Fixed an issue where, when enabling HA in Amoro and switching between primary and standby nodes, the data loaded by the primary node was inconsistent with the database.

Author: 张文领

amoro-ams/src/main/java/org/apache/amoro/server/AmoroServiceContainer.java

@@ -136,7 +136,14 @@ public static void main(String[] args) {
         } catch (Exception e) {
           LOG.error("AMS start error", e);
         } finally {
-          service.disposeOptimizingService();
+          try {
+            service.disposeOptimizingService();
+          } catch (Exception e) {
+            LOG.warn("AMS dispose error", e);
+          } finally {
+            // if HA enabled, make sure the dispose complete signal is sent to ZK
+            service.signalDisposeComplete();
+          }
         }
       }
     } catch (Throwable t) {
@@ -153,6 +160,10 @@ public void waitFollowerShip() throws Exception {
     haContainer.waitFollowerShip();
   }
 
+  public void signalDisposeComplete() {
+    haContainer.signalDisposeComplete();
+  }
+
   public void startRestServices() throws Exception {
     EventsManager.getInstance();
     MetricManager.getInstance();

amoro-ams/src/main/java/org/apache/amoro/server/HighAvailabilityContainer.java

@@ -21,6 +21,7 @@
 import org.apache.amoro.client.AmsServerInfo;
 import org.apache.amoro.config.Configurations;
 import org.apache.amoro.properties.AmsHAProperties;
+import org.apache.amoro.shade.thrift.org.apache.commons.lang3.StringUtils;
 import org.apache.amoro.shade.zookeeper3.org.apache.curator.framework.CuratorFramework;
 import org.apache.amoro.shade.zookeeper3.org.apache.curator.framework.CuratorFrameworkFactory;
 import org.apache.amoro.shade.zookeeper3.org.apache.curator.framework.api.transaction.CuratorOp;
@@ -45,6 +46,8 @@ public class HighAvailabilityContainer implements LeaderLatchListener {
   private final CuratorFramework zkClient;
   private final String tableServiceMasterPath;
   private final String optimizingServiceMasterPath;
+  // path to signal that this node has completed ams dispose
+  private final String disposeCompletePath;
   private final AmsServerInfo tableServiceServerInfo;
   private final AmsServerInfo optimizingServiceServerInfo;
   private volatile CountDownLatch followerLatch;
@@ -59,6 +62,7 @@ public HighAvailabilityContainer(Configurations serviceConfig) throws Exception
       String haClusterName = serviceConfig.getString(AmoroManagementConf.HA_CLUSTER_NAME);
       tableServiceMasterPath = AmsHAProperties.getTableServiceMasterPath(haClusterName);
       optimizingServiceMasterPath = AmsHAProperties.getOptimizingServiceMasterPath(haClusterName);
+      disposeCompletePath = AmsHAProperties.getMasterReleaseConfirmPath(haClusterName);
       ExponentialBackoffRetry retryPolicy = new ExponentialBackoffRetry(1000, 3, 5000);
       this.zkClient =
           CuratorFrameworkFactory.builder()
@@ -90,6 +94,7 @@ public HighAvailabilityContainer(Configurations serviceConfig) throws Exception
       zkClient = null;
       tableServiceMasterPath = null;
       optimizingServiceMasterPath = null;
+      disposeCompletePath = null;
       tableServiceServerInfo = null;
       optimizingServiceServerInfo = null;
       // block follower latch forever when ha is disabled
@@ -102,6 +107,8 @@ public void waitLeaderShip() throws Exception {
     if (leaderLatch != null) {
       leaderLatch.await();
       if (leaderLatch.hasLeadership()) {
+        waitPreviousLeaderDisposeComplete();
+
         CuratorOp tableServiceMasterPathOp =
             zkClient
                 .transactionOp()
@@ -134,6 +141,83 @@ public void waitFollowerShip() throws Exception {
     LOG.info("Became the follower of AMS");
   }
 
