Method, apparatus, mlops system, and recording medium for constructing mlops system environment

The automated MLOps system addresses integration and scalability challenges by dynamically configuring container orchestration and optimizing hyperparameters, resulting in efficient and flexible AI model deployment.

WO2026235270A1PCT designated stage Publication Date: 2026-11-12GANGNEUNG WONJU NAT UNIV IND ACAD COOPERATION GROUP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/KR2026/002956
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2026-02-10
Filing Date
2026-02-23
Publication Date
2026-11-12

AI Technical Summary

Technical Problem

Conventional MLOps platforms face inefficiencies due to manual integration of distributed and heterogeneous components, high entry barriers, version incompatibility, and lack of flexibility in scalability and maintenance, leading to complex management and system instability.

Method used

An automated method for constructing an MLOps system environment that detects available resources in real-time, dynamically configures a container orchestration cluster, and automatically installs and configures machine learning execution modules and workflow modules, while optimizing hyperparameters and defining service specifications to ensure seamless integration and scalability.

Benefits of technology

The solution enables efficient, scalable, and flexible deployment of AI models by automatically constructing an optimized computing infrastructure without user intervention, reducing complexity, and ensuring stable operation and resource utilization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure KR2026002956_12112026_PF_FP_ABST
    Figure KR2026002956_12112026_PF_FP_ABST
Patent Text Reader

Abstract

The present invention relates to a technology for automatically constructing an MLOps system environment, and more particularly, to a method, an apparatus, a system, and a recording medium for detecting available resources of a target system, dynamically configuring a container orchestration cluster based on the detected resources, and automatically installing and configuring machine learning execution modules and workflow management modules on the cluster.
Need to check novelty before this filing date? Find Prior Art

Description

METHOD, APPARATUS, MLOPS SYSTEM, AND RECORDING MEDIUM FOR CONSTRUCTING MLOPS SYSTEM ENVIRONMENT

[0001] The present invention relates to an MLOps (Machine Learning Operations) system for supporting the development and operation of artificial intelligence models, and more particularly, to a method, an apparatus, an MLOps system, and a recording medium for constructing MLOps system environment that supports the establishment of an execution environment for artificial intelligence–related software.

[0002] MLOps (Machine Learning Operations) is a methodology for automating and systematizing the entire lifecycle of machine learning models, including development, deployment, and operation. MLOps encompasses processes such as data collection and preprocessing, model design, hyperparameter optimization and automated training, performance evaluation, and deployment of trained models to API servers or edge devices. Through continuous maintenance of model quality and rapid deployment, MLOps is regarded as a core technology for ensuring the reliability and scalability of artificial intelligence systems.

[0003] However, conventional approaches to constructing MLOps platforms expose several structural limitations and technical challenges.

[0004] First, inefficiencies arise due to manual integration among distributed and heterogeneous components. Most existing systems are built by individually introducing data repositories, experiment management tools, pipeline orchestration tools, and model deployment tools. As interface integration among these tools is performed manually, the complexity of integrated management increases exponentially, thereby hindering seamless end-to-end pipeline connectivity.

[0005] Second, there exists a high entry barrier associated with the installation and configuration of individual tools. Each open-source component typically requires a complex setup process demanding advanced technical expertise, resulting in excessive consumption of time and resources for infrastructure construction. Consequently, initial system deployment becomes impractical or significantly delayed without highly skilled specialists.

[0006] Third, version incompatibility among components and software dependency conflicts frequently occur. When multiple open-source software components are combined, version discrepancies among specific modules often lead to overall system instability, and diagnosing and resolving the root causes of such compound errors becomes extremely difficult.

[0007] Finally, there is a lack of flexibility in scalability and maintenance in response to hardware resource changes. When system infrastructure is expanded or upgraded, environment variables and configuration settings of previously deployed components must be manually readjusted. Such rigid structures impede agile adaptation to evolving computing environments and increase the maintenance burden during system operation.

[0008] Accordingly, there is a need for an intelligent automated construction technology capable of minimizing user intervention while autonomously detecting available hardware resources of a target system and dynamically configuring optimized container clusters and MLOps pipelines based on the detected results. In particular, an integrated MLOps environment construction technology is required that sequentially installs modularized execution tools to fundamentally prevent dependency conflicts among open-source components and comprehensively establishes a data storage mirroring environment to ensure post-deployment operational stability.

[0009] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can automatically build a computing infrastructure optimized for an individual hardware environment by detecting available resources of a target system in real time and dynamically configuring a container orchestration cluster based on the detected resources.

[0010] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can resolve complex interdependencies by setting priorities among a plurality of machine learning execution modules and workflow modules installed on the container orchestration cluster and sequentially performing automatic installation and configuration without user intervention.

[0011] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can automatically link and activate interfaces between internal resources of the container orchestration cluster and external storage in order to store and manage at least a portion of training data, model files, and model artifacts of an artificial intelligence model.

[0012] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can provide a customized execution environment by dynamically generating container build files based on available resources and improve efficiency of system resource utilization by automatically releasing resources unnecessarily occupied after completion of pipeline tasks.

[0013] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can reduce the complexity of performance improvement tasks of an artificial intelligence model by automatically configuring a module for optimizing hyperparameters to enhance performance of the artificial intelligence model.

