fix: return deserialized operation results

Author: zhongkechen

docs/migration-1.x-to-2.x.md

@@ -185,27 +185,29 @@ assertThrows(IllegalStateException.class, future::get);
 What changes in practice:
 
 - Serialization problems now fail on first execution instead of surfacing later on replay.
+- Operation results can be returned after a SerDes round-trip so first execution matches replay.
 - Custom `SerDes` implementations must be able to deserialize SDK-managed values they serialize.
 - Child-context results are validated consistently, including virtual child-context paths.
 
 This is usually a correctness improvement, but it can surface previously hidden `SerDes` bugs during upgrade.
 
 ### New opt-out configuration
 
-If your workload is very performance-sensitive and you need to skip the extra validation deserialize pass, you can opt out:
+If your workload is very performance-sensitive and you need to skip the extra deserialize pass, you can opt out:
 
 ```java
 @Override
 protected DurableConfig createConfiguration() {
     return DurableConfig.builder()
-            .withSerializationRoundTripValidation(false)
+            .withDeserializeAfterSerialization(false)
             .build();
 }
 ```
 
 Use that carefully:
 
-- Disabling validation can hide serialization bugs until replay.
+- Disabling this can hide serialization bugs until replay.
+- First execution may return the raw result shape instead of the replay result shape.
 - Custom `SerDes` implementations are still expected to be round-trip safe.
 
 ## Recommended Validation After Upgrading
@@ -226,6 +228,6 @@ Most upgrades are straightforward:
 - Logger metadata moves to `executionArn`, `operationId`, and `operationName`
 - Replay-sensitive logging becomes per-context, `isReplaying()` moves to `DurableContext`, and step logs are no longer replay-suppressed
 - Validation failures now throw `IllegalStateException`
-- Serialization round-trip problems surface earlier by default, with an opt-out via `withSerializationRoundTripValidation(false)`
+- Serialization round-trip problems surface earlier by default, with an opt-out via `withDeserializeAfterSerialization(false)`
 
 If you update those areas first, the `1.x` to `2.x` migration should be low risk.

examples/src/main/java/software/amazon/lambda/durable/examples/general/CustomConfigExample.java

@@ -31,7 +31,7 @@
  *   <li>Automatic region detection with fallback to us-east-1 for testing environments
  *   <li>Environment variable credentials provider
  *   <li>Custom SerDes with snake_case property naming
- *   <li>Optional round-trip validation toggle for performance-sensitive workloads
+ *   <li>Optional post-serialization deserialization toggle for performance-sensitive workloads
  * </ul>
  */
 public class CustomConfigExample extends DurableHandler<String, String> {
@@ -69,8 +69,8 @@ protected DurableConfig createConfiguration() {
         return DurableConfig.builder()
                 .withDurableExecutionClient(durableClient)
                 .withSerDes(customSerDes)
-                // Disable the extra deserialize pass if your workload is sensitive to the added validation cost.
-                .withSerializationRoundTripValidation(false)
+                // Disable the extra deserialize pass if your workload is sensitive to the added cost.
+                .withDeserializeAfterSerialization(false)
                 .build();
     }
 

sdk/src/main/java/software/amazon/lambda/durable/DurableConfig.java

@@ -98,7 +98,7 @@ public final class DurableConfig {
     private final LoggerConfig loggerConfig;
     private final PollingStrategy pollingStrategy;
     private final Duration checkpointDelay;
-    private final boolean validateSerializationRoundTrip;
+    private final boolean deserializeAfterSerialization;
     private final PluginRunner pluginRunner;
 
     private DurableConfig(Builder builder) {
@@ -110,7 +110,7 @@ private DurableConfig(Builder builder) {
         this.loggerConfig = Objects.requireNonNullElseGet(builder.loggerConfig, LoggerConfig::defaults);
         this.pollingStrategy = Objects.requireNonNullElse(builder.pollingStrategy, PollingStrategies.Presets.DEFAULT);
         this.checkpointDelay = Objects.requireNonNullElseGet(builder.checkpointDelay, () -> Duration.ofSeconds(0));
-        this.validateSerializationRoundTrip = builder.validateSerializationRoundTrip;
+        this.deserializeAfterSerialization = builder.deserializeAfterSerialization;
         this.pluginRunner = builder.plugins.isEmpty() ? PluginRunner.noOp() : new PluginRunner(builder.plugins);
 
         validateConfiguration();
@@ -189,16 +189,16 @@ public Duration getCheckpointDelay() {
     }
 
