
Defense of the dissertation of Kussepova Lazzat for the degree of Doctor of Philosophy (PhD) in the educational program «8D06102 - Computer Science»
L.N. Gumilyov Eurasian National University, a dissertation defense for the degree of Doctor of Philosophy (PhD) by Kussepova Lazzat on the topic «Models and algorithms for efficient container application development and deployment» to the educational program «8D06102 – Computer Science».
The dissertation was carried out at the «Department of Artificial Intelligence Technologies» of L.N. Gumilyov Eurasian National University.
The language of defense is kazakh
Official reviewers:
Nurbolat Tasbolatuly – Doctor of Philosophy PhD, Associate Professor, Director of the Department of Science at Astana International University (Astana, Republic of Kazakhstan);
Zhanar Oralbekova – Doctor of Philosophy (PhD), Associate Professor of the School of Artificial Intelligence and Data Science, Astana IT University (Astana, Republic of Kazakhstan).
Temporary members of the Dissertation Committee:
Zhanibek Kozhirbayev – Doctor of Philosophy (PhD), Associate Professor, Leading Researcher at the Laboratory of Computational Materials Science, Center for Energy and Advanced Materials Science, National Laboratory Astana, Nazarbayev University (Astana, Republic of Kazakhstan).
Aigul Adamova – Doctor of Philosophy (PhD), Associate Professor, Dean of the School of Software Engineering at Astana IT University (Astana, Republic of Kazakhstan).
Umidjon Djumayozov – Doctor of Technical Sciences (DSc), Associate Professor, and Professor in the Department of Software Engineering at Samarkand State University named after Sharof Rashidov (Samarkand, Republic of Uzbekistan).
Scientific advisors:
Ademi Ospanova – Candidate of Physical and Mathematical Sciences, Associate Professor of the Department of Information Security, L.N. Gumilyov Eurasian National University (Astana, Republic of Kazakhstan);
Fariza Tebueva – Doctor of Physical and Mathematical Sciences, Associate Professor, and Professor in the Department of Computational Mathematics and Cybernetics at the North Caucasus Federal University (Stavropol, Russia).
The defense will take place on December 29, 2026, at 12:00 PM in the Dissertation Council in the field of study “8D061 — Information and Communication Technologies” (“8D06103 — Information Systems”, “6D070300 — Information Systems”) at L.N. Gumilyov Eurasian National University.
Link: https://teams.microsoft.com/meet/454382166261453?p=xDABUBs0czxymZAQez
Address: Astana, street Pushkina, 11, Educational building No. 2, room No. 222.
Abstract (English): the dissertation work by Kussepova Lazzat Educational programme “8D06102 - Computer Science” Submitted for the degree of Doctor of Philosophy (PhD) on the topic “Models and algorithms for efficient container application development and deployment” submitted for the degree of Doctor of Philosophy (PhD) under the educational program «8D06102-Computer Science» ABSTRACT Relevance of the research. Containerization and orchestration technologies have become the primary approach to deploying modern cloud and local applications. However, as the scale of a cluster increases, decisions on deployment, scaling, and load balancing become more complex. Existing orchestration systems make these decisions mainly through static rules and reactive thresholds. This approach becomes ineffective when telemetry is incomplete and noisy, workloads are non-stationary, and execution platforms are heterogeneous. Reactive scaling leads to a decline in service quality between the increase in workload and the time a new replica becomes ready; cold-start latency makes a substantial contribution in serverless environments; when scaling, placement, and balancing control loops operate without coordination, improvements in individual algorithms do not translate into improvements in overall system quality. Therefore, developing models and algorithms that explicitly account for telemetry uncertainty, adapt to workload changes, and integrate control loops for coordinated operation is a relevant scientific task. This task must be considered across the entire life cycle of a containerized application — from configuration creation to deployment and runtime management. Research aim ‒ To develop models and algorithms for the creation, deployment, and adaptive management of containerized applications in cloud services and local infrastructures, and to experimentally substantiate the conditions for their application. In accordance with the stated aims, the following objectives must be accomplished in the dissertation: 1. To analyse methods for the creation, deployment, and adaptive management of containerized applications, and identify their limitations and criteria for evaluating effectiveness; 2. To develop a method for the automatic generation, verification, and correction of container configurations at the creation stage, and comparatively evaluate image build strategies; 3. To develop an algorithm for the adaptive selection of an execution environment at the deployment stage, taking workload characteristics and resource requirements into account, and evaluate its effectiveness; 4. To develop models and algorithms for adaptive management at the execution stage that account for incomplete telemetry