[0014] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can improve deployment flexibility by defining service specifications separately from execution environments, and can provide a non-disruptive operation environment without data loss during inference services by establishing an object storage mirroring environment based on real-time available capacity information.

[0015] It is an object of the present invention to provide a method, an apparatus, an MLOps system, and a recording medium for constructing an MLOps system environment, which can dynamically perform an entire process of constructing an MLOps system environment without user intervention by executing a single file including a script.

[0016] According to one embodiment of the present invention for achieving the above-described objects, a method for constructing an MLOps (Machine Learning Operations) system environment includes:

[0017] (a) detecting available resources of a target system, and dynamically configuring a container orchestration cluster in which a plurality of nodes are logically clustered so as to be optimized for system specifications based on the detected available resources; and

[0018] (b) automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with the resource configuration of the cluster.

[0019] In this case, the step (a) may include:

[0020] (a1) generating a deployment file in which detection logic for collecting available resource states and hardware specification information of the system is defined; and

[0021] (a2) executing the generated deployment file to dynamically determine and configure the container orchestration cluster in correspondence with the collected specification information.

[0022] Further, the step (b) may include sequentially performing automatic installation and configuration according to priorities set based on interdependencies among the plurality of machine learning execution modules and workflow management modules, without user intervention.

[0023] Further, the workflow management module may be configured to:

[0024] (i) support a preprocessing process for input data, a training and evaluation process for an artificial intelligence model, and an inference process using a trained AI model; and

[0025] (ii) support definition and allocation of data storage resources in response to requests, and deployment of the allocated resources in the form of PODs.

[0026] Further, the plurality of machine learning execution modules may include an object storage, wherein the object storage may be configured such that interfaces between internal resources of the container orchestration cluster and external storage are automatically linked and activated in order to store and manage at least a portion of training data, model files, and model artifacts of an artificial intelligence model.

[0027] Further, the step (b) may include generating execution configuration information including at least one of service port information, user account information, and storage volume information in order to construct the object storage environment, and generating and activating the object storage in a containerized form based on the generated execution configuration information.

[0028] Further, the plurality of machine learning execution modules may include an interactive code execution module, and the step (b) may include:

[0029] (b1) dynamically generating a container build file including base image information and library configuration information for execution targets based on the detected available resources of the system;

[0030] (b2) generating a user-customized container image based on the generated container build file; and

[0031] (b3) registering the generated container image in a repository and activating the container image so as to be usable in conjunction with the workflow management module.

[0032] Further, the plurality of workflow management modules may include a pipeline management module, and the step (b) may include:

[0033] (b3) generating a pipeline configuration file defining each stage of machine learning execution as an independent container; and

[0034] (b4) configuring management logic to dynamically release unnecessary resources by monitoring system resources allocated at a completion time of pipeline tasks to be executed in the future.

[0035] Further, the plurality of machine learning execution modules may include a model optimization module for automatically optimizing hyperparameters in order to enhance performance of an artificial intelligence model, and the step (b) may include:

[0036] (b5) installing a plurality of optimization modules for searching an optimal combination of hyperparameters based on at least one of a plurality of optimization algorithms; and

[0037] (b6) configuring the model optimization module to be executed in a distributed search environment in conjunction with the workflow management module.

[0038] Further, the plurality of machine learning execution modules may include a lifecycle-based experiment and tracking management module for models, and the step (b) may include:

[0039] (b7) constructing a base environment having reproducible and consistent characteristics based on library and version information for lifecycle-based experiment and tracking management of models; and

[0040] (b8) configuring an API interface between a specific artificial intelligence model and the base environment such that experiment-related data of the specific artificial intelligence model is automatically transmitted to and recorded in the base environment.

[0041] Further, the plurality of machine learning execution modules may include a model deployment and service module for deploying and servicing trained models, and the step (b) may include:

[0042] (b9) configuring a service configuration module for defining an application service specification including internal dependencies and execution paths of the trained model;

[0043] (b10) configuring an execution environment configuration module for defining an infrastructure execution environment for operating the model for which the service specification is defined in a containerized form; and

[0044] (b11) configuring a service module having interface logic for servicing the containerized model in an online or local-based manner in association with the execution environment.

[0045] Further, the method may further include:

[0046] (b12) configuring a mirroring environment for the object storage based on real-time detected available storage capacity information so that data processing according to inference requests is continuously performed without interruption, thereby ensuring data availability.

[0047] Further, the method may be configured such that an entire MLOps system environment is automatically constructed without user intervention by executing a single file including a script for performing all processes including the step (a) and the step (b).

[0048] According to one embodiment of the present invention, an apparatus for constructing an MLOps system environment includes:

[0049] an interface unit configured to support transmission of environment construction commands to a system; and

[0050] a processor configured to detect, or support detection of, available resources of a target system, dynamically configure a container orchestration cluster in which a plurality of nodes are logically clustered so as to be optimized for system specifications based on the detected available resources, and automatically install and configure at least one of a plurality of machine learning execution modules and workflow management modules on the configured container orchestration cluster in association with the resource configuration of the cluster.