     /**
-     * Gets whether serialized operation data should be immediately deserialized to verify round-trip compatibility.
+     * Gets whether serialized operation data should be deserialized immediately after serialization.
      *
-     * <p>When enabled, the SDK validates serialized operation results and exceptions before checkpointing them. This
-     * catches incompatible SerDes behavior early at the cost of an extra deserialize pass. Defaults to true, and custom
-     * SerDes implementations are still expected to be round-trip safe even if this validation is disabled.
+     * <p>When enabled, the SDK returns deserialized operation results so first execution matches replay, and validates
+     * serialized exceptions before checkpointing them. Defaults to true, and custom SerDes implementations are still
+     * expected to be round-trip safe even if this behavior is disabled.
      *
-     * @return true when round-trip serialization validation is enabled
+     * @return true when serialized data should be deserialized immediately
      */
-    public boolean shouldValidateSerializationRoundTrip() {
-        return validateSerializationRoundTrip;
+    public boolean shouldDeserializeAfterSerialization() {
+        return deserializeAfterSerialization;
     }
 
     /**
@@ -308,7 +308,7 @@ public static final class Builder {
         private LoggerConfig loggerConfig;
         private PollingStrategy pollingStrategy;
         private Duration checkpointDelay;
-        private boolean validateSerializationRoundTrip = true;
+        private boolean deserializeAfterSerialization = true;
         private List<DurableExecutionPlugin> plugins = new ArrayList<>();
 
         public Builder() {}
@@ -420,18 +420,16 @@ public Builder withCheckpointDelay(Duration duration) {
         }
 
         /**
-         * Controls whether the SDK immediately deserializes serialized results and exceptions to verify they can be
-         * read back before checkpointing.
+         * Controls whether the SDK immediately deserializes serialized operation data after serialization.
          *
-         * <p>This validation is enabled by default. Disable it only to avoid the extra deserialize pass when the
-         * additional safety check is too expensive for your workload. Custom SerDes implementations are still expected
-         * to round-trip SDK-managed values correctly.
+         * <p>This is enabled by default. When enabled, operation results are returned after a SerDes round-trip so
+         * first execution returns the same value shape as replay, and exceptions are checked before checkpointing.
          *
-         * @param validateSerializationRoundTrip true to validate serialized data with an immediate deserialize pass
+         * @param deserializeAfterSerialization true to deserialize serialized data immediately
          * @return This builder
          */
-        public Builder withSerializationRoundTripValidation(boolean validateSerializationRoundTrip) {
-            this.validateSerializationRoundTrip = validateSerializationRoundTrip;
+        public Builder withDeserializeAfterSerialization(boolean deserializeAfterSerialization) {
+            this.deserializeAfterSerialization = deserializeAfterSerialization;
             return this;
         }
 

sdk/src/main/java/software/amazon/lambda/durable/DurableHandler.java

@@ -117,7 +117,7 @@ public DurableConfig getConfiguration() {
      *         .withDurableExecutionClient(durableClient)
      *         .withSerDes(customSerDes)  // Optional: custom SerDes for user data
      *         .withExecutorService(customExecutor)  // Optional: custom thread pool
-     *         .withSerializationRoundTripValidation(false)  // Optional: skip extra validation deserialize pass
+     *         .withDeserializeAfterSerialization(false)  // Optional: skip immediate deserialize pass
      *         .build();
      * }
      * }</pre>

sdk/src/main/java/software/amazon/lambda/durable/operation/ChildContextOperation.java