and changing workloads, and experimentally substantiate the effectiveness and conditions of application of forecasting, scaling, cold-start reduction, and multi-agent management approaches. Object of the research – The processes of creating, deploying, and managing containerized applications at runtime in cloud and local infrastructure. Subject of the research – Models, algorithms, and software tools for the adaptive management of containerized applications. Research methods. The study employed the theory of Markov decision processes (including the framework of partially observable processes), reinforcement learning and multi-agent learning methods, contextual bandit theory, deep learning and time-series forecasting methods, formal verification, and SMT solvers. Mathematical statistics methods were used to process the results: analysis of variance, parametric and non-parametric tests, corrections for multiple comparisons, and effect size estimation. Experiments were conducted through simulation modelling and as computational experiments using real production traces. Software tools. The development of algorithms and software tools was supported by the use of Python and its libraries. Scientific novelty. The scientific novelty of the work is as follows: the resource management problem in a container environment was formalized as a partially observable Markov decision process; telemetry noise is explicitly accounted for through an observation model, while the control decision relies on a belief state rather than the actual state of the system, thereby improving management quality under noisy measurements; a workload forecasting architecture incorporating an attention mechanism into a bidirectional LSTM network was compared, and a proactive scaling algorithm based on its forecast and trend was considered; an algorithm for reducing cold starts in a serverless environment was proposed: it learns to select the cache size according to the workload structure and uses less memory than static caching policies; the cluster management problem was formulated as a multi-agent system divided into functional roles; the trade-off between service quality and cost is regulated through Lagrangian relaxation within a constrained Markov process; the platform meta-selection problem was formulated as a contextual bandit problem with drift, and an algorithm with exponential forgetting of old statistics was proposed; this algorithm adapts faster when the workload profile changes; a method for verifying container configurations was proposed: it combines the results of static, semantic, and machine-learning-based analyses into a single numerical score and is integrated into a closed loop with language-model-based generation. Main results submitted for defence: 1. Formalizing the resource management problem as a partially observable Markov decision process and solving it on the basis of a belief state substantially reduces service-level violations compared with a threshold policy when telemetry is noisy. 2. A bidirectional recurrent network with an Attention mechanism provides a statistically significant advantage over baseline architectures in forecasting cloud workloads and reducing large forecasting errors. 3. Adjusting the cache size through reinforcement learning makes it possible to find a trade-off point between the proportion of cold starts and memory consumption in a serverless environment that static policies cannot achieve. 4. Managing a cluster through a multi-agent system with role separation reduces service-level violations, while the associated increase in cost is controlled through a budget constraint within a constrained Markov process. 5. A contextual bandit algorithm that exponentially reduces the weight of old statistics (EF LinUCB) selects a platform more effectively than conventional bandit strategies and static rules when the workload profile changes. 6. Integrating three-level configuration verification with language-model-based generation in a closed loop increases the proportion of semantic errors detected before deployment. Practical significance of the research: The developed models and algorithms can be applied in container cluster management systems, CI/CD pipelines, and the operation of cloud platforms (Appendix A). Incorporating the verification system and generation agent into the pre-deployment stage enables configuration errors to be detected before they reach the production environment. The scaling and caching algorithms are intended to reduce resource consumption. The results of the work can also be used in teaching cloud technologies, distributed systems, and machine learning. Implementation of the results. An information system implementation certificate dated 13 October 2026 was obtained, confirming the practical application of the results of the work at L.N. Gumilyov Eurasian National University. Approbation of the dissertation results. The main results of the dissertation research were presented at the following international conferences: 1. Облачные технологии в свете долгосрочной перспективы для бизнес-организаций // Фундаментальные проблемы информационной безопасности в условиях цифровой трансформации (FISP-2020): сб. докл. II Всерос. науч. конф. (с приглаш. зарубеж. ученых). – Ставрополь: СКФУ, 2020. – С. 201-206. (30 November 2020). 