[changelog][cdc] Add client side handling of new A/A version swap messages

Author: xunyin8

clients/da-vinci-client/src/main/java/com/linkedin/davinci/consumer/ChangelogClientConfig.java

@@ -81,6 +81,17 @@ public class ChangelogClientConfig<T extends SpecificRecord> {
    */
   private PubSubConsumerAdapterFactory<? extends PubSubConsumerAdapter> pubSubConsumerAdapterFactory;
   private PubSubContext pubSubContext;
+  private boolean versionSwapByControlMessageEnabled = false;
+  /**
+   * Client region name used for filtering version swap messages from other regions in A/A setup. The client will only
+   * react to version swap messages with the same source region as the client region name.
+   */
+  private String clientRegionName = "";
+  /**
+   * Total region count used for version swap in A/A setup. Each subscribed partition need to receive this many
+   * corresponding version swap messages before it can safely go to the new version to ensure data completeness.
+   */
+  private int totalRegionCount = 1;
 
   public ChangelogClientConfig(String storeName) {
     this.innerClientConfig = new ClientConfig<>(storeName);
@@ -327,6 +338,33 @@ public boolean isStateful() {
     return this.isStateful;
   }
 
+  public boolean isVersionSwapByControlMessageEnabled() {
+    return this.versionSwapByControlMessageEnabled;
+  }
+
+  public ChangelogClientConfig setVersionSwapByControlMessageEnabled(boolean isVersionSwapByControlMessageEnabled) {
+    this.versionSwapByControlMessageEnabled = isVersionSwapByControlMessageEnabled;
+    return this;
+  }
+
+  public String getClientRegionName() {
+    return this.clientRegionName;
+  }
+
+  public ChangelogClientConfig setClientRegionName(String clientRegionName) {
+    this.clientRegionName = clientRegionName;
+    return this;
+  }
+
+  public int getTotalRegionCount() {
+    return this.totalRegionCount;
+  }
+
+  public ChangelogClientConfig setTotalRegionCount(int totalRegionCount) {
+    this.totalRegionCount = totalRegionCount;
+    return this;
+  }
+
   public static <V extends SpecificRecord> ChangelogClientConfig<V> cloneConfig(ChangelogClientConfig<V> config) {
     ChangelogClientConfig<V> newConfig = new ChangelogClientConfig<V>().setStoreName(config.getStoreName())
         .setLocalD2ZkHosts(config.getLocalD2ZkHosts())
@@ -354,7 +392,10 @@ public static <V extends SpecificRecord> ChangelogClientConfig<V> cloneConfig(Ch
         // Store version should not be cloned
         .setStoreVersion(null)
         // Is stateful config should not be cloned
-        .setIsStateful(false);
+        .setIsStateful(false)
+        .setVersionSwapByControlMessageEnabled(config.isVersionSwapByControlMessageEnabled())
+        .setClientRegionName(config.getClientRegionName())
+        .setTotalRegionCount(config.getTotalRegionCount());
     return newConfig;
   }
 

clients/da-vinci-client/src/main/java/com/linkedin/davinci/consumer/VeniceAfterImageConsumerImpl.java

@@ -3,6 +3,7 @@
 import com.linkedin.davinci.repository.NativeMetadataRepositoryViewAdapter;
 import com.linkedin.venice.exceptions.VeniceException;
 import com.linkedin.venice.kafka.protocol.ControlMessage;
+import com.linkedin.venice.kafka.protocol.VersionSwap;
 import com.linkedin.venice.kafka.protocol.enums.ControlMessageType;
 import com.linkedin.venice.pubsub.api.DefaultPubSubMessage;
 import com.linkedin.venice.pubsub.api.PubSubConsumerAdapter;
@@ -73,7 +74,8 @@ protected VeniceAfterImageConsumerImpl(
         storeRepository,
         storeName,
         changelogClientConfig.getConsumerName(),
-        this.changeCaptureStats);
+        this.changeCaptureStats,
+        changelogClientConfig.isVersionSwapByControlMessageEnabled());
   }
 
   @Override
@@ -178,6 +180,100 @@ protected static void adjustSeekCheckPointsBasedOnHeartbeats(
     }
   }
 
