Where to adapt dynamic service compositions

Bo Jiang, Professor, Bo Jiang, W. K. Chan, Zhenyu Zhang, T. H. TseWWW 2009 · DOI

TL;DR

This paper introduces a novel technique to automatically identify "cracks" – specific service invocations within a dynamic service composition that are most impactful for adaptation. By continuously monitoring service Quality of Service (QoS) and analyzing historical performance, their method pinpoints these critical junctures, enabling more efficient and effective maintenance of complex web service workflows.

What Problem Does This Paper Address?

Modern business processes are increasingly built by composing various web services. These services, and the way they are composed, can change over time. When a service composition fails to meet its Quality of Service (QoS) requirements (like response time or reliability), it needs to be adapted. However, large and complex compositions make it difficult to pinpoint which specific service invocation is causing the problem and would yield the biggest improvement if adapted. Simply replacing a problematic service might not be enough if the underlying issue lies in how it's integrated or if other services are also contributing to the poor performance. This paper addresses the challenge of efficiently identifying these critical adaptation points, or "cracks," within dynamic service compositions.

Methods

The core of the proposed technique involves two main steps:

  1. Service Monitor and QoS Collection: The system intercepts service invocations within a composite service at the middleware tier. For each invocation, it records the sequence of services called and collects a vector of QoS metrics (e.g., response time, success rate) for the invoker service. This creates a historical log of service interactions and their associated QoS.
  2. Detection of Cracks: Using the collected QoS data, the technique calculates a "crack-sensitivity" score for each service. This score quantifies how consistently a service receives higher (or lower, depending on normalization) QoS values compared to others in the composition. Statistical methods, including Analysis of Variance (ANOVA) and multiple statistical comparison procedures, are employed to determine if certain services exhibit significantly different QoS performance patterns. Services identified as consistently performing poorly or exceptionally well are flagged as potential cracks. The paper then identifies specific invocation endpoints within these services that are still active in the latest composition as the actual adaptation points.

Datasets

Not discussed in the paper. The evaluation uses a simulation experiment based on a real-life insurance application for damage claim service, with 10 instances of each service type.

Key Insights and Ideas

  • The "Crack" Concept: The paper introduces the idea of "cracks" as specific service invocation points that are most critical for adaptation. This moves beyond simply identifying a problematic service to pinpointing the exact interaction that needs attention.
  • QoS Monitoring is Key: Continuously monitoring QoS at the invocation level, rather than just at the composite service level, is crucial for diagnosing issues.
  • Statistical Analysis for Impact: Using statistical methods like ANOVA helps to objectively identify services with significantly different performance characteristics, reducing guesswork in adaptation.
  • Pareto Principle in Service Adaptation: The authors implicitly suggest that focusing adaptation efforts on these identified "cracks" aligns with the Pareto Principle (80-20 rule), where a small number of critical points yield the most significant improvements.

Key Results

The paper's evaluation, using a simulated insurance claim scenario, demonstrated the effectiveness of their crack-detection technique.

  • An ANOVA test confirmed that the crack-sensitivity means of services differed significantly at a 0.1% significance level.
  • A multiple comparison procedure further showed that a specific service instance (G1, a garage service) was significantly more crack-sensitive than all other 49 services.
  • Based on this analysis, the "Assign garage" endpoint within the Europe Assist service was identified as a crack, indicating it as a prime candidate for adaptation. The p-value for the comparison between G1 and another garage service (G2) was extremely small (7.9578 × 10⁻⁶), reinforcing the statistical significance.

Limitations and Future Work

Limitations:

  • The paper acknowledges that the current technique might not be effective in all scenarios and suggests further work on improving the evaluation metrics to reduce false positive reports.
  • The current approach focuses on identifying cracks based on historical QoS data. The effectiveness of adapting at these identified points needs further validation.

Future Work:

  • Developing a more effective and lightweight online metrics evaluation strategy.
  • Exploring different sampling strategies for long-running services to efficiently utilize historical data.
  • Investigating ontology-based service composition (e.g., OWL-S) and testing adaptation patterns.

Summary updated

BibTeX
@inproceedings{professor2009where,
  author = {Bo Jiang, Professor and Bo Jiang and W. K. Chan and Zhenyu Zhang and T. H. Tse},
  title = {{Where to adapt dynamic service compositions}},
  booktitle = {The Web Conference},
  year = {2009},
  doi = {10.1145/1526709.1526888},
}
Figure