@@ -149,7 +149,7 @@ private void executeChildContext() {
     }
 
     private void handleChildContextSuccess(T result) {
-        var serialized = serializeResult(result);
+        var serializedResult = serializeAndDeserializeResult(result);
 
         if (replayChildren.get() || isVirtual || parentOperation != null && parentOperation.isOperationCompleted()) {
             // Skip checkpointing if
@@ -158,10 +158,10 @@ private void handleChildContextSuccess(T result) {
             // - replaying a SUCCEEDED child with replayChildren=true — skip checkpointing.
             // - nestingType is FLAT
             // Mark the completableFuture completed so get() doesn't block waiting for a checkpoint response.
-            cachedOperationResult.set(DeserializedOperationResult.succeeded(result));
+            cachedOperationResult.set(DeserializedOperationResult.succeeded(serializedResult.deserialized()));
             markAlreadyCompleted();
         } else {
-            checkpointSuccess(result, serialized);
+            checkpointSuccess(serializedResult.deserialized(), serializedResult.serialized());
         }
     }
 

sdk/src/main/java/software/amazon/lambda/durable/operation/MapOperation.java

@@ -178,21 +178,22 @@ protected void replay(Operation existing) {
     @Override
     protected void handleCompletion(ConcurrencyCompletionStatus concurrencyCompletionStatus) {
         this.cachedResult = constructMapResult(concurrencyCompletionStatus);
-        var serialized = serializeResult(cachedResult);
-        var serializedBytes = serialized.getBytes(StandardCharsets.UTF_8);
+        var serializedResult = serializeAndDeserializeResult(cachedResult);
+        this.cachedResult = serializedResult.deserialized();
+        var serializedBytes = serializedResult.serialized().getBytes(StandardCharsets.UTF_8);
 
         if (serializedBytes.length < LARGE_RESULT_THRESHOLD) {
             sendOperationUpdate(OperationUpdate.builder()
                     .action(OperationAction.SUCCEED)
                     .subType(getSubType().getValue())
-                    .payload(serialized));
+                    .payload(serializedResult.serialized()));
         } else {
             // Large result: checkpoint with stripped payload + replayChildren flag
-            var strippedResult = serializeResult(stripMapResult(cachedResult));
+            var strippedResult = serializeAndDeserializeResult(stripMapResult(cachedResult));
             sendOperationUpdate(OperationUpdate.builder()
                     .action(OperationAction.SUCCEED)
                     .subType(getSubType().getValue())
-                    .payload(strippedResult)
+                    .payload(strippedResult.serialized())
                     .contextOptions(
                             ContextOptions.builder().replayChildren(true).build()));
         }

sdk/src/main/java/software/amazon/lambda/durable/operation/ParallelOperation.java

@@ -78,13 +78,15 @@ protected void handleCompletion(ConcurrencyCompletionStatus concurrencyCompletio
         int skippedCount = items.size() - succeededCount - failedCount;
         cachedResult = new ParallelResult(
                 items.size(), succeededCount, failedCount, skippedCount, concurrencyCompletionStatus, statuses);
+        var serializedResult = serializeAndDeserializeResult(cachedResult);
+        cachedResult = serializedResult.deserialized();
 
         // Branches added after checkpoint will not exist in the checkpointed result, but they'll be in the returned
         // value from get() method.
         sendOperationUpdate(OperationUpdate.builder()
                 .action(OperationAction.SUCCEED)
                 .subType(getSubType().getValue())
-                .payload(serializeResult(cachedResult))
+                .payload(serializedResult.serialized())
                 .contextOptions(ContextOptions.builder().replayChildren(true).build()));
     }
 

sdk/src/main/java/software/amazon/lambda/durable/operation/SerializableDurableOperation.java

@@ -34,6 +34,8 @@
 public abstract class SerializableDurableOperation<T> extends BaseDurableOperation implements DurableFuture<T> {
     private static final Logger logger = LoggerFactory.getLogger(SerializableDurableOperation.class);
 
+    protected record SerializedResult<T>(String serialized, T deserialized) {}
+
     private final TypeToken<T> resultTypeToken;
     private final SerDes resultSerDes;
 
