ABSTRACT Background Preserving the quality and safety of perishable products requires continuous monitoring and reliable traceability. Although the Internet of Things (IoT) enables real‐time data collection, multi‐organizational supply chains lack a common mechanism for assigning data custody while maintaining transparency, integrity, and performance. Objective This study proposes and evaluates an architecture integrating IoT, edge/fog computing, and hybrid storage—an off‐chain traditional database combined with a permissioned blockchain—to monitor and trace perishable products. Methods A prototype was implemented using IoT devices and simulators, edge and fog components, and hybrid storage. High‐volume sensor data and critical records were stored off‐chain in MongoDB, while their corresponding hashes were stored on‐chain using Hyperledger Fabric. Four controlled experiments assessed insertion response time, the impact of increasing sensors and edge devices, blockchain queue performance under burst workloads, and blockchain storage consumption. The hybrid approach was compared with MongoDB‐only and Hyperledger Fabric‐only storage. Results Hybrid storage achieved insertion up to six times faster than blockchain‐only storage. Response times increased with simultaneous requests and additional edge devices, while asynchronous ordered insertion prevented transaction conflicts during bursts. The prototype achieved 18.5 transactions per second, below the 65 estimated for an illustrative supply‐chain scenario. Blockchain storage grew approximately 8 MB per 100 records, reaching about 1 GB for 12,800 hashes. Conclusion The prototype demonstrates the feasibility of combining off‐chain storage, permissioned blockchain records, and edge/fog processing to provide verifiable traceability while reducing on‐chain load. Larger‐scale, real‐world evaluations and storage‐management strategies remain necessary.
ABSTRACT Background Code smells are an indicator that something is misplaced in software systems, which can reduce maintainability and quality of software applications. In the literature, there is much research about code smells where numerous strategies for automating code smell detection have been developed to improve software quality. Objective The purpose of this work is to provide a review of search‐based, heuristic‐based, machine learning‐based, deep learning‐based, and hybrid‐based code smell detection algorithms. Method Concerning the main goals of this research, we have found 38 primary studies. We gathered relevant studies published on this topic between 2017 and 2024. These articles' data were extracted according to some criteria, including code smells, machine learning methods, programming languages, dataset size, evaluation strategy, and statistical tests. Results Machine learning‐based code smell detection methods have been suggested in most empirical investigations. We found that machine learning and deep learning are the most popular approaches for predicting code smells. Researchers typically employ support vector machine and decision tree algorithms. The Arcelli Fontana and Zanoni benchmark dataset was the most frequently investigated dataset. Most of the research community's attention has been focused on a small number of smells, including blob, feature envy, long method, and data class. Researchers also pay more attention to code smells like Long Method and Feature Envy. Deep learning techniques are increasingly used, and most scientists utilize source code metrics as indicators. The standard performance measurements mostly used are F‐measure, recall, and precision. Conclusions The study provides an overview of existing approaches and highlights current research trends in code smell detection, particularly the increasing use of machine learning and deep learning techniques.
ABSTRACT Background Blockchain, as a decentralized and distributed ledger, shows significant application potential and commercial value in the Internet of Things (IoT). However, the dynamicity of IoT nodes introduces significant challenges to traditional blockchain systems. Specifically, in dynamic environments involving node joins/leaves, mobility, and state changes, traditional solutions often suffer from poor security, significant consensus latency, and severe storage overhead. Although blockchain for IoT has been extensively surveyed, prior reviews fail to provide a comprehensive evaluation because the existing literature considers security, consensus, and scalability only as separate issues, while overlooking node dynamics as a fundamental dimension. Objective This paper aims to provide a systematic investigation of blockchain adaptation to node dynamics in IoT and to establish a comprehensive evaluation perspective for analyzing the impacts of node dynamics on security, consensus, and scalability. Methods To this end, this paper systematically investigates these challenges through three contributions. Firstly, a unified taxonomy of node dynamics is proposed to analyze the impacts on security, consensus, and scalability. Secondly, it synthesizes representative solutions across trust evaluation, consensus algorithms, and scalability techniques. Thirdly, it introduces a three‐dimensional evaluation approach to comprehensively analyze these solutions in terms of performance, security, and adaptability. Results The evaluation results illuminate a clear trade‐off among these three dimensions, demonstrating that improving only one aspect often degrades the overall system balance in dynamic environments. Conclusion Finally, this paper introduces critical research limitations and future research directions to provide practical suggestions for effectively deploying blockchain in highly dynamic IoT environments.