[0051] A non-transitory computer-readable recording medium storing computer-executable instructions for constructing an MLOps (Machine Learning Operations) system environment according to an embodiment of the present invention for achieving the above-described object is configured such that the instructions cause a computer to perform:

[0052] (a) detecting available resources of a target system to be constructed, and dynamically configuring a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications; and

[0053] (b) automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with a resource configuration of the cluster.

[0054] An MLOps (Machine Learning Operations) system according to an embodiment of the present invention for achieving the above-described object includes:

[0055] an interface unit configured to receive environment construction commands; and

[0056] a processor configured to detect or support detection of available resources of a system, dynamically configure a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications, and automatically install and configure at least one of a plurality of machine learning execution modules and workflow management modules on the configured container orchestration cluster in association with a resource configuration of the cluster.

[0057] The present invention has the effect of automatically constructing a computing infrastructure optimized for an individual hardware environment by detecting available resources of a target system in real time and dynamically configuring a container orchestration cluster based thereon.

[0058] The present invention has the effect of resolving complex interdependency issues by setting priorities among a plurality of machine learning execution modules and workflow modules installed on the container orchestration cluster, and sequentially performing automatic installation and configuration without user intervention.

[0059] The present invention has the effect of enabling automatic linkage and activation of interfaces between internal resources of the container orchestration cluster and external storage in order to store and manage at least a portion of training data, model files, and model artifacts of an artificial intelligence model.

[0060] The present invention has the effect of providing a customized execution environment by dynamically generating container build files based on available resources, and improving efficiency of system resource utilization by automatically releasing unnecessarily occupied resources upon completion of pipeline tasks.

[0061] The present invention has the effect of reducing the difficulty of performance improvement tasks for artificial intelligence models by automatically configuring a module for optimizing hyperparameters to enhance model performance.

[0062] The present invention has the effect of improving deployment flexibility by defining service specifications and execution environments separately, and providing a non-disruptive operation environment without data loss even during inference services by constructing an object storage mirroring environment based on real-time available capacity information.

[0063] The present invention has the effect of dynamically performing an entire process of constructing an MLOps system environment automatically without user intervention by executing a single file including a script.

[0064] FIG. 1 is a block diagram illustrating a configuration of an MLOps system environment construction apparatus according to an embodiment of the present invention.

[0065] FIG. 2 is a block diagram illustrating a configuration of an MLOps system according to an embodiment of the present invention.

[0066] FIG. 3 is a diagram illustrating a concept of a method for constructing an MLOps system environment according to an embodiment of the present invention.

[0067] FIG. 4 is a diagram illustrating a method for constructing an MLOps system environment according to an embodiment of the present invention.

[0068] FIG. 5 is a flowchart illustrating a method for constructing an MLOps system environment according to an embodiment of the present invention.

[0069] The best mode for carrying out the present invention is described in the following embodiments.

[0070] Hereinafter, various embodiments of the present invention will be described with reference to the accompanying drawings.

[0071] In this process, the thicknesses of lines shown in the drawings and the sizes of components may be exaggerated for clarity and convenience of explanation. In addition, terms described below are defined in consideration of functions in the present invention, and may vary depending on intentions or customary practices of users or operators. Therefore, definitions of such terms should be construed based on the overall contents of this specification.

[0072] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that a person having ordinary skill in the art to which the present invention pertains can readily carry out the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, parts irrelevant to the description are omitted in the drawings for clarity, and like reference numerals designate like elements throughout the specification.

[0073] Throughout the specification, when any part is described as “including” a component, it does not exclude other components unless explicitly stated otherwise, and may include additional components.

[0074] Terms and words used in this specification and the claims should not be construed as being limited to their ordinary or dictionary meanings, but should be interpreted in accordance with meanings and concepts consistent with the technical spirit of the present invention, based on the principle that an inventor may appropriately define the concepts of terms in order to describe the invention in the best possible manner.

[0075] Accordingly, the embodiments described in this specification and the configurations illustrated in the drawings are merely some preferred embodiments of the present invention and do not represent all technical aspects of the present invention, and it should be understood that various equivalents and modifications capable of replacing them may exist at the time of filing of the present application.

[0076] In addition, for the purpose of facilitating understanding of the invention, the accompanying drawings are not necessarily drawn to scale, and dimensions of some components may be exaggerated. Further, the same reference numerals may be assigned to the same components in different embodiments.

[0077] Although terms such as “first” and “second” are used to describe various components, such components are not limited by such terms. These terms are merely used to distinguish one component from another, and unless otherwise explicitly stated, a first component may be a second component.

[0078] Throughout the specification, unless explicitly stated otherwise, each component may be singular or plural.

[0079] In the specification, the expression “supporting OO” means that one device performs one or more operations that enable another device to perform a specific operation, including but not limited to providing data, transmitting control commands, performing authentication, providing a user interface, performing computations, transmitting script files, controlling transmission of script files, or performing other related operations.

[0080] FIG. 1 is a block diagram illustrating a configuration of an MLOps system environment construction apparatus according to an embodiment of the present invention.

[0081] Referring to FIG. 1, an MLOps system environment construction apparatus 100 according to an embodiment of the present invention includes an interface unit 110 and a processor 120.

