Implement detection and potential mitigation of recovery failure cycles

src/DurableTask.Netherite/OrchestrationService/NetheriteOrchestrationServiceSettings.cs

@@ -206,6 +206,11 @@ public class NetheriteOrchestrationServiceSettings
         /// </summary>
         public int PartitionStartupTimeoutMinutes { get; set; } = 15;
 
+        /// <summary>
+        /// If true, disables the prefetching during replay.
+        /// </summary>
+        public bool DisablePrefetchDuringReplay { get; set; } = false;
+
         /// <summary>
         /// Allows attaching additional checkers and debuggers during testing.
         /// </summary>

src/DurableTask.Netherite/StorageLayer/Faster/AzureBlobs/BlobManager.cs

@@ -42,6 +42,8 @@ partial class BlobManager : ICheckpointManager, ILogCommitManager
         BlobUtilsV12.BlockBlobClients eventLogCommitBlob;
         BlobLeaseClient leaseClient;
 
+        BlobUtilsV12.BlockBlobClients checkpointCompletedBlob;
+
         BlobUtilsV12.BlobDirectory pageBlobPartitionDirectory;
         BlobUtilsV12.BlobDirectory blockBlobPartitionDirectory;
 
@@ -419,6 +421,7 @@ public async Task StartAsync()
 
                 this.eventLogCommitBlob = this.blockBlobPartitionDirectory.GetBlockBlobClient(CommitBlobName);
                 this.leaseClient = this.eventLogCommitBlob.WithRetries.GetBlobLeaseClient();
+                this.checkpointCompletedBlob = this.blockBlobPartitionDirectory.GetBlockBlobClient(this.GetCheckpointCompletedBlobName());
 
                 AzureStorageDevice createDevice(string name) =>
                      new AzureStorageDevice(name, this.blockBlobPartitionDirectory.GetSubDirectory(name), this.pageBlobPartitionDirectory.GetSubDirectory(name), this, true);
@@ -1057,10 +1060,11 @@ IEnumerable<Guid> ICheckpointManager.GetLogCheckpointTokens()
 
         internal async Task<bool> FindCheckpointsAsync(bool logIsEmpty)
         {
-            BlobUtilsV12.BlockBlobClients checkpointCompletedBlob = default;
+            string jsonString = null;
+            DateTimeOffset lastModified = default;
+
             try
             {
-                string jsonString = null;
 
                 if (this.UseLocalFiles)
                 {
@@ -1076,24 +1080,24 @@ internal async Task<bool> FindCheckpointsAsync(bool logIsEmpty)
                 else
                 {
                     var partDir = this.blockBlobPartitionDirectory;
-                    checkpointCompletedBlob = partDir.GetBlockBlobClient(this.GetCheckpointCompletedBlobName());
 
                     await this.PerformWithRetriesAsync(
                         semaphore: null,
                         requireLease: true,
                         "BlockBlobClient.DownloadContentAsync",
                         "FindCheckpointsAsync",
                         "",
-                        checkpointCompletedBlob.Name,
+                        this.checkpointCompletedBlob.Name,
                         1000,
                         true,
                         failIfReadonly: false, 
                         async (numAttempts) =>
                         {
                             try
                             {
-                                Azure.Response<BlobDownloadResult> downloadResult = await checkpointCompletedBlob.WithRetries.DownloadContentAsync();
+                                Azure.Response<BlobDownloadResult> downloadResult = await this.checkpointCompletedBlob.WithRetries.DownloadContentAsync();
                                 jsonString = downloadResult.Value.Content.ToString();
+                                lastModified = downloadResult.Value.Details.LastModified;
                                 this.CheckpointInfoETag = downloadResult.Value.Details.ETag;
                                 return 1;
                             }
@@ -1105,22 +1109,58 @@ await this.PerformWithRetriesAsync(
                         });
                 }
 