Objective Long-running software systems suffer from aging, characterized by rising performance degradation, resource depletion, and elevated failure rates. Conventional rejuvenation techniques, which often depend on predetermined restart intervals, model the system as simply healthy, deteriorated, or failed. However, practical evidence from a variety of fields indicates that wear and defect buildup frequently accelerate nonlinearly, necessitating more adaptable mathematical explanations.Method Compared to simpler two-parameter or linear models, an extended three-parameter power-law model which consists of a scaling factor, exponent, and offset is employed here to more accurately depict this behavior. The proposed framework enables adaptive rejuvenation policies triggered by observed conditions rather than rigid schedules by continuously monitoring a degradation metric tuned to this model. Renewal theory with rewards can be used to enhance rejuvenation timing and analytically evaluate steady-state unavailability.Results By accurately simulating real-world dynamics, numerical results show that these degradation-aware, threshold-based policies outperform fixed interval approaches by accurately modelling real-world dynamics, particularly in scenarios with increasing aging.Conclusions Together, the extended degradation model and the proposed rejuvenation policies provide a unified analytical framework that improves system availability and reduces long-run operational costs. This study shows that alert-based strategies consistently outperform risk-based policies because they allow for early, data-driven maintenance across various software aging conditions.
Purpose Workflow task scheduling performs a vital role in optimizing performance in cloud-fog computing systems, namely, the combination of centralized cloud servers and decentralized fog nodes to support Internet of Things (IoT) applications. Efficient scheduling guarantees the lowest delay, energy consumption, execution time, and cost. However, owing to the heterogeneous and dynamic characteristics of such environments, task scheduling has been recognized as an NP-hard problem, in which finding the global optimum solution is computationally infeasible.Methodology To mitigate these problems, our proposed solution is the hybridization of the artificial bee colony (ABC) algorithm and the butterfly swarm optimization (BSO) algorithm as the HABSO algorithm for scheduling workflow tasks. The ABC algorithm, which is known for its global search capabilities, is used to explore possible solutions within the wide and complex space of cloud-fog systems, whereas BSO aids in enhancing the local search capabilities and fine-tuning of solutions to ensure rapid convergence and the best possible performance. This hybrid algorithm provides a good balance between scheduling objectives for the task to be finished, a factor that minimizes the total execution time (TET) and energy consumption (EC) of the computational resources.Findings The simulation results indicate the superiority of the proposed HABSO hybrid algorithm over classical state-of-the-art metaheuristic algorithms, with average reductions of 14.35% (ABC), 18.59% (BSO), 14.02% (GWO), 16.26% (WOA), 13.41% (PSO-WOA), and 12.01% (GWO-MSOA) for the TET. Similarly, energy savings of 14.33%, 11.62%, 12.72%, 12.73%, 13.07%, and 8.85% were obtained for the ABC, BSO, GWO, WOA, PSO-WOA, and GWO-MSOA, respectively. Furthermore, sensitivity evaluation, convergence behaviour and statistical evaluations were conducted.Originality The proposed algorithm provides better performance at different workloads and task dependencies; therefore, it has the advantage of improving task scheduling for cloud-fog computing systems.