[0082] The interface unit 110 is configured to transmit an environment construction command for constructing an MLOps system environment to a system 200. The environment construction command may include an initial configuration request, a reconfiguration request, or a configuration change request for the MLOps environment.

[0083] In one embodiment, the interface unit 110 supports the system 200 to sequentially perform installation by transmitting to the MLOps system 200 a file including a script for automatically performing an entire process of constructing the MLOps system environment, such that the system 200 reads the script and retrieves installation files from an external code repository (see reference numeral 300 in FIG. 3).

[0084] The interface unit 110 may be configured to output, under control of the processor 120, progress status, configuration results, or error information generated during the MLOps environment construction process to the outside, thereby supporting a user or an external system to monitor and manage the environment construction status.

[0085] The interface unit 110 may include a communication module for processing wired or wireless communications. The communication module may include, for example, a network interface card (NIC) supporting Ethernet, a wireless local area network (Wi-Fi), Bluetooth, or the like. In addition, the interface unit 110 may include an input / output (I / O) control module configured to convert direct control commands from a user into electrical signals and transmit the converted signals to the processor 120.

[0086] The processor 120 controls overall operations of the apparatus 100 and performs computation and control functions for supporting the system 200 to automatically construct an MLOps system environment.

[0087] In one embodiment, the processor 120 controls the interface unit 110 to transmit to the system 200 a single file including a script for automatically performing an entire process of system environment construction, such that the system 200 executes the script to retrieve installation files from an external code repository 300 and sequentially install the MLOps system environment.

[0088] Through this, the processor 120 supports detecting or supporting detection of available resources of a target system, and supports dynamically configuring a container orchestration cluster (200-1 in FIG. 3) in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications. In addition, the processor 120 may support automatically installing and configuring at least one of a plurality of machine learning execution modules 200-2 and workflow management modules 200-3 on the configured container orchestration cluster 200-1 in association with a resource configuration of the cluster.

[0089] The processor 120 may include a hardware configuration including one or more central processing units (CPUs), graphics processing units (GPUs), signal processing units, or similar computing units, and a software configuration including an operating system, firmware, application programs, and artificial intelligence algorithms executed on the hardware. The processor 120 is configured to perform data processing, control, analysis, and decision-making functions in association with a memory and a storage unit, and specific configurations and operation methods thereof may be variously modified depending on an implementation environment and an application purpose.

[0090] The MLOps system environment construction apparatus 100 may further include, in addition to the interface unit 110 and the processor 120 described above, a power supply unit for supplying power, a control unit for operation control and state management, and other components for assisting operations of the apparatus, although not illustrated in the drawings. Such components may be implemented as a single module or distributed into a plurality of modules, and configurations and connection methods thereof may be variously modified depending on an application environment and a usage purpose of the apparatus.

[0091] FIG. 2 is a block diagram illustrating a configuration of an MLOps system according to an embodiment of the present invention, and FIG. 3 is a diagram illustrating an operational concept of the MLOps system environment construction apparatus according to an embodiment of the present invention.

[0092] Referring to FIG. 2, an MLOps system 200 according to an embodiment of the present invention includes an interface unit 210, a processor 220, a memory unit 230, and a storage unit 240. In addition, the MLOps system 200 includes a plurality of parallel processor resources 200-1a, 200-1b, …, 200-1n (where n is a natural number), and a plurality of terminals 400-1, 400-2, …, 400-m (where m is a natural number) configured to request tasks from the system 200.

[0093] The interface unit 210 receives, from the apparatus 100, a script file for automatically performing an entire process of constructing an MLOps system environment, and receives MLOps environment installation files from an external code repository 300. The interface unit 210 supports processing of control commands and configuration information for MLOps environment construction within the apparatus by transmitting the received data to the processor 220.

[0094] The interface unit 210 may be configured to output, under control of the processor 220, progress status, configuration results, or error information generated during the MLOps environment construction process to the outside, thereby supporting a user or an external system to monitor and manage the environment construction status. In addition, the interface unit 210 may support transmitting information related to construction and operation of the MLOps environment to the apparatus 100 for output.

[0095] The interface unit 210 may include a communication module for processing wired or wireless communications. The communication module may include, for example, a network interface card (NIC) supporting Ethernet, a wireless local area network (Wi-Fi), Bluetooth, or the like. Further, the interface unit 210 may include an input / output (I / O) control module configured to convert direct control commands from a user into electrical signals and transmit the converted signals to the processor 220.

[0096] The processor 220 may be configured to control construction of an MLOps environment based on data received through the interface unit 210. The processor 220 may detect available resources of a system or identify system resource availability information received from the apparatus 100, and dynamically configure a container orchestration cluster (see reference numeral 200-1 in FIG. 3) in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications.

[0097] In addition, the processor 220 may automatically install and configure at least one of a plurality of machine learning execution modules 200-2 and workflow management modules 200-3 on the configured container orchestration cluster 200-1 in association with a resource configuration of the cluster. Accordingly, the processor 220 may control automatic construction of the MLOps environment without user intervention.