-                if (jsonString == null)
-                {
-                    return false;
-                }
-                else
-                {
-                    // read the fields from the json to update the checkpoint info
-                    JsonConvert.PopulateObject(jsonString, this.CheckpointInfo);
-                    return true;
-                }
             }
             catch (Exception e)
             {
-                this.HandleStorageError(nameof(FindCheckpointsAsync), "could not find any checkpoint", checkpointCompletedBlob.Name, e, true, this.PartitionErrorHandler.IsTerminated);
+                this.HandleStorageError(nameof(FindCheckpointsAsync), "could not find any checkpoint", this.checkpointCompletedBlob.Name, e, true, this.PartitionErrorHandler.IsTerminated);
+                throw;
+            }
+
+            if (jsonString == null)
+            {
+                return false;
+            }
+
+            try
+            {
+                // read the fields from the json to update the checkpoint info
+                JsonConvert.PopulateObject(jsonString, this.CheckpointInfo);
+            }
+            catch (JsonException e)
+            {
+                this.HandleStorageError(nameof(FindCheckpointsAsync), "could not parse json file describing last checkpoint", this.checkpointCompletedBlob.Name, e, true, false);
                 throw;
             }
+
+            if (this.CheckpointInfo.RecoveryAttempts > 0 || DateTimeOffset.UtcNow - lastModified > TimeSpan.FromMinutes(5))
+            {
+                this.CheckpointInfo.RecoveryAttempts++;
+
+                this.TraceHelper.FasterProgress($"Incremented recovery attempt counter to {this.CheckpointInfo.RecoveryAttempts} in {this.checkpointCompletedBlob.Name}.");
+
+                await this.WriteCheckpointMetadataAsync();
+
+                if (this.CheckpointInfo.RecoveryAttempts > 3 && this.CheckpointInfo.RecoveryAttempts < 30)
+                {
+                    this.TraceHelper.BoostTracing = true;
+                }
+            }
+
+            return true;
+        }
+
+        public async Task ClearRecoveryAttempts()
+        {
+            if (this.CheckpointInfo.RecoveryAttempts > 0)
+            {
+                this.CheckpointInfo.RecoveryAttempts = 0;
+
+                this.TraceHelper.BoostTracing = false;
+
+                await this.WriteCheckpointMetadataAsync();
+
+                this.TraceHelper.FasterProgress($"Cleared recovery attempt counter in {this.checkpointCompletedBlob.Name}.");
+            }
         }
 
         void ICheckpointManager.CommitIndexCheckpoint(Guid indexToken, byte[] commitMetadata)
@@ -1436,7 +1476,7 @@ await this.PerformWithRetriesAsync(
             }
         }
 
-        internal async Task FinalizeCheckpointCompletedAsync()
+        internal async Task WriteCheckpointMetadataAsync()
         {
             var jsonText = JsonConvert.SerializeObject(this.CheckpointInfo, Formatting.Indented);
             if (this.UseLocalFiles)

src/DurableTask.Netherite/StorageLayer/Faster/AzureBlobs/CheckpointInfo.cs

@@ -10,25 +10,46 @@ namespace DurableTask.Netherite.Faster
     [JsonObject]
     class CheckpointInfo
     {
+        /// <summary>
+        /// The FasterKV token for the last index checkpoint taken before this checkpoint.
+        /// </summary>
         [JsonProperty]
         public Guid IndexToken { get; set; }
 
+        /// <summary>
+        /// The FasterKV token for this checkpoint.
+        /// </summary>
         [JsonProperty]
         public Guid LogToken { get; set; }
 
+        /// <summary>
+        /// The FasterLog position for this checkpoint.
+        /// </summary>
         [JsonProperty]
         public long CommitLogPosition { get; set; }
 
+        /// <summary>
+        /// The input queue (event hubs) position for this checkpoint.
+        /// </summary>
         [JsonProperty]
         public long InputQueuePosition { get; set; }
 
+        /// <summary>
+        /// If the input queue position is a batch, the position within the batch.
+        /// </summary>
         [JsonProperty]
         public int InputQueueBatchPosition { get; set; }
 
+        /// <summary>
+        /// The input queue fingerprint for this checkpoint.
+        /// </summary>
         [JsonProperty]
         public string InputQueueFingerprint { get; set; }
 
-        [JsonProperty]
-        public long NumberInstances { get; set; }
+        /// <summary>
+        /// The number of recovery attempts that have been made for this checkpoint.
+        /// </summary>
+        //[JsonProperty]
+        public int RecoveryAttempts { get; set; }
     }
 }

src/DurableTask.Netherite/StorageLayer/Faster/FasterKV.cs

@@ -380,7 +380,7 @@ void RunTask() {
 
         public override async Task FinalizeCheckpointCompletedAsync(Guid guid)
         {
-            await this.blobManager.FinalizeCheckpointCompletedAsync();
+            await this.blobManager.WriteCheckpointMetadataAsync();
 
             if (this.cacheDebugger == null)
             {
@@ -739,7 +739,7 @@ public override async Task RunPrefetchSession(IAsyncEnumerable<TrackedObjectKey>
             long lastReport = 0;
             void ReportProgress(int elapsedMillisecondsThreshold)
             {
-                if (stopwatch.ElapsedMilliseconds - lastReport >= elapsedMillisecondsThreshold)
+                if (stopwatch.ElapsedMilliseconds - lastReport >= elapsedMillisecondsThreshold || this.TraceHelper.BoostTracing)
                 {
                     this.blobManager.TraceHelper.FasterProgress(
                         $"FasterKV PrefetchSession {sessionId} elapsed={stopwatch.Elapsed.TotalSeconds:F2}s issued={numberIssued} pending={maxConcurrency - prefetchSemaphore.CurrentCount} hits={numberHits} misses={numberMisses}");

src/DurableTask.Netherite/StorageLayer/Faster/FasterTraceHelper.cs

@@ -25,7 +25,9 @@ public FasterTraceHelper(ILogger logger, LogLevel logLevelLimit, ILogger perform
             this.partitionId = (int) partitionId;
         }
 
