 // there is no matching index it will eventually hit index 0 which should cause the follower to
288 // force an install snapshot upon failure to remove the conflicting indexes due to indexes 0 and 1
289 // being in the prior snapshot and not the log.
291 // We also add another voting follower actor into the mix even though it shoildn't affect the
293 ServerConfigurationPayload persistedServerConfig = new ServerConfigurationPayload(Arrays.asList(
294 new ServerInfo(leaderId, true), new ServerInfo(follower1Id, false),
295 new ServerInfo(follower2Id, true), new ServerInfo("downPeer", false)));
296 ReplicatedLogImplEntry persistedServerConfigEntry = new ReplicatedLogImplEntry(0, currentTerm,
297 persistedServerConfig);
299 InMemoryJournal.clear();
300 InMemoryJournal.addEntry(leaderId, 1, new UpdateElectionTerm(currentTerm, leaderId));
301 InMemoryJournal.addEntry(leaderId, 2, persistedServerConfigEntry);
302 InMemoryJournal.addEntry(follower2Id, 1, persistedServerConfigEntry);
304 DefaultConfigParamsImpl follower2ConfigParams = newFollowerConfigParams();
305 follower2ConfigParams.setCustomRaftPolicyImplementationClass(DisableElectionsRaftPolicy.class.getName());
306 follower2Actor = newTestRaftActor(follower2Id, TestRaftActor.newBuilder().peerAddresses(
307 ImmutableMap.of(leaderId, testActorPath(leaderId), follower1Id, follower1Actor.path().toString())).
308 config(follower2ConfigParams).persistent(Optional.of(false)));
309 TestRaftActor follower2Instance = follower2Actor.underlyingActor();
310 follower2Instance.waitForRecoveryComplete();
311 follower2CollectorActor = follower2Instance.collectorActor();
313 peerAddresses = ImmutableMap.of(follower1Id, follower1Actor.path().toString(),
314 follower2Id, follower2Actor.path().toString());
316 createNewLeaderActor();
319 assertEquals("Leader term", currentTerm, leaderContext.getTermInformation().getCurrentTerm());
320 assertEquals("Leader journal lastIndex", -1, leaderContext.getReplicatedLog().lastIndex());
321 assertEquals("Leader commit index", -1, leaderContext.getCommitIndex());
323 // Add new log entries to the leader - several more than the prior log entries
325 expSnapshotState.add(sendPayloadData(leaderActor, "zero-1"));
326 expSnapshotState.add(sendPayloadData(leaderActor, "one-1"));
327 expSnapshotState.add(sendPayloadData(leaderActor, "two-1"));
328 expSnapshotState.add(sendPayloadData(leaderActor, "three-1"));
329 expSnapshotState.add(sendPayloadData(leaderActor, "four-1"));
331 MessageCollectorActor.expectMatching(leaderCollectorActor, ApplyState.class, expSnapshotState.size());
332 MessageCollectorActor.expectMatching(follower2CollectorActor, ApplyState.class, expSnapshotState.size());
335 assertEquals("Leader journal lastIndex", lastIndex, leaderContext.getReplicatedLog().lastIndex());
336 assertEquals("Leader commit index", lastIndex, leaderContext.getCommitIndex());
337 assertEquals("Leader snapshot index", -1, leaderContext.getReplicatedLog().getSnapshotIndex());
338 assertEquals("Leader replicatedToAllIndex", -1, leaderInstance.getCurrentBehavior().getReplicatedToAllIndex());
340 // Re-enable AppendEntries to the follower. The follower's log will be out of sync and it should
341 // should force the leader to install snapshot to re-sync the entire follower's log and state.
343 followerInstance.stopDropMessages(AppendEntries.class);
344 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, SnapshotComplete.class);
346 assertEquals("Follower term", currentTerm, follower1Context.getTermInformation().getCurrentTerm());
347 assertEquals("Follower journal lastIndex", lastIndex, follower1Context.getReplicatedLog().lastIndex());
348 assertEquals("Follower journal lastTerm", currentTerm, follower1Context.getReplicatedLog().lastTerm());
349 assertEquals("Follower commit index", lastIndex, follower1Context.getCommitIndex());
350 assertEquals("Follower applied state", expSnapshotState, followerInstance.getState());
352 testLog.info("testFollowerResyncWithMoreLeaderLogEntriesAndDownPeerAfterNonPersistentLeaderRestart ending");
356 public void testFollowerLeaderStateChanges() {
357 testLog.info("testFollowerLeaderStateChanges");
359 ActorRef roleChangeNotifier = factory.<MessageCollectorActor>createTestActor(
360 MessageCollectorActor.props().withDispatcher(Dispatchers.DefaultDispatcherId()),
361 factory.generateActorId("roleChangeNotifier"));
362 follower1Builder.roleChangeNotifier(roleChangeNotifier);
364 setupLeaderAndNonVotingFollower();
366 ((DefaultConfigParamsImpl)follower1Context.getConfigParams()).setElectionTimeoutFactor(2);
367 ((DefaultConfigParamsImpl)follower1Context.getConfigParams()).