[0098] In order to detect available resources of a target system, the processor 220 may collect system resource information and hardware specification information such as CPU utilization, memory capacity, storage capacity, network bandwidth, and availability and quantity of parallel processor resources. The resource information may be collected in real time or over a predetermined period, and based on the collected resource information, the processor 220 may determine a configuration type and scale of the container orchestration cluster 200-1 that can be constructed in the corresponding system environment.

[0099] Herein, the parallel processor resources refer to computing resources for performing parallel processing of artificial intelligence tasks, and may include at least one hierarchical unit selected from: cores or threads as computation units within a single chip; graphics processing units (GPUs); neural processing units (NPUs); tensor processing units (TPUs); accelerator chips including at least one of the foregoing processing units; boards or modules on which a plurality of accelerator chips are mounted; and nodes or servers in which a plurality of modules are connected via a network. The parallel processor resources may be allocated to artificial intelligence tasks individually or in combination, and may be utilized in logically integrated or logically partitioned forms.

[0100] In addition, the processor 220 may dynamically configure a container orchestration cluster 200-1 in which a plurality of nodes are logically grouped so as to be optimized for system specifications based on the detected available resource information. In this case, configuration of the cluster 200-1 may take into account the number of nodes, role assignment of each node, resource allocation schemes, scheduling policies, and the like.

[0101] The processor 220 may automatically install and configure at least one of a plurality of machine learning execution modules 200-2 and workflow management modules 200-3 on the configured container orchestration cluster 200-1.

[0102] In this case, the machine learning execution modules 200-2 may provide execution functions for performing at least a portion of development, training, evaluation, and inference of artificial intelligence models, and the workflow management modules 200-3 may be configured to manage an execution order, dependencies, and execution flows of machine learning tasks.

[0103] The processor 220 may analyze interdependency relationships between the machine learning execution modules 200-2 and the workflow management modules 200-3, and may control automatic determination of an installation and configuration sequence based on analysis results. Accordingly, the respective modules may be installed and configured in an appropriate order without intervention of a user 10, thereby enabling consistent construction of the overall MLOps system environment.

[0104] In addition, the processor 220 may automatically configure, in association with a resource configuration of the container orchestration cluster 200-1, resource usage ranges, data access methods, and execution conditions to be used by each machine learning execution module 200-2 or workflow management module 200-3. Through this, an MLOps execution environment suitable for resource characteristics of a target system may be efficiently constructed.

[0105] Meanwhile, the processor 220 may sequentially perform construction of the MLOps system environment by executing a script configured to execute the MLOps system environment construction according to a predetermined order that takes interdependencies into account. In this case, the container orchestration cluster 200-1 and a machine learning execution environment may be collectively constructed without requiring repetitive configuration inputs from a user.

[0106] The processor 220 may generate a deployment file in which resource detection logic is defined in order to collect available resource states and hardware specification information of a target system. The deployment file may be generated as a deployable resource executable in a container orchestration environment, and may include detection logic for collecting available resource information such as CPU resources, memory resources, storage resources, network resources, and accelerator resources (e.g., GPUs) on the system.

[0107] In generating the deployment file, the processor 220 may dynamically configure deployment settings for enabling the detection logic to be normally executed in the corresponding system environment by referring to at least a portion of an operating environment of the target system, a type of operating system, a container runtime environment, and predefined detection policies. Accordingly, the deployment file may function not merely as a configuration file but as an executable object that actively collects actual resource states of the system at an execution time.

[0108] The processor 220 may execute the generated deployment file on the container orchestration cluster 200-1. In this case, the collected resource information may be real-time information or information accumulated over a predetermined period, and the processor 220 may analyze resource characteristics of a target system based on the collected information.

[0109] The processor 220 may dynamically determine a configuration scheme of the container orchestration cluster 200-1 in response to the resource information collected as a result of execution of the deployment file. For example, based on the collected resource information, at least a portion of a number of nodes constituting the cluster 200-1, role assignment of each node, resource allocation criteria, scheduling policies, and scalability may be determined.

[0110] As described above, instead of applying predefined fixed cluster configuration information, the present invention dynamically determines and configures the container orchestration cluster 200-1 based on actual system specification information collected at an execution time through the deployment file, thereby automatically providing an MLOps execution foundation optimized for resource characteristics of a target system.

[0111] In one embodiment, the container orchestration cluster 200-1 may include a Kubernetes cluster. The processor 220 may monitor node resources through a deployment file or automatically detect system specifications based on specification-based determination logic, and may select a Kubernetes installation tool most suitable for the detected system specifications from among k3s, kind, kubeadm, and minikube to automatically perform installation and initial configuration.

[0112] The workflow management module 200-3 is configured to manage an overall flow of machine learning tasks, and may be set to support a data preprocessing process for input data, a training and evaluation process of an artificial intelligence model, and an inference process using the trained artificial intelligence model. The workflow management module 200-3 may be configured to define and manage step-by-step flows and execution conditions of machine learning tasks such that respective processes are executed independently while being interlinked with one another.

[0113] In one embodiment, in order to support storage and access of data used in machine learning tasks, the workflow management module 200-3 may define data storage resources and be configured to receive allocation of storage resources from the container orchestration cluster 200-1 in response to requests for the defined storage resources. In this case, the allocated data storage resources may be deployed in the form of pods in a container orchestration environment, and may be configured to be continuously accessible while preprocessing, training, evaluation, and inference processes are being executed.