Quantum computing has entered the noisy intermediate-scale quantum (NISQ) era, where limited qubit numbers, short coherence times, and high error rates pose significant challenges to reliable large-scale execution. Efficient scheduling and resource allocation across heterogeneous quantum hardware are therefore crucial for maximizing system throughput, fidelity, and fairness. In this work, we propose a hardware-aware reinforcement learning framework for quantum task scheduling in distributed NISQ systems. Our design explicitly models qubit-level variability, including connectivity degree, coherence times, error rates, and throughput, while integrating task-level constraints such as deadlines, priorities, and concurrency requirements. We formulate the scheduling problem as a constrained optimization task and instantiate it as a Markov decision process (MDP), enabling reinforcement learning agents to learn adaptive strategies. Specifically, we implement deep Q-network (DQN) and proximal policy optimization (PPO) agents, and compare them against heuristic and random baselines. Experimental results demonstrate that PPO consistently outperforms DQN and heuristic methods, achieving higher task completion rates, fewer deadline violations, and more robust adaptation across different reward configurations. This work bridges quantum hardware modeling with reinforcement learning-based scheduling, providing a practical pathway for resource optimization in distributed quantum computing environments.
Purpose In the fields of requirements engineering and problem-oriented requirements analysis, successful conflict identification and resolution are essential to the development and deployment of software systems. The purpose of this paper is to identify conflicting requirements using the Enhanced Pairwise Comparison Method (E-PCM) applied to a real-world mobile phone case study within the Jackson problem frame framework.Method This paper enhances the traditional PCM by incorporating a full Natural Language Processing (NLP) pipeline comprising entity extraction, parsing rules, and conflict identification using the spaCy library. The methodology further introduces a modified contribution table to represent identified conflicts, and employs empirical evaluation with a structured conflict scaling system (0-5) in place of fuzzy logic for attribute weighting.Results The results indicate potential conflicts between requirements, categorized as no, minimal, moderate, substantial, high, and extreme conflict, each assigned a corresponding level from 0 to 5. The E-PCM model achieves a precision of 1.0 (100%) across all nine requirements, with recall ranging from 0.5 to 1.0 and an F1-Score ranging from 0.66 to 1.0. The overall accuracy of E-PCM is 87%, outperforming the cluster-based approach (85.71%). Inter-rater agreement among 78 undergraduate participants was substantial (k = 0.769), and Fisher's Exact Test validated the NLP-based conflict predictions (p <= 0.05 for 13 out of 15 sub-RQs). These findings provide stakeholders with actionable insights to inform decision-making and requirement prioritization.Conclusion This paper addresses a significant gap in requirements engineering by introducing the first NLP-augmented pairwise comparison method natively aligned with the Jackson problem frame. Unlike existing approaches, E-PCM combines entity extraction, semantic parsing rules, and empirical conflict scaling to provide a domain-independent, scalable, and data-driven conflict identification framework acknowledged both in the literature and by practitioners.
Introduction Defining an experimental configuration is a complex decision problem for early-stage researchers, who must map goals, constraints, and requirements onto datasets, algorithms, and parameter settings that directly affect experimental outcomes. Existing scientific workflow engines improve execution and reproducibility; however, they rarely capture the decision rationale behind configuration choices, which is needed to inform future selections.Method We propose a context-aware decision-support framework that formalises experiment configuration as a structured and sequential decision problem. The framework combines three components: a semantic Knowledge Graph (KG) storing historical configurations, contextual attributes, and decision rationale; an MDP-based Option Explorer that filters the KG under user-defined constraints and ranks feasible configurations by expected cumulative reward; and a Graphical User Interface for specifying constraints, inspecting ranked alternatives, and providing structured feedback. Unlike existing workflow systems, the framework explicitly separates user-defined context from automated reasoning, producing an interpretable ranked list rather than a single opaque recommendation. We evaluated the framework in a user study with 90 MSc- and PhD-level researchers performing a model-selection task, using a synthetic dataset of one million experimental configurations under three levels of contextual detail.Results Compared with manual search, the framework reduced decision time (up to 68%), reduced perceived difficulty (up to 36%), and increased user satisfaction (up to 43%) under the constrained condition.Conclusion By formalising the link between experimental context and probabilistic decision ranking, the framework improves reproducibility and scalability of decision support in scientific experimentation.
