Correctly record SLM state during master failover

server/src/main/java/org/elasticsearch/snapshots/RegisteredPolicySnapshots.java

@@ -44,7 +44,7 @@
  * cluster state as custom metadata. When a snapshot is started by SLM, it is added to this set. Upon completion,
  * is it removed. If a snapshot does not record its failure in SnapshotLifecycleStats, likely due to a master shutdown,
  * it will not be removed from the registered set. A subsequent snapshot will then find that a registered snapshot
- * is no longer running and will infer that it failed, updating SnapshotLifecycleStats accordingly.
+ * is no longer running and update SnapshotLifecycleStats based on the status of the snapshot.
  */
 public class RegisteredPolicySnapshots implements Metadata.ProjectCustom {
 

x-pack/plugin/slm/src/internalClusterTest/java/org/elasticsearch/xpack/slm/SLMSnapshotBlockingIntegTests.java

@@ -15,6 +15,7 @@
 import org.elasticsearch.action.admin.cluster.snapshots.restore.TransportRestoreSnapshotAction;
 import org.elasticsearch.action.admin.cluster.snapshots.status.SnapshotStatus;
 import org.elasticsearch.action.admin.cluster.snapshots.status.SnapshotsStatusResponse;
+import org.elasticsearch.action.support.master.AcknowledgedResponse;
 import org.elasticsearch.cluster.SnapshotsInProgress;
 import org.elasticsearch.cluster.health.ClusterHealthStatus;
 import org.elasticsearch.cluster.routing.UnassignedInfo;
@@ -72,6 +73,7 @@ public class SLMSnapshotBlockingIntegTests extends AbstractSnapshotIntegTestCase
     private static final String NEVER_EXECUTE_CRON_SCHEDULE = "* * * 31 FEB ? *";
 
     static final String REPO = "my-repo";
+    List<String> masterNodeNames = null;
     List<String> dataNodeNames = null;
 
     @Override
@@ -85,7 +87,7 @@ protected Settings nodeSettings(int nodeOrdinal, Settings otherSettings) {
     @Before
     public void ensureClusterNodes() {
         logger.info("--> starting enough nodes to ensure we have enough to safely stop for tests");
-        internalCluster().startMasterOnlyNodes(2);
+        masterNodeNames = internalCluster().startMasterOnlyNodes(2);
         dataNodeNames = internalCluster().startDataOnlyNodes(2);
         ensureGreen();
     }
@@ -329,6 +331,185 @@ public void testRetentionWithMultipleRepositories() throws Exception {
         testUnsuccessfulSnapshotRetention(randomBoolean());
     }
 