[0114] Accordingly, the workflow management module 200-3 may support reliable storage, sharing, and utilization of data required at each stage of machine learning tasks, thereby enabling machine learning task flows to be efficiently and consistently performed on the cluster 200-1.

[0115] In one embodiment, the workflow management module 200-3 may include Kubeflow. Kubeflow is an open-source platform designed to facilitate construction and operation of machine learning pipelines based on Kubernetes, and is a tool capable of automating an entire process from training to deployment of machine learning models while fully leveraging scalability and stability of a Kubernetes-based environment.

[0116] In addition, the plurality of machine learning execution modules may include an object storage. The object storage is configured to store and manage at least a portion of training data, model files, and model artifacts of an artificial intelligence model, and may be set to efficiently manage various data generated or used during execution of machine learning tasks.

[0117] In addition, the object storage may be configured such that interfaces between internal resources of the container orchestration cluster 200-1 and external storage are automatically linked and activated. Accordingly, the object storage is activated through an interface with machine learning execution modules 200-2 executed in the cluster environment, and data required for preprocessing, training, evaluation, and inference processes of machine learning tasks may be reliably stored and accessed.

[0118] In one embodiment, the processor 220 may generate execution configuration information including at least a portion of service port information, user account information, and storage volume information in order to construct an object storage environment. The execution configuration information may define basic execution conditions required for the object storage to normally operate in a container orchestration environment.

[0119] In addition, the processor 220 may control creation and activation of the object storage in a container form based on the generated execution configuration information. Accordingly, the object storage is executed on the container orchestration cluster 200-1 and may provide data storage and management functions in association with the machine learning execution modules 200-2 and the workflow management module 200-3.

[0120] In one embodiment, the object storage may include a MinIO object storage. The MinIO object storage is a cloud-native object storage system and is an open-source object storage for storing and retrieving data. The processor 220 may configure required execution configuration information using Docker Compose in order to execute the MinIO object storage in a container form. The execution configuration information configured by Docker Compose may include at least a portion of service port information, environment variables, user account information, and data storage volume information. Accordingly, the system 200 may be configured to execute and manage the MinIO object storage in a local environment or a container orchestration cluster environment using Docker Compose.

[0121] In addition, the plurality of machine learning execution modules may include an interactive code execution module. The interactive code execution module is an execution environment that supports a user 10 to immediately confirm code execution results for development of machine learning models, data analysis, and experiments, and may be configured in a container-based manner.

[0122] The processor 220 may dynamically generate a container build file including base image information and library configuration information suitable for a target execution environment based on detected available resources of the system. In the container build file, execution environment information required for normal operation of the interactive code execution module may be defined.

[0123] In addition, the processor 220 may generate a user-customized container image based on the generated container build file. In this case, the generated container image may be configured to correspond to resource characteristics of a target system and user-defined settings.

[0124] The processor 220 may register the generated container image in a repository and activate the container image to be usable in association with the workflow management module 200-3. Accordingly, the interactive code execution module may be executed on the container orchestration cluster 200-1 and utilized for development and experimentation processes in conjunction with machine learning task flows.

[0125] In one embodiment, the interactive code execution module may include a notebook module. The notebook is an interactive code execution environment used in machine learning, data analysis, and visualization tasks, and is a development tool that allows code, descriptions (markdown), visualizations, and results to be executed and viewed on a single screen. The processor 220 may generate a Docker file defining a base image, essential libraries, and configuration information for an application to be executed in the notebook environment. In addition, the processor 220 may generate a Docker image required for the notebook environment based on contents defined in the Docker file. Further, the processor 220 may upload the generated Docker image to a Docker Hub and configure Kubeflow to use the Docker image. Accordingly, a user-customized development environment capable of easily performing data preprocessing, analysis, model training, and experiments may be provided in a container-based manner.

[0126] In addition, the plurality of workflow management modules 200-3 may include a pipeline management module. The pipeline management module is configured to systematically manage a plurality of task stages performed during execution of machine learning tasks, and may support independent execution of each stage of the machine learning execution.

[0127] The processor 220 may generate a pipeline configuration file that defines each stage of the machine learning execution as an independent container. The pipeline configuration file may include container definition information corresponding to respective machine learning execution stages and an execution order between the stages, such that machine learning tasks including preprocessing, training, evaluation, and inference are configured and executed as a single pipeline.

[0128] In addition, the processor 220 may configure management logic capable of monitoring system resources allocated at a completion time of pipeline tasks to be executed in the future, and dynamically releasing resources that are no longer required for pipeline tasks whose execution has been completed. By such management logic, resources of the container orchestration cluster 200-1 may be efficiently reclaimed and reused even after execution of the pipeline.

[0129] Accordingly, the pipeline management module may improve overall resource utilization efficiency of an MLOps environment by structuring and managing machine learning execution tasks on a stage-by-stage basis and automatically cleaning up unnecessary system resources after execution is completed.

[0130] In one embodiment, the processor 220 may generate and register a Kubeflow Pipeline definition file to control construction of an artificial intelligence workflow from an initial stage. Through this, pipeline tasks that are currently being executed or have been completed may be cleaned up, and unnecessary resources may be released.