Background The integration of Artificial Intelligence of Things (AIoT) has significantly advanced anomaly detection and predictive maintenance in industrial systems.Objective This study proposes an AIoT-based framework for firefighting pump anomaly analysis using an Improved Particle Swarm Optimization (IPSO) and hybrid Bidirectional Long Short-Term Memory-Convolutional Neural Network (BiLSTM-CNN) model.Method IPSO is employed to optimize hyperparameters, enhancing the model's accuracy and convergence efficiency in detecting diverse failure modes. To evaluate robustness, the framework is tested under multi-frequency sampling schemes with 30 s, 1 min, 5 min, and 10 min, analyzing the trade-off between data granularity and computational cost. The proposed system is further validated through deployment in a real pump manufacturing company in Taiwan, demonstrating its practical applicability in industrial environments. The main contributions include: (i) an IPSO-optimized BiLSTM-CNN model for improved anomaly detection, (ii) a systematic multi-frequency evaluation framework, and (iii) a real-world AIoT deployment for predictive maintenance.Results Experimental results show that the proposed approach outperforms conventional models (RNN, CNN, GRU, LSTM, BiRNN, BiLSTM, and CNN-BiLSTM), achieving up to 95.32% improvement in validation loss, 97.29% reduction in MSE, and over 6% gains in prediction and recall indices.Conclusions These results demonstrate the robustness and effectiveness of the proposed framework for reliable firefighting pump maintenance.
Purpose Accurate software defect prediction (SDP) is critical to the success of any software project. Earlier studies have largely used static, semantic or structural features either in isolation or in pairs, offering a partial view of the source code. In reality, static features depict the statistical characteristics, semantic features depict the context and structural features depict data and control dependencies of the code. We propose a strong SDP model that integrates three types of features, achieving a holistic view of the source code, ultimately thereby enabling more robust and generalizable predictions.Methods First, the model extracts the static features from the open source PROMISE repository, semantic features from the Abstract Syntax Tree via CodeBERT followed by BiGRU and structural features from the Program Dependency Graph via Graph Convolutional Network. Second, feature alignment is performed for the fixed-set representation of the three types of features using global attention pooling. Third, feature fusion is done for joint feature representation, followed by the application of additive attention to select the most suitable features. To handle the class imbalance scenario, cost-sensitive gradient boosting is applied to penalize misclassifications more heavily. At last, the final feature set is fed to a classifier for defect prediction.Results Experiments conducted on eleven open source datasets reveal that the proposed unified feature representation approach achieves substantial performance improvements over the state-of-the-art models. Moreover, the Wilcoxon signed-rank test offers statistical validation for the relevance of these enhancements.Conclusion The integration of static, semantic and structural information results in a more holistic representation of source code, which substantially enhances defect prediction performance. The proposed approach addresses the partial view constraints of earlier approaches and offers strong potential for establishing a more reliable and robust SDP across various application domains.
Traditional DevOps pipelines often struggle with scalability, adaptability, and intelligence, particularly in distributed microservices and hybrid cloud environments, where reactive monitoring, static resource allocation, and manual interventions contribute to frequent failures, longer recovery times, and inefficient resource utilization. This study proposes an adaptive ML-Driven DevOps (ML-DevOps) framework designed to transform reactive CI/CD pipelines into proactive, self-optimizing systems through predictive analytics, anomaly detection, reinforcement learning, and intelligent resource optimization. The framework integrates five core components: a predictive analytics engine, hybrid anomaly detection system, reinforcement learning agent, resource optimizer, and orchestration layer for compatibility with mainstream DevOps tools. A 21-month evaluation was conducted across 25 organizations representing five industries, with over 78,000 deployment events analyzed. The framework demonstrated substantial improvements in deployment reliability, recovery efficiency, and resource management, consistently reducing failures, accelerating recovery, and optimizing infrastructure use across diverse industries and organizational scales. By embedding machine learning intelligence throughout the software delivery lifecycle, the ML-DevOps framework advances DevOps from reactive automation to intelligent, autonomous operation. Its modular, plug-and-play design ensures practical integration into existing toolchains, making it a scalable and domain-agnostic solution. Future work will explore explainability, federated learning, and lightweight edge deployment to enhance transparency and adaptability.