+    // Test that SLM stats and lastSuccess/lastFailure are correctly updated with master shutdown
+    public void testSLMWithMasterShutdown() throws Exception {
+        final String indexName = "test";
+        final String policyName = "test-policy";
+        int clusterSize = masterNodeNames.size() + dataNodeNames.size();
+        indexRandomDocs(indexName, 20);
+        createRepository(REPO, "mock");
+
+        createSnapshotPolicy(
+            policyName,
+            "snap",
+            NEVER_EXECUTE_CRON_SCHEDULE,
+            REPO,
+            indexName,
+            true,
+            false,
+            new SnapshotRetentionConfiguration(TimeValue.ZERO, null, null)
+        );
+
+        // block snapshot from completing
+        blockMasterFromFinalizingSnapshotOnIndexFile(REPO);
+
+        // first SLM execution
+        final String snapshotName = executePolicy(policyName);
+        final String initialMaster = internalCluster().getMasterName();
+        waitForBlock(initialMaster, REPO);
+
+        // restart master
+        internalCluster().restartNode(initialMaster);
+        ensureStableCluster(clusterSize);
+        awaitNoMoreRunningOperations();
+
+        // ensure snapshot is completed successfully after master failover
+        assertBusy(() -> {
+            final SnapshotInfo snapshotInfo;
+            try {
+                GetSnapshotsResponse snapshotsStatusResponse = clusterAdmin().prepareGetSnapshots(TEST_REQUEST_TIMEOUT, REPO)
+                    .setSnapshots(snapshotName)
+                    .get();
+                snapshotInfo = snapshotsStatusResponse.getSnapshots().get(0);
+            } catch (SnapshotMissingException sme) {
+                throw new AssertionError(sme);
+            }
+            assertEquals(SnapshotState.SUCCESS, snapshotInfo.state());
+        }, 30L, TimeUnit.SECONDS);
+        assertSnapshotSuccessful(snapshotName);
+
+        // the SLM policy metadata has not been updated due to master shutdown
+        assertBusy(() -> {
+            SnapshotLifecyclePolicyItem policy = client().execute(
+                GetSnapshotLifecycleAction.INSTANCE,
+                new GetSnapshotLifecycleAction.Request(TEST_REQUEST_TIMEOUT, TEST_REQUEST_TIMEOUT, policyName)
+            ).get().getPolicies().getFirst();
+            assertNull(policy.getLastSuccess());
+            assertNull(policy.getLastFailure());
+            assertEquals(0, policy.getPolicyStats().getSnapshotFailedCount());
+            assertEquals(0, policy.getPolicyStats().getSnapshotTakenCount());
+        });
+
+        // 2nd SLM execution, it should pick up the last missing stats
+        String snapshotSecond = executePolicy(policyName);
+
+        awaitNoMoreRunningOperations();
+        assertSnapshotSuccessful(snapshotSecond);
+
+        // stats should have 2 successful snapshots, 1 from the new snapshot and 1 from previous success
+        assertBusy(() -> {
+            SnapshotLifecyclePolicyItem policy = client().execute(
+                GetSnapshotLifecycleAction.INSTANCE,
+                new GetSnapshotLifecycleAction.Request(TEST_REQUEST_TIMEOUT, TEST_REQUEST_TIMEOUT, policyName)
+            ).get().getPolicies().getFirst();
+            assertNull(policy.getLastFailure());
+            assertNotNull(policy.getLastSuccess());
+            assertEquals(snapshotSecond, policy.getLastSuccess().getSnapshotName());
+            assertEquals(0, policy.getPolicyStats().getSnapshotFailedCount());
+            assertEquals(2, policy.getPolicyStats().getSnapshotTakenCount());
+        });
+    }
+
+    public void testSLMWithMasterShutdownAndDeletedSnapshot() throws Exception {
+        final String indexName = "test";
+        final String policyName = "test-policy";
+        int clusterSize = masterNodeNames.size() + dataNodeNames.size();
+        indexRandomDocs(indexName, 20);
+        createRepository(REPO, "mock");
+
+        createSnapshotPolicy(
+            policyName,
+            "snap",
+            NEVER_EXECUTE_CRON_SCHEDULE,
+            REPO,
+            indexName,
+            true,
+            false,
+            new SnapshotRetentionConfiguration(TimeValue.ZERO, null, null)
+        );
+
+        // block snapshot from completing
+        blockMasterFromFinalizingSnapshotOnIndexFile(REPO);
+
+        // first SLM execution
+        final String snapshotName = executePolicy(policyName);
+        final String initialMaster = internalCluster().getMasterName();
+        waitForBlock(initialMaster, REPO);
+
+        // restart master