2. Қазіргі заманғы бұлттық есептеулер стандарттары мен сервистері // International Conference of L.N. Gumilyov Eurasian National University (Astana, 2021 – 12 April). 3. Cloud Services Projects to Support Small and Medium-Sized Businesses and Approaches for their Commercialization // 2022 13th National Conference with International Participation (ELECTRONICA) (2022): 4 pp. (online). Discussion and implementation of the research results: 2 articles in journals indexed in Scopus and Web of Science: 1. Ospanova A., Zharkimbekova A., Kussepova L., Tokkuliyeva A., Kokkoz M. Cloud Service for Protecting Computer Networks of Enterprises Using Intelligent Hardware and Software Devices, Based on Raspberry Pi Microcomputers // Acta Polytechnica Hungarica 19(4) (2022): 85-103 pp. (Quartile - Q3, percentile - 80). 2. Omarbekova A., Kussepova L., Zulkhazhav A., Altaibek M., Bekmanova G., Nazyrova A., Lamasheva Zh. Monitoring and comparative analysis of containerized environments // Frontiers in Computer Science 8 (2026): 1886850, 1-22 pp. https://doi.org/10.3389/fcomp.2026.1886850 (Quartile - Q1, percentile - 71) 3 articles in journals included in the list of scientific publications recommended by the Committee for Quality Assurance in Science and Higher Education of the Ministry of Science and Higher Education of the Republic of Kazakhstan: 1. Кусепова Л.Т., Оспанова А.Б, Назырова А.Е., Кусепова Г.Т. Развертывание контейнерных приложений с помощью машинного обучения: обзор и анализ // Вестник КазУТБ Учредители: Казахский университет технологии и бизнеса. – 2024. – Т. 2. – №. 23. 2. Кусепова Л. Т., Кусепова Г. Т. АНАЛИЗ И СРАВНЕНИЕ ИНСТРУМЕНТОВ ОРКЕСТРАЦИИ КОНТЕЙНЕРОВ В ЛОКАЛЬНОЙ И ОБЛАЧНОЙ ИНФРАСТРУКТУРЕ //Вестник Казахстанско-Британского технического университета. – 2026. – Т. 23. – №. 1. – С. 22-36. 3. Kussepova L.T., Ospanova A.B., Tebueva F.B., Amandykkyzy A., Nazyrova A.E., Tuleshov Y.A. An overview of modern cloud container technologies and a comparative analysis of their orchestration tools // Bulletin of the National Engineering Academy of the Republic of Kazakhstan. – 2026. – №1 (99). – P. 112-125. Structure and scope of the dissertation. The dissertation is written in Kazakh and consists of an introduction, two sections, a conclusion, a list of references comprising 247 sources, and appendices. The introduction substantiates the relevance of the research topic and defines the aim and objectives of the work, the object and subject of the research, and the methods used. The scientific novelty, practical significance of the research, and main results submitted for defence are formulated. Information on the approbation, publication, and practical implementation of the results is provided. The first section examines the scientific and theoretical foundations of creating and deploying containerized applications in cloud and local infrastructures. Architectural models of cloud infrastructure, container technologies, and orchestration platforms are systematized. A comparative analysis of algorithms for container placement, scaling, load balancing, and resource utilization optimization is conducted. The limitations of existing approaches under incomplete telemetry and changing workloads are identified, and the need to develop adaptive management models is substantiated. Criteria for evaluating the effectiveness of the proposed solutions and the experimental research methodology are established. The second section develops and experimentally evaluates models and algorithms for adaptive management during the creation, deployment, and execution of containerized applications. The resource management problem is formalized as a partially observable Markov decision process, and a belief-state-based decision-making approach is considered. A bidirectional LSTM architecture with an attention mechanism for workload forecasting and the SLOPE+ proactive scaling algorithm are investigated. The CASR-PPO algorithm for reducing cold starts in a serverless environment and approaches to multi-agent management of container infrastructure are evaluated. For the adaptive selection of an execution platform, the EF-LinUCB algorithm with an exponential reduction in the weight of old observations is considered. The Con2ValInt method for three-level verification of container configurations is integrated into a verification and correction cycle based on large language models and retrieval-augmented generation. Container image build strategies are compared, and the computational costs and conditions of application of the developed components are identified. An architecture based on telemetry and feedback is proposed to integrate the components into a “creation – deployment – adaptive management” cycle. The conclusion summarizes the main results obtained in relation to the research objectives and formulates the scientific and practical significance of the proposed models and algorithms. Based on the experimental results, the conditions and limitations of their application are clarified, and the relationship between service quality, resource utilization, and computational costs is evaluated. The appendices present comparative descriptions of container scheduling algorithms and copyright certificates for the software tools developed.
