aws / aws/aws-durable-execution-sdk-java

[Bug]: waitForCondition exhaustion can produce FAILED invocation without ErrorObject

オープン
#634 コメント 0 件 リアクション 0 件 担当者 0 名 GitHub で見る
bug java pkg:sdk
主要言語
Java
スター
28
フォーク
11
平均マージ
1日 8時間
マージ済み PR(30日)
47

説明

### Expected Behavior

When `waitForCondition` exhausts its configured attempts, the SDK should:

1. Checkpoint the `WaitForCondition` step as `FAILED` with a non-null `ErrorObject`.
2. Checkpoint an enclosing synchronous child context as `FAILED` with a non-null `ErrorObject`.
3. Return a root `DurableExecutionInvocationOutput` with `Status=FAILED` and a non-null `ErrorObject`.

This ensures replay preserves useful diagnostics and any service/emulator consuming the invocation result can reliably record the execution as failed.

### Actual Behavior

The built-in wait strategies throw `WaitForConditionFailedException(String)`, whose `DurableOperationException` state has a null `ErrorObject`.

That null propagates through the failure path:

- `WaitForConditionOperation.handleCheckFailure()` reuses `DurableOperationException.getErrorObject()` without a null fallback.
- `ChildContextOperation.handleChildContextFailure()` does the same.
- `DurableExecutor.buildErrorObject()` returns `DurableOperationException.getErrorObject()` directly, even when it is null.

The final invocation output can therefore be equivalent to:

```json
{
"Status": "FAILED"
}
```

The Java `LocalDurableTestRunner` preserves the `FAILED` status, but the error is absent. When used through SAM Local, this combines with an emulator bug that currently converts `FAILED` with no error into `ExecutionSucceeded`.

### Steps to Reproduce

Use a synchronous child context containing a condition that never stops polling:

```java
var runner = LocalDurableTestRunner.create(String.class, (input, ctx) ->
ctx.runInChildContext("child", String.class, child ->
child.waitForCondition(
"poll",
String.class,
(state, stepCtx) -> WaitForConditionResult.continuePolling(state),
WaitForConditionConfig.builder()
.waitStrategy(WaitStrategies.fixedDelay(
2, Duration.ofSeconds(1)))
.build())));

var result = runner.runUntilComplete("test");
```

The local runner reports `FAILED`, but the root error is absent. Running the equivalent handler through SAM Local can produce this history:

```text
StepFailed (WaitForCondition)
ContextFailed (RunInChildContext)
InvocationCompleted
ExecutionSucceeded
```

### SDK Version

`2.1.0`; the behavior is also present on current `2.1.1-SNAPSHOT` / `main` as of 2026-08-19.

### Java Version

17

### Is this a regression?

No known working version.

### Last Working Version

N/A

### Proposed Fix

Always fall back to serializing the thrown exception when an SDK exception has no embedded error:

```java
if (e instanceof DurableOperationException operationException
&& operationException.getErrorObject() != null) {
return operationException.getErrorObject();
}
return ExceptionHelper.buildErrorObject(e, serDes);
```

Apply equivalent null fallbacks when checkpointing failures in `WaitForConditionOperation` and `ChildContextOperation`, so operation history and replay also retain the error.

Add an integration test for max-attempt exhaustion inside synchronous `runInChildContext` that asserts:

- the step is failed with error details;
- the child context is failed with error details;
- the root invocation is `FAILED` with an error.

### Related Issue

The SAM Local emulator status-handling defect exposed by this null error is tracked in aws/aws-durable-execution-sdk-python#656.

コントリビューションガイド

コントリビューションガイドを開く

調査の方向性

WaitForConditionOperation.handleCheckFailure()、ChildContextOperation.handleChildContextFailure()、DurableExecutor.buildErrorObject() から着手し、その後 LocalDurableTestRunner で runInChildContext の同期的な再現を実行します。null エラー時のフォールバックと、最大試行回数の使い果たしを対象とする統合テストを追加します。step、child context、root invocation がすべてエラー詳細付きで FAILED になれば完了です。

索引モデルが issue の本文から書いたものです。

評価

技術スタック
java
領域
backend, testing
issue の種類
バグ
難易度
4/5
見積もり時間
3〜5日
活発さ
活発
明瞭さ
明確に書かれている
初心者へのやさしさ
68/100

新しい issue をメールで受け取る

初心者向けの GitHub issue を短くまとめたダイジェスト。