+  public void waitPreviousLeaderDisposeComplete() throws Exception {
+    // 1、Create the path if it does not exist, to ensure it exists for future primary and standby
+    // node switchover.
+    if (zkClient.checkExists().forPath(disposeCompletePath) == null) {
+      createPathIfNeeded(disposeCompletePath);
+    }
+
+    // 2、Determine if there is a previous leader, or if it is different from current node.
+    boolean hasPreviousOtherLeader = false;
+    try {
+      byte[] masterData = zkClient.getData().forPath(tableServiceMasterPath);
+      if (masterData != null && masterData.length > 0) {
+        String masterInfoInZkNode = new String(masterData, StandardCharsets.UTF_8);
+        if (!StringUtils.isEmpty(masterInfoInZkNode)) {
+          try {
+            // If data cannot be parsed correctly, it indicates that the AMS service is starting for
+            // the first time.
+            AmsServerInfo previousLeaderInfo =
+                JacksonUtil.parseObject(masterInfoInZkNode, AmsServerInfo.class);
+            if (previousLeaderInfo != null) {
+              // If parsing succeeds, check if it's different from current node
+              String currentInfoStr = JacksonUtil.toJSONString(tableServiceServerInfo);
+              LOG.debug(
+                  "Current node info JSON: {}, ZK node info JSON: {}",
+                  currentInfoStr,
+                  masterInfoInZkNode);
+              if (!masterInfoInZkNode.equals(currentInfoStr)) {
+                hasPreviousOtherLeader = true;
+              } else {
+                LOG.debug(
+                    "Previous leader is the same as current node (self-restart)."
+                        + " No need to wait for dispose signal.");
+              }
+            }
+          } catch (Exception e) {
+            LOG.warn(
+                "Failed to parse master info from ZooKeeper: {}, treating as no previous leader",
+                masterInfoInZkNode,
+                e);
+            // If parsing fails, treat as no previous leader
+          }
+        }
+      }
+    } catch (KeeperException.NoNodeException e) {
+      // No previous leader node found, indicating that this is the first startup of ams.
+      LOG.debug("No previous leader node found, indicating that this is the first startup of ams.");
+    }
+
+    if (!hasPreviousOtherLeader) {
+      LOG.debug("No previous other master detected, start service immediately.");
+      return;
+    }
+
+    // If disposeCompletePath exists, the following scenarios may occur:
+    // 1) A primary-standby node switchover occurs, and the former primary
+    //      node has not completed the AMS dispose operation.
+    // 2) The previous primary node is unreachable due to network issues.
+    // 3) No primary-standby node switchover occurred, but ZK retains
+    // information about the previous primary node.
+    long startTime = System.currentTimeMillis();
+    int maxWaitTime = 30000; // 30s
+    while (System.currentTimeMillis() - startTime <= maxWaitTime) {
+      // At this point, the disposeCompletePath does not exist,
+      // indicating that the previous master node has completed
+      // the AMS service shutdown operation and deleted the path.
+      if (zkClient.checkExists().forPath(disposeCompletePath) == null) {
+        LOG.info("Previous leader has completed dispose. Proceeding.");
+        return;
+      }
+    }
+
+    LOG.debug(
+        "Timeout ({}ms) waiting for previous other leader to signal dispose complete. Proceeding anyway. "
+            + "This might indicate the previous leader is unresponsive.",
+        maxWaitTime);
+  }
+
   public void close() {
     if (leaderLatch != null) {
       try {
@@ -163,6 +247,27 @@ public void notLeader() {
     followerLatch.countDown();
   }
 
+  /**
+   * In HA mode, when the AMS service is stopped, delete the existing `disposeCompletePath` file
+   * from ZK to indicate that the AMS service has been terminated.
+   */
+  public void signalDisposeComplete() {
+    // when HA is disabled, do nothing
+    if (zkClient == null) {
+      return;
+    }
+
+    try {
+      if (zkClient.checkExists().forPath(disposeCompletePath) != null) {
+        zkClient.delete().forPath(disposeCompletePath);
+        return;
+      }
+      LOG.debug("ams dispose complete signal written.");
+    } catch (Exception e) {
+      LOG.warn("Failed to write dispose complete signal", e);
+    }
+  }
+
   private AmsServerInfo buildServerInfo(String host, int thriftBindPort, int restBindPort) {
     AmsServerInfo amsServerInfo = new AmsServerInfo();
     amsServerInfo.setHost(host);