[0131] In addition, the plurality of machine learning execution modules 200-2 may include a model optimization module configured to automatically optimize hyperparameters for improving performance of an artificial intelligence model. The model optimization module is configured to automatically search for and adjust values of hyperparameters that affect performance during a training process of the machine learning model, and may utilize various optimization techniques. Here, hyperparameters are configuration values that affect performance of an artificial intelligence model, and hyperparameter optimization refers to a process of detecting and adjusting combinations of hyperparameters that maximize accuracy of the model or minimize loss in order to improve performance of the artificial intelligence model.

[0132] The processor 220 may install a plurality of optimization modules for searching for an optimal combination of hyperparameters based on at least one of a plurality of optimization algorithms. The optimization modules may be configured to generate combinations of hyperparameters according to different search strategies and to comparatively analyze model training results for each combination.

[0133] In addition, the processor 220 may enable the model optimization module to be executed in a distributed search environment in association with the workflow management module 200-3. Accordingly, model training and evaluation tasks for a plurality of hyperparameter combinations may be distributed and performed in parallel across a plurality of nodes of the container orchestration cluster 200-1, thereby improving a time required for hyperparameter search and efficiency of system resource utilization.

[0134] As described above, by automating and configuring a hyperparameter search process in a scalable manner through association with the workflow management module, the model optimization module may effectively support performance improvement of artificial intelligence models in an MLOps environment.

[0135] In one embodiment, the model optimization module may include an AutoML module. AutoML (Automated Machine Learning) is a system that automates an entire machine learning process so that models can be developed even by non-experts, and is a technology that automatically performs processes from data preprocessing to model selection and hyperparameter tuning.

[0136] In one embodiment, the plurality of model optimization modules may include a first optimization module, a second optimization module, a third optimization module, and a fourth optimization module. The first optimization module may be Goptuna, which performs distributed optimization in a Kubernetes environment to provide efficient search strategies. The second optimization module may be Hyperopt, which performs efficient hyperparameter optimization using a Tree-structured Parzen Estimator algorithm. The third optimization module may be a Scikit optimizer, which performs hyperparameter optimization using a Bayesian Optimization algorithm and is capable of finding high-performance hyperparameters with a relatively small number of trials. The fourth optimization module may be Optuna, which enables rapid application of AutoML with lightweight and intuitive configuration. Through this, hyperparameters of an artificial intelligence model may be automatically searched and optimal model performance may be derived.

[0137] In addition, the plurality of machine learning execution modules 200-2 may include an experiment and tracking management module for performing experiment management and tracking over an entire lifecycle of an artificial intelligence model. The experiment and tracking management module is configured to systematically record and manage various experiment information generated during training, evaluation, optimization, and deployment processes of the model, and may support ensuring reproducibility and consistency of a model development process.

[0138] The processor 220 may construct a base environment having reproducible and consistent characteristics based on library configuration information and version information used for experiment and tracking management over an entire lifecycle of a model. The base environment is defined as an execution environment in which model experiments can be repeatedly performed under identical experimental conditions, thereby facilitating comparison and verification of experiment results.

[0139] In addition, the processor 220 may configure an API interface between a specific artificial intelligence model and the base environment such that experiment-related data generated during an experiment process of the specific artificial intelligence model, including training histories, performance indicators, parameter information, metrics, and model artifacts, are automatically transmitted to and recorded in the base environment. Accordingly, information generated during the model experiment process is automatically collected and stored, thereby enabling systematic tracking and management of a full lifecycle history of the model.

[0140] As described above, the experiment and tracking management module may effectively support quality management and continuous improvement of artificial intelligence models in an MLOps environment by providing records for an entire model development process and a reproducible execution environment.

[0141] In one embodiment, the experiment and tracking management module may include an MLflow module. The MLflow module is an open-source platform that helps manage machine learning projects in a more systematic manner and supports an entire lifecycle from model experimentation to deployment, and is an MLOps tool that supports machine learning experiment tracking, model storage, reproducibility, and deployment.

[0142] In this case, the processor 220 may generate a requirements file defining libraries and version information required for constructing an environment for executing an MLflow server, and may consistently install a project environment using the requirements file. In addition, the processor 220 may automate construction of a base environment for executing the MLflow server and paths and settings for storing experiment results.

[0143] In addition, the plurality of machine learning execution modules 200-2 may include a model deployment and serving module for deploying and servicing trained artificial intelligence models. The model deployment and serving module is configured to stably provide a trained model in an actual usage environment, and may be defined to include dependencies required for model execution and service interfaces. The model deployment and serving module may include a service configuration module, an execution environment configuration module, and a service module.

[0144] The processor 220 may configure a service configuration module for defining an application service specification including at least a portion of internal dependencies of a trained model, a model deployment environment, endpoint information, and execution path information. The application service specification may include libraries, configuration information, and data processing flows required for executing the model, thereby supporting consistent operation of the model in a service environment.

[0145] In addition, the processor 220 may configure an execution environment configuration module for defining an infrastructure execution environment for operating the model for which the service specification is defined in a containerized form. The execution environment configuration module may define resource allocation and execution conditions required for the model service, thereby enabling the model service to be stably executed on the container orchestration cluster 200-1.