+        internalCluster().restartNode(initialMaster);
+        ensureStableCluster(clusterSize);
+        awaitNoMoreRunningOperations();
+
+        // ensure snapshot is completed successfully after master failover
+        assertBusy(() -> {
+            final SnapshotInfo snapshotInfo;
+            try {
+                GetSnapshotsResponse snapshotsStatusResponse = clusterAdmin().prepareGetSnapshots(TEST_REQUEST_TIMEOUT, REPO)
+                    .setSnapshots(snapshotName)
+                    .get();
+                snapshotInfo = snapshotsStatusResponse.getSnapshots().get(0);
+            } catch (SnapshotMissingException sme) {
+                throw new AssertionError(sme);
+            }
+            assertEquals(SnapshotState.SUCCESS, snapshotInfo.state());
+        }, 30L, TimeUnit.SECONDS);
+        assertSnapshotSuccessful(snapshotName);
+
+        // the SLM policy metadata has not been updated due to master shutdown
+        assertBusy(() -> {
+            SnapshotLifecyclePolicyItem policy = client().execute(
+                GetSnapshotLifecycleAction.INSTANCE,
+                new GetSnapshotLifecycleAction.Request(TEST_REQUEST_TIMEOUT, TEST_REQUEST_TIMEOUT, policyName)
+            ).get().getPolicies().getFirst();
+            assertNull(policy.getLastSuccess());
+            assertNull(policy.getLastFailure());
+            assertEquals(0, policy.getPolicyStats().getSnapshotFailedCount());
+            assertEquals(0, policy.getPolicyStats().getSnapshotTakenCount());
+        });
+
+        // delete the snapshot, simulate missing snapshot from repo
+        assertBusy(() -> {
+            AcknowledgedResponse response = clusterAdmin().prepareDeleteSnapshot(TEST_REQUEST_TIMEOUT, REPO, snapshotName).get();
+            assertTrue(response.isAcknowledged());
+        });
+
+        // 2nd SLM execution, it should pick up the last missing stats
+        String snapshotSecond = executePolicy(policyName);
+
+        awaitNoMoreRunningOperations();
+        assertSnapshotSuccessful(snapshotSecond);
+
+        // stats should have 1 successful and 1 failed snapshot, the deleted snapshot is inferred failure
+        assertBusy(() -> {
+            SnapshotLifecyclePolicyItem policy = client().execute(
+                GetSnapshotLifecycleAction.INSTANCE,
+                new GetSnapshotLifecycleAction.Request(TEST_REQUEST_TIMEOUT, TEST_REQUEST_TIMEOUT, policyName)
+            ).get().getPolicies().getFirst();
+            assertNotNull(policy.getLastSuccess());
+            assertEquals(snapshotSecond, policy.getLastSuccess().getSnapshotName());
+            assertNotNull(policy.getLastFailure());
+            assertEquals(snapshotName, policy.getLastFailure().getSnapshotName());
+            assertEquals(1, policy.getPolicyStats().getSnapshotFailedCount());
+            assertEquals(1, policy.getPolicyStats().getSnapshotTakenCount());
+        });
+    }
+
+    private void assertSnapshotSuccessful(String snapshotName) throws Exception {
+        assertBusy(() -> {
+            final SnapshotInfo snapshotInfo;
+            try {
+                GetSnapshotsResponse snapshotsStatusResponse = clusterAdmin().prepareGetSnapshots(TEST_REQUEST_TIMEOUT, REPO)
+                    .setSnapshots(snapshotName)
+                    .get();
+                snapshotInfo = snapshotsStatusResponse.getSnapshots().get(0);
+            } catch (SnapshotMissingException sme) {
+                throw new AssertionError(sme);
+            }
+            assertEquals(SnapshotState.SUCCESS, snapshotInfo.state());
+        });
+    }
+
     private void testUnsuccessfulSnapshotRetention(boolean partialSuccess) throws Exception {
         final String indexName = "test-idx";
         final String policyId = "test-policy";

x-pack/plugin/slm/src/main/java/org/elasticsearch/xpack/slm/SnapshotLifecycleService.java

@@ -232,6 +232,17 @@ public static String getJobId(SnapshotLifecyclePolicyMetadata policyMeta) {
         return policyMeta.getPolicy().getId() + "-" + policyMeta.getVersion();
     }
 
+    /**
+     * Gets the policy name from a job id, which is expected to be in the format
+     * {@code <policyid>-<version>}. This method extracts the policy id by
+     * removing the version part (the last part after the last dash).
+     */
+    public static String getPolicyId(String jobId) {
+        int lastDashIndex = jobId.lastIndexOf('-');
+        assert lastDashIndex != -1 : "Invalid job id format: " + jobId;
+        return jobId.substring(0, lastDashIndex);
+    }
+
     /**
      * Cancel all scheduled snapshot jobs
      */