+  /**
+   * Similar to {@link #internalSeekToEndOfPush} this can also be optimized later for a faster find.
+   */
+  @Override
+  protected CompletableFuture<Void> internalFindNewVersionCheckpoints(
+      String oldVersionTopic,
+      String newVersionTopic,
+      long generationId,
+      Set<Integer> partitions) {
+    if (partitions.isEmpty()) {
+      return CompletableFuture.completedFuture(null);
+    }
+    return CompletableFuture.supplyAsync(() -> {
+      boolean lockAcquired = false;
+      Map<Integer, VeniceChangeCoordinate> checkpoints = new HashMap<>();
+      try {
+        synchronized (internalSeekConsumer) {
+          PubSubConsumerAdapter consumerAdapter = internalSeekConsumer.get();
+          consumerAdapter.batchUnsubscribe(consumerAdapter.getAssignment());
+          Map<PubSubTopicPartition, List<DefaultPubSubMessage>> polledResults;
+          Map<Integer, Boolean> versionSwapConsumedPerPartitionMap = new HashMap<>();
+          for (Integer partition: partitions) {
+            versionSwapConsumedPerPartitionMap.put(partition, false);
+          }
+          List<PubSubTopicPartition> topicPartitionList = getPartitionListToSubscribe(
+              partitions,
+              Collections.EMPTY_SET,
+              pubSubTopicRepository.getTopic(newVersionTopic));
+
+          for (PubSubTopicPartition topicPartition: topicPartitionList) {
+            consumerAdapter.subscribe(topicPartition, PubSubSymbolicPosition.EARLIEST);
+          }
+
+          // Poll until we receive the desired version swap message in the new version topic for each partition
+          LOGGER.info(
+              "Polling for version swap messages in: {} with generation id: {} for partitions: {}",
+              newVersionTopic,
+              generationId,
+              partitions);
+          while (!versionSwapConsumedPerPartitionMap.values().stream().allMatch(x -> x)) {
+            polledResults = consumerAdapter.poll(5000L);
+            for (Map.Entry<PubSubTopicPartition, List<DefaultPubSubMessage>> entry: polledResults.entrySet()) {
+              PubSubTopicPartition pubSubTopicPartition = entry.getKey();
+              List<DefaultPubSubMessage> messageList = entry.getValue();
+              for (DefaultPubSubMessage message: messageList) {
+                if (message.getKey().isControlMessage()) {
+                  ControlMessage controlMessage = (ControlMessage) message.getValue().getPayloadUnion();
+                  ControlMessageType controlMessageType = ControlMessageType.valueOf(controlMessage);
+                  if (controlMessageType.equals(ControlMessageType.VERSION_SWAP)) {
+                    VersionSwap versionSwap = (VersionSwap) controlMessage.getControlMessageUnion();
+                    // In theory just matching the generation id and source region should be sufficient but just to be
+                    // safe we will match all fields
+                    if (versionSwap.getGenerationId() == generationId
+                        && versionSwap.getSourceRegion().toString().equals(clientRegionName)
+                        && oldVersionTopic.equals(versionSwap.getOldServingVersionTopic().toString())
+                        && newVersionTopic.equals(versionSwap.getNewServingVersionTopic().toString())) {
+                      LOGGER.info(
+                          "Found corresponding version swap message for partition: {}",
+                          pubSubTopicPartition.getPartitionNumber());
+                      versionSwapConsumedPerPartitionMap.put(pubSubTopicPartition.getPartitionNumber(), true);
+                      VeniceChangeCoordinate coordinate = new VeniceChangeCoordinate(
+                          pubSubTopicPartition.getPubSubTopic().getName(),
+                          message.getPosition(),
+                          pubSubTopicPartition.getPartitionNumber());
+                      checkpoints.put(pubSubTopicPartition.getPartitionNumber(), coordinate);
+                      // We are done with this partition
+                      consumerAdapter.unSubscribe(pubSubTopicPartition);
+                      break;
+                    }
+                  }
+                }
+              }
+            }
+          }
+          LOGGER.info(
+              "Found all version swap messages in: {} with generation id: {} for partitions: {}",
+              newVersionTopic,
+              generationId,
+              partitions);
+        }
+        // We cannot change the subscription here because the consumer might not finish polling all the messages in the
+        // old version topic yet. We can acquire the lock and update the VersionSwapMessageState.
+        subscriptionLock.writeLock().lock();
+        lockAcquired = true;
+        versionSwapMessageState.setNewTopicCheckpoints(checkpoints);
+      } finally {
+        if (lockAcquired) {
+          subscriptionLock.writeLock().unlock();
+        }
+      }
+      return null;
+    }, seekExecutorService);
+  }
+
   protected CompletableFuture<Void> internalSeekToEndOfPush(
       Set<Integer> partitions,
       PubSubTopic targetTopic,