Proceedings of the International scientific and practical conference ―Modern Science: Challenges and Perspectives‖ (February 9-11, 2026) / Publisher website: www.naukainfo.com. - London, United Kingdom, 2026. - 121 p.

52 Table 2. Engineering challenges of deploying recommendation algorithms in e-commerce Engineering challenge Nature of the problem Typical engineering solutions Integration into distributed architectures The need for interaction between recommendation modules and microservices, product catalogues, and payment systems API-oriented integration, containerisation, Kubernetes, service mesh Streaming data processing Continuous streams of user behavioural events require real-time processing Apache Kafka, Spark Streaming, Flink, real-time feature stores Low-latency inference Strict requirements for recommendation response time (<100 ms) Caching, approximate nearest neighbour search, GPU/TPU inference Data privacy and regulatory compliance Processing of personal data is subject to GDPR, CCPA, and national regulations Data anonymisation, differential privacy, federated learning Model lifecycle management (MLOps) Model degradation due to data drift and the need for retraining and quality control CI/CD for machine learning, monitoring, automated retraining, A/B testing Source: compiled based on [1; 5] The presented comparative analysis indicates that the engineering challenges of deploying recommendation algorithms extend beyond purely algorithmic optimisation and encompass architectural, infrastructural, and regulatory aspects. The integration of recommender systems into distributed microservices architectures requires advanced cloud infrastructure and orchestration of computational resources. Streaming data processing and stringent low-latency inference requirements necessitate the use of specialised stream analytics technologies and hardware- accelerated computation. Compliance with data privacy requirements imposes additional constraints on model architectures, thereby stimulating the adoption of privacy-preserving and federated learning methods. Maintaining the lifecycle of recommender models requires the implementation of MLOps approaches that support