Introduction Requirements engineering plays a crucial role in the software development lifecycle, encompassing the elicitation, analysis, specification, and validation of requirements. Inefficiencies in any of these processes can lead to delays, budget overruns, and even project failure. This paper explores the integration of requirements reuse and recommender systems to enhance the elicitation process by leveraging historical project data and stakeholder interaction patterns.Methods The proposed methodology incorporates a hybrid approach that combines collaborative filtering and content-based filtering to recommend relevant requirements to stakeholders. A dynamic weighting framework adjusts the contributions of these two approaches based on the availability of data. In situations with insufficient qualified data, the approach relies more heavily on content-based filtering to address challenges such as data sparsity and the cold-start problem. To enhance the semantic similarity between requirements, the method aggregates GloVe word vectors with domain-specific TF-IDF scores to identify software engineering-specific vocabulary.Results Experimental evaluation using a benchmark dataset demonstrates that the proposed hybrid approach significantly improves the prediction accuracy of relevant requirements recommendations, compared to traditional methods.Conclusion The integration of requirements reuse with a recommender system that combines collaborative and content-based filtering offers an effective solution to streamline the elicitation process, mitigate risks of overlooking critical requirements, and save time during the evaluation and selection of requirements. The proposed method improves the efficiency and accuracy of requirements engineering, especially in contexts with limited data availability.
Objective To design a secure Federated Learning (FL) framework for Internet of Medical Things (IoMT) that protects sensitive patient data from both classical and quantum attacks.Methods Proposed the QSFedMA-IoMT protocol integrating quantum and classical security techniques. Utilized entanglement-based E91 protocol for generating a highly secure root key to establish trust. Applied BB84 protocol for efficient generation of per-round session keys during FL updates. Incorporated classical cryptographic scheme AES-GCM for secure communication. Employed privacy-enhancing techniques such as norm-clipping and Gaussian noise to mitigate information leakage during model training.Results Our work demonstrates robust resistance against both classical and quantum adversaries, while enhancing data privacy through secure key distribution and differential privacy mechanisms. It ensures the integrity of model updates within the federated learning process and achieves an effective balance between strong security guarantees and computational efficiency, making it well-suited for IoMT environments.Conclusion The QSFedMA-IoMT protocol delivers a robust and practical hybrid framework for securing federated learning in healthcare systems. By integrating E91 and BB84 protocols, it strengthens key management and trust establishment. The combination of quantum security with classical privacy-preserving techniques ensures resilience, scalability, and efficiency. Overall, this work provides a promising direction for secure and privacy-aware federated learning in next-generation IoMT applications.
Background Topology and Orchestration Specification for Cloud Applications (TOSCA) is a specification language used for modelling topology and orchestration of cloud applications. This language particularly allows the description of workflows that can be used for specifying management tasks such as (un)deployment plans.Motivations This textual language for describing TOSCA workflows does not provide any visual notation for graphically designing or observing the corresponding workflows. Moreover, this specification language is error-prone and can be source of mistakes during the writing of the management plans.Methods In this article, we propose a transformation from TOSCA to the graphical Business Process Model and Notation (BPMN), which allows the visualisation of TOSCA workflows. We also provide automated verification techniques for analysing TOSCA (un)deployment workflows in terms of functional and architectural properties as well as execution times. The transformation and verification steps are achieved in a fully automated way.Results This approach computes BPMN models and verification results within a reasonable time on realistic applications.