[0146] Further, the processor 220 may configure a service module including interface logic for providing the containerized model as an online-based service or a local-based service in association with the execution environment. Accordingly, inference services in response to external requests may be provided through the execution environment, and the model may be utilized in various service forms.

[0147] In addition, the processor 220 may configure a mirroring environment of an object storage based on real-time detected available storage capacity information so that data processing in response to inference requests is performed without interruption. The mirroring environment may be configured such that data used during a model service process is distributed and stored across a plurality of storage systems.

[0148] Accordingly, even when a failure or insufficient capacity occurs in a specific storage system, data access may be continuously maintained through the mirrored object storage, thereby ensuring continuity of model inference services and data availability. Such a mirroring configuration may serve to improve stability during model deployment and servicing processes.

[0149] In one embodiment, the model deployment and serving module may include a BentoML module. BentoML is an open-source MLOps framework that facilitates easy deployment of machine learning models, and is a tool that enables trained machine learning models to be rapidly packaged and deployed as APIs.

[0150] In this case, the processor 220 may configure a Bento file defining an environment for deploying an artificial intelligence model, dependent packages, execution paths, and endpoints. In addition, the processor 220 may generate an environment definition file for executing a model API service configured based on BentoML as a Docker container by using Docker Compose. Further, various execution conditions such as network settings, port mapping, service names, and volumes may be defined through the environment definition file. The above-described service module is a module including logic for loading a trained artificial intelligence model and deploying the model in the form of a REST API, and may be implemented to return processing results of the model when inference requests are received from outside.

[0151] The processor 220 may include a hardware configuration including one or more central processing units (CPUs), graphics processing units (GPUs), signal processing units, or similar computing units, and a software configuration including an operating system, firmware, application programs, and artificial intelligence algorithms executed on the hardware. The processor 220 is configured to perform data processing, control, analysis, and decision-making functions in association with the memory unit 230 and the storage unit 240, and specific configurations and operation methods thereof may be variously modified depending on an implementation environment and an application purpose.

[0152] The MLOps system 100 may further include, in addition to the interface unit 210 and the processor 220 described above, a storage unit for storing data, a power supply unit for supplying power, a control unit for operation control and state management, and other components for assisting operations of the apparatus, although not illustrated in the drawings. Such components may be implemented as a single module or distributed into a plurality of modules, and configurations and connection methods thereof may be variously modified depending on an application environment and a usage purpose of the apparatus.

[0153] FIG. 4 is a diagram illustrating a concept of constructing an MLOps system environment according to an embodiment of the present invention.

[0154] Referring to FIG. 4, an external code repository 300, such as GitHub, stores and manages MLOps system environment construction files. According to an environment construction command from a user 10 through the MLOps system environment construction apparatus 100, an MLOps system environment construction file is retrieved from the external code repository 300, a script is executed in a target system 200, and a Kubernetes cluster, Kubeflow, a MinIO object storage, a notebook, a pipeline, AutoML, MLflow, and BentoML are sequentially and automatically installed on the target system 200.

[0155] FIG. 5 is a flowchart illustrating a method for constructing an MLOps system environment according to an embodiment of the present invention.

[0156] Referring to FIG. 5, the method for constructing an MLOps system environment according to an embodiment of the present invention includes: detecting available resources of a target system (S510); dynamically configuring a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications (S520); and automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with a resource configuration of the cluster (S530).

[0157] A non-transitory computer-readable recording medium according to an embodiment of the present invention, in which computer-executable instructions for constructing an MLOps system environment are recorded, is configured such that the instructions cause a computer to perform:

[0158] (a) detecting available resources of a target system, and dynamically configuring a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications; and

[0159] (b) automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with a resource configuration of the cluster.

[0160] A non-transitory computer-readable recording medium according to an embodiment of the present invention stores instructions readable and executable by a processor of a computer, a server, or an electronic device, and the instructions may include functions for controlling the electronic device to perform training and inference of an artificial intelligence model. The recording medium may be implemented as a hard disk drive (HDD), a solid state drive (SSD), a flash memory, a magnetic tape, an optical disk, or a combination thereof, and has non-transitory storage characteristics in which stored information is retained even when power supply is cut off.

[0161] Although the present invention has been described above with reference to limited embodiments and drawings, the present invention is not limited thereto, and various modifications and variations may be made by a person having ordinary skill in the art to which the present invention pertains within the scope of the technical spirit of the present invention and the equivalent scope of the claims set forth below.

[0162] The present invention is industrially applicable to artificial intelligence model development and operation, cloud computing environments, and automated management of machine learning systems.

Claims

1.A method for constructing an MLOps (Machine Learning Operations) system environment, comprising:(a) detecting available resources of a target system, and dynamically configuring a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications; and(b) automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with the resource configuration of the cluster.2.The method according to claim 1,wherein the step (a) comprises:(a1) generating a deployment file in which detection logic for collecting available resource states and hardware specification information of the system is defined; and(a2) executing the generated deployment file to dynamically determine and configure the container orchestration cluster so as to correspond to the collected specification information