@@ -95,17 +97,18 @@ protected T deserializeResult(String result) {
     }
 
     /**
-     * Serializes the result to a string.
+     * Serializes the result and returns the value that should be exposed to callers.
+     *
+     * <p>Use this for operations that cache a first-execution result instead of reading it back from checkpoint data.
+     * This keeps first execution consistent with replay when a SerDes normalizes or otherwise changes the value.
      *
      * @param result the result to serialize
-     * @return the serialized string
+     * @return the serialized string and the deserialized result
      */
-    protected String serializeResult(T result) {
+    protected SerializedResult<T> serializeAndDeserializeResult(T result) {
         var serialized = resultSerDes.serialize(result);
-        if (shouldValidateSerializationRoundTrip()) {
-            deserializeResult(serialized);
-        }
-        return serialized;
+        var deserialized = shouldDeserializeAfterSerialization() ? deserializeResult(serialized) : result;
+        return new SerializedResult<>(serialized, deserialized);
     }
 
     /**
@@ -117,15 +120,15 @@ protected String serializeResult(T result) {
     @SuppressWarnings("ThrowableNotThrown")
     protected ErrorObject serializeException(Throwable throwable) {
         var error = ExceptionHelper.buildErrorObject(throwable, resultSerDes);
-        if (shouldValidateSerializationRoundTrip()) {
+        if (shouldDeserializeAfterSerialization()) {
             deserializeException(error);
         }
         return error;
     }
 
-    private boolean shouldValidateSerializationRoundTrip() {
+    private boolean shouldDeserializeAfterSerialization() {
         var config = getContext().getDurableConfig();
-        return config == null || config.shouldValidateSerializationRoundTrip();
+        return config == null || config.shouldDeserializeAfterSerialization();
     }
 
     /**

sdk/src/main/java/software/amazon/lambda/durable/operation/StepOperation.java

@@ -144,9 +144,11 @@ private void checkpointStarted() {
     }
 
     private void handleStepSucceeded(T result) {
+        var serializedResult = serializeAndDeserializeResult(result);
+
         // Send SUCCEED
         var successUpdate =
-                OperationUpdate.builder().action(OperationAction.SUCCEED).payload(serializeResult(result));
+                OperationUpdate.builder().action(OperationAction.SUCCEED).payload(serializedResult.serialized());
 
         // sendOperationUpdate must be synchronous here. When waiting for the return of this call,
         // the context threads waiting for the result of this step operation will be wakened up and registered.

sdk/src/main/java/software/amazon/lambda/durable/operation/WaitForConditionOperation.java

@@ -128,15 +128,15 @@ private void executeCheckLogic(T currentState, int attempt) {
                     // Execute check function in user executor
                     WaitForConditionResult<T> result = checkFunc.apply(currentState, stepContext);
 
-                    // Serialize/deserialize round-trip on the value to ensure state is checkpoint-safe
-                    var serializedState = serializeResult(result.value());
-                    T deserializedValue = deserializeResult(serializedState);
+                    // Normalize the value through SerDes so first execution matches replay.
+                    var serializedState = serializeAndDeserializeResult(result.value());
+                    T deserializedValue = serializedState.deserialized();
 
                     if (result.isDone()) {
                         // Condition met — checkpoint SUCCEED
                         var successUpdate = OperationUpdate.builder()
                                 .action(OperationAction.SUCCEED)
-                                .payload(serializedState);
+                                .payload(serializedState.serialized());
                         sendOperationUpdate(successUpdate);
                     } else {
                         // Compute delay from strategy
@@ -145,7 +145,7 @@ private void executeCheckLogic(T currentState, int attempt) {
                         // Checkpoint RETRY with delay
                         var retryUpdate = OperationUpdate.builder()
                                 .action(OperationAction.RETRY)
-                                .payload(serializedState)
+                                .payload(serializedState.serialized())
                                 .stepOptions(StepOptions.builder()
                                         .nextAttemptDelaySeconds(Math.toIntExact(delay.toSeconds()))
                                         .build());