Background DevOps pipelines have become the primary vehicle for operationalizing software architecture decisions; however, their design and evolution remain largely ad hoc and tool-specific. This disconnect weakens traceability from architectural intent to runtime automation, complicates change impact analysis, and increases the risk of configuration errors. Although model-driven engineering (MDE) has been proposed to support CI/CD adoption, existing approaches typically focus on individual tools or isolated pipeline fragments and lack a unified, reusable foundation for systematic transformation.Aims This paper aims to introduce a unified DevOps Pipeline Meta-Model (DP2M) and an architecture-to-pipeline transformation framework that enables the derivation of executable DevOps pipelines directly from software architecture models, while ensuring traceability and supporting systematic reuse.Materials and Methods A mixed-methods approach is employed, combining: (i) a systematic mapping of MDE-for-DevOps literature; (ii) a cross-vendor analysis of industrial pipeline specification languages across Jenkins (Declarative and Scripted), GitHub Actions, GitLab CI, Azure Pipelines, CircleCI, Travis CI, Google Cloud Build, and AWS CodePipeline; and (iii) semi-structured interviews with practitioners. From this, a taxonomy of pipeline artifacts and concerns-covering build, test, deployment, security, compliance, and observability-is derived, along with quality-driven requirements for pipeline modeling. These are formalized into the DP2M meta-model and a catalog of reusable transformation patterns with defined rules and constraints.Results The proposed DP2M captures a technology-agnostic representation of DevOps pipelines with explicit traceability links to architectural elements and decisions. A prototype toolchain implements the framework and generates executable pipelines across multiple CI/CD platforms. Evaluation through realistic case studies demonstrates expressiveness across heterogeneous toolchains, preservation of architectural intent, reduction of duplication, and improved handling of DevSecOps concerns as first-class modeling constructs.Discussions The findings highlight the limitations of existing tool-centric approaches and demonstrate how a unified meta-model combined with formal transformation patterns can bridge the gap between architecture and pipeline implementation. The approach supports traceability, facilitates change impact analysis, and enables controlled co-evolution of architecture and pipeline models across diverse environments.Conclusions This work presents a practical and scalable path toward architecture-centric, model-driven DevOps pipelines. By enabling analyzable, evolvable, and reusable pipelines across projects and platforms, the proposed framework advances the integration of software architecture and DevOps practices while addressing key challenges in traceability, consistency, and automation.
Objective Converting binary integers to variable-length decimal strings is a fundamental operation in computing. Conventional fast approaches rely on recursive division and small lookup tables. The goal of this work is to develop a significantly faster method for this task.Methods We propose a SIMD-based algorithm that leverages integer multiply-add instructions available on recent AMD and Intel processors. Our method eliminates lookup tables entirely and computes multiple quotients and remainders in parallel. Additionally, we introduce a dual-variant design with dynamic selection that adapts to input characteristics: a branch-heavy variant optimized for homogeneous digit-length distributions and a branch-light variant for heterogeneous datasets.Results Our single-core algorithm consistently outperforms all competing methods across the full range of integer sizes. It runs 1.4-2 & times; faster than the closest competitor and 2-4 & times; faster than the C++ standard library function 'std::to_chars' across tested workloads.Conclusion The proposed SIMD-based approach with dual-variant dynamic selection provides a substantial performance improvement for integer-to-decimal conversion, delivering superior speed without relying on traditional lookup tables.
Background The rapid adoption of low-code (LC) and no-code (NC) tools has transformed the software development landscape, offering faster development cycles and enabling non-programmers to contribute to software creation.Objective This study systematically maps the effects of LC/NC tools on developers, addressing the critical need to understand whether these tools serve as a valuable aid or a potential threat to their roles.Methods By analyzing how LC/NC tools support development processes, the challenges they present, and their potential to replace traditional developers, this research fills a key knowledge gap by performing a systematic mapping study.Results The findings reveal that while LC/NC tools simplify certain tasks, they do not replace the need for professional developers to handle complex technical challenges, change management, and system integration.Conclusions Rather than posing a threat, these tools complement and enhance developers' expertise, empowering them to focus on more advanced aspects of software creation. This study provides valuable insights for the professional developer community and the broader software industry, advocating for the thoughtful integration of LC/NC tools as powerful aids rather than competitors to professional development.
