Introduction
As web services grow in complexity and interconnectivity, diagnosing runtime failures becomes increasingly difficult. Traditional Model-Based Diagnosis (MBD) relies on precise formal logic but often lacks the flexibility to handle missing system knowledge. Conversely, statistical methods leverage historical data but frequently miss "semantic faults" that violate business logic despite appearing statistically valid. This paper proposes a hybrid approach that combines formal Business Process Execution Language (BPEL) descriptions with historical execution data to improve diagnostic accuracy.
Methodology
The proposed framework utilizes three probabilistic sub-models: the Behavior Transition Model (BTM), the Message Transition Model (MTM), and the Behavior-Message Emission Model (BMM). By mapping BPEL processes to these models and refining them with successful historical traces, the system balances formal logic with real-world experience using a weighting scheme ($wt_1, wt_2$).
When a failure occurs, the system executes a diagnosis algorithm to identify the root cause. It first checks message transition probabilities against a threshold ($\epsilon$). Then, it employs the Viterbi algorithm to determine the most likely "correct" behavior sequence. By comparing the actual failed trace to this inferred ideal path, the system localizes discrepancies as faults.
Experimental Setup
The researchers evaluated the hybrid model using a simulated environment with synthetic web service compositions ranging from 20 to 100 nodes. Real-world Quality of Service (QoS) attributes from the QWS database were used to simulate realistic fault conditions, allowing for a robust assessment of the model's performance in distributed environments.
Results and Discussion
The "Diagnosis Model" (DM) was compared against a baseline established by Yan’s method. The results demonstrate that the hybrid approach consistently achieves higher accuracy across all scales of service compositions.
Key insights from the evaluation include:
- Holistic Trace Analysis: By considering the entire execution trace rather than isolated steps, the model effectively identifies control-flow faults.
- Robustness to Noise: The use of a probability threshold ($\epsilon$) allows the model to ignore non-critical variations, focusing only on significant failures.
- Scalability: The performance advantage of the DM remains stable as the number of service nodes increases, indicating its suitability for large-scale applications.
Conclusion
The hybrid model-based approach offers a significant improvement in diagnosing web service composition faults by bridging the gap between formal models and historical data. Future work will focus on optimizing the weighting of these data sources and validating the model in live, heterogeneous production environments to assess real-time computational overhead.
Summary updated