368 setHeartBeatInterval(FiniteDuration.apply(100, TimeUnit.MILLISECONDS));
370 MessageCollectorActor.clearMessages(roleChangeNotifier);
371 follower1Actor.tell(ElectionTimeout.INSTANCE, ActorRef.noSender());
372 followerInstance.startDropMessages(AppendEntries.class);
374 LeaderStateChanged leaderStateChanged = MessageCollectorActor.expectFirstMatching(roleChangeNotifier,
375 LeaderStateChanged.class);
376 assertEquals("getLeaderId", null, leaderStateChanged.getLeaderId());
378 MessageCollectorActor.clearMessages(roleChangeNotifier);
379 followerInstance.stopDropMessages(AppendEntries.class);
381 leaderStateChanged = MessageCollectorActor.expectFirstMatching(roleChangeNotifier,
382 LeaderStateChanged.class);
383 assertEquals("getLeaderId", leaderId, leaderStateChanged.getLeaderId());
386 private void createNewLeaderActor() {
387 expSnapshotState.clear();
388 leaderActor = newTestRaftActor(leaderId, TestRaftActor.newBuilder().peerAddresses(peerAddresses).
389 config(leaderConfigParams).persistent(Optional.of(false)));
390 leaderInstance = leaderActor.underlyingActor();
391 leaderCollectorActor = leaderInstance.collectorActor();
392 waitUntilLeader(leaderActor);
393 leaderContext = leaderInstance.getRaftActorContext();
396 private void setupLeaderAndNonVotingFollower() {
397 snapshotBatchCount = 100;
400 // Set up a persisted ServerConfigurationPayload with the leader voting and the follower non-voting.
402 ServerConfigurationPayload persistedServerConfig = new ServerConfigurationPayload(Arrays.asList(
403 new ServerInfo(leaderId, true), new ServerInfo(follower1Id, false)));
404 ReplicatedLogImplEntry persistedServerConfigEntry = new ReplicatedLogImplEntry(0, initialTerm,
405 persistedServerConfig);
407 InMemoryJournal.addEntry(leaderId, 1, new UpdateElectionTerm(initialTerm, leaderId));
408 InMemoryJournal.addEntry(leaderId, 2, persistedServerConfigEntry);
409 InMemoryJournal.addEntry(follower1Id, 1, new UpdateElectionTerm(initialTerm, leaderId));
410 InMemoryJournal.addEntry(follower1Id, 2, persistedServerConfigEntry);
412 DefaultConfigParamsImpl followerConfigParams = newFollowerConfigParams();
413 follower1Actor = newTestRaftActor(follower1Id, follower1Builder.peerAddresses(
414 ImmutableMap.of(leaderId, testActorPath(leaderId))).config(followerConfigParams).
415 persistent(Optional.of(false)));
417 peerAddresses = ImmutableMap.<String, String>builder().
418 put(follower1Id, follower1Actor.path().toString()).build();
420 leaderConfigParams = newLeaderConfigParams();
421 leaderActor = newTestRaftActor(leaderId, TestRaftActor.newBuilder().peerAddresses(peerAddresses).
422 config(leaderConfigParams).persistent(Optional.of(false)));
424 followerInstance = follower1Actor.underlyingActor();
425 follower1CollectorActor = followerInstance.collectorActor();
427 leaderInstance = leaderActor.underlyingActor();
428 leaderCollectorActor = leaderInstance.collectorActor();
430 leaderContext = leaderInstance.getRaftActorContext();
431 follower1Context = followerInstance.getRaftActorContext();
433 waitUntilLeader(leaderActor);
435 // Verify leader's context after startup
437 currentTerm = initialTerm + 1;
438 assertEquals("Leader term", currentTerm, leaderContext.getTermInformation().getCurrentTerm());
439 assertEquals("Leader server config", Sets.newHashSet(persistedServerConfig.getServerConfig()),
440 Sets.newHashSet(leaderContext.getPeerServerInfo(true).getServerConfig()));
441 assertEquals("Leader isVotingMember", true, leaderContext.isVotingMember());
443 // Verify follower's context after startup
445 MessageCollectorActor.expectFirstMatching(follower1CollectorActor, AppendEntries.class);
446 assertEquals("Follower term", currentTerm, follower1Context.getTermInformation().getCurrentTerm());
447 assertEquals("Follower server config", Sets.newHashSet(persistedServerConfig.getServerConfig()),
448 Sets.newHashSet(follower1Context.getPeerServerInfo(true).getServerConfig()));
449 assertEquals("FollowerisVotingMember", false, follower1Context.isVotingMember());