Introduction Open-source software (OSS) projects thrive on collaborative communities, forming a critical backbone of modern technology. Despite community contributions shaping open-source development, a gap remains in understanding how community interactions impact project longevity. To address this, our study investigates how discussion volume (as a measure of engagement), sentiment expressed in discussions, issue resolution time, and issue characteristics influence both the resolution process and, crucially, OSS project lifespan.Methods Using a comprehensive GitHub issue dataset, we applied robust statistical analyses and effect size measures across various project lifecycle stages.Results Our findings reveal that while discussion volume weakly correlates with resolution time and sentiment's practical impact is negligible, a counter-intuitive pattern emerges: longer-lived projects consistently exhibit extended median issue resolution times, in contrast to faster initial resolutions observed in shorter-lived projects. Analysis of issue labels suggests this is because shorter-lived projects tend to address beginner-friendly issues, whereas enduring projects confront more complex, core development tasks. These findings emphasize that OSS project longevity depends less on rapid initial issue resolution and more on structured, sustained engagement with increasingly complex contributions-underscoring the vital role of qualitative communication and the nature of work in fostering robust communities.
Background Quantum support vector machines and quantum kernel methods have emerged as promising approaches within quantum machine learning, with the goal of leveraging quantum computing to enhance classification performance and computational efficiency. This review systematically surveys recent advances in QSVM and kernel-based quantum classifiers, and analyzes their algorithmic frameworks, experimental implementations, and practical challenges.Methods We systematically examine QSVM approaches, including three schemes based on HHL algorithm, Hadamard test, and variational optimization, alongside multi-class extension and quantum feature mapping.Result Findings indicate that QSVM can offer theoretical speed-ups in specific settings, particularly when combined with quantum feature mappings that encode data into high-dimensional Hilbert spaces. However, current implementations remain constrained by hardware limitations, and a lack of large-scale general validation. Most studies focus on proof-of-principle experiments with limited real-world applicability.Conclusion While promising, QSVM requires further work on scalability, noise resilience, and real-world integration. Future research should focus on robust algorithms and empirical studies.
Introduction As machine learning (ML)-enabled systems become increasingly prevalent across industries, the engineering challenges of deploying and maintaining them in production have emerged as critical. The existing engineering knowledge base often derives from conceptual frameworks or case studies conducted by large technology companies, leaving a gap in the empirical nuances of deployment practices across a wide range of engineering contexts.Objective This study aims to investigate real-world ML deployment workflows and inference architectures to identify common patterns, contextual variations, and underlying trade-offs that influence how practitioners operationalize ML models in production.Methods We conducted a multi-case study of eight ML systems across sectors, including advertising, finance, healthcare, manufacturing, and software platforms. We collected data through semi-structured interviews with practitioners and the various inference architectures. We applied thematic analysis to extract recurring patterns across the deployment lifecycle, from model versioning to inference serving.Results Our findings reveal five core deployment themes: model versioning and storage, quality assurance, monitoring, model packaging, and inference serving. We identify a maturity gradient from manual, ad hoc practices to automated CI/CD pipelines and further highlight architectural trade-offs between tightly coupled vs loosely coupled model deployment patterns. Other key findings include the rarity of model drift monitoring and the use of hybrid inference serving patterns to balance latency and scalability requirements.Conclusions Although ML domains and deployment architectures vary, we observed recurrent practices and strategic trade-offs that provide an empirical foundation for developing more standardized, context-aware ML engineering processes. Our findings provide actionable guidance for practitioners and identify under-explored areas for future research.