-        public bool IsTracingAtMostDetailedLevel => this.logLevelLimit == LogLevel.Trace;
+        public bool IsTracingAtMostDetailedLevel => this.logLevelLimit == LogLevel.Trace || this.BoostTracing;
+
+        public bool BoostTracing { get; set; }
 
         // ----- faster storage layer events
 
@@ -139,7 +141,7 @@ public void FasterProgress(Func<string> constructString)
 
         public void FasterStorageProgress(string details)
         {
-            if (this.logLevelLimit <= LogLevel.Trace)
+            if (this.logLevelLimit <= LogLevel.Trace || this.BoostTracing)
             {
                 this.logger.LogTrace("Part{partition:D2} {details}", this.partitionId, details);
                 EtwSource.Log.FasterStorageProgress(this.account, this.taskHub, this.partitionId, details, TraceUtils.AppName, TraceUtils.ExtensionVersion);

src/DurableTask.Netherite/StorageLayer/Faster/LogWorker.cs

@@ -207,7 +207,7 @@ protected override async Task Process(IList<PartitionUpdateEvent> batch)
             }
         }
 
-        public async Task ReplayCommitLog(long from, StoreWorker worker)
+        public async Task ReplayCommitLog(long from, StoreWorker worker, bool enablePrefetch)
         {
             // this procedure is called by StoreWorker during recovery. It replays all the events
             // that were committed to the log but are not reflected in the loaded store checkpoint.
@@ -220,7 +220,11 @@ public async Task ReplayCommitLog(long from, StoreWorker worker)
 
                 var fetchTask = this.FetchEvents(from, replayChannel.Writer, prefetchChannel.Writer);
                 var replayTask = Task.Run(() => this.ReplayEvents(replayChannel.Reader, worker));
-                var prefetchTask = Task.Run(() => worker.RunPrefetchSession(prefetchChannel.Reader.ReadAllAsync(this.cancellationToken)));
+
+                if (enablePrefetch)
+                {
+                    var prefetchTask = Task.Run(() => worker.RunPrefetchSession(prefetchChannel.Reader.ReadAllAsync(this.cancellationToken)));
+                }
 
                 await fetchTask;
                 await replayTask;

src/DurableTask.Netherite/StorageLayer/Faster/PartitionStorage.cs

@@ -20,9 +20,6 @@ class PartitionStorage : IPartitionState
         readonly ILogger logger;
         readonly ILogger performanceLogger;
         readonly MemoryTracker memoryTracker;
-        //readonly CloudStorageAccount storageAccount;
-        //readonly string localFileDirectory;
-        //readonly CloudStorageAccount pageBlobStorageAccount;
 
         Partition partition;
         BlobManager blobManager;
@@ -193,7 +190,10 @@ async Task TerminationWrapper(Task what)
                     if (this.log.TailAddress > (long)this.storeWorker.CommitLogPosition)
                     {
                         // replay log as the store checkpoint lags behind the log
-                        await this.TerminationWrapper(this.storeWorker.ReplayCommitLog(this.logWorker));
+
+                        bool disablePrefetch = this.blobManager.CheckpointInfo.RecoveryAttempts > 6 || this.settings.DisablePrefetchDuringReplay;
+
+                        await this.TerminationWrapper(this.storeWorker.ReplayCommitLog(this.logWorker, prefetch: !disablePrefetch));
                     }
                 }
                 catch (OperationCanceledException) when (this.partition.ErrorHandler.IsTerminated)
@@ -215,6 +215,8 @@ async Task TerminationWrapper(Task what)
                 }
 
                 this.TraceHelper.FasterProgress("Recovery complete");
+
+                await this.blobManager.ClearRecoveryAttempts();
             }
             this.blobManager.FaultInjector?.Started(this.blobManager);
             return this.storeWorker.InputQueuePosition;

src/DurableTask.Netherite/StorageLayer/Faster/StoreWorker.cs

@@ -681,16 +681,16 @@ protected override async Task Process(IList<PartitionEvent> batch)
             return target;
         }
 
-        public async Task ReplayCommitLog(LogWorker logWorker)
+        public async Task ReplayCommitLog(LogWorker logWorker, bool prefetch)
         {
             var startPosition = this.CommitLogPosition;
-            this.traceHelper.FasterProgress($"Replaying log from {startPosition}");
+            this.traceHelper.FasterProgress($"Replaying log from {startPosition} prefetch={prefetch}");
 
             var stopwatch = new System.Diagnostics.Stopwatch();
             stopwatch.Start();
 
             this.effectTracker.IsReplaying = true;
-            await logWorker.ReplayCommitLog(startPosition, this);
+            await logWorker.ReplayCommitLog(startPosition, this, prefetch);
             stopwatch.Stop();
             this.effectTracker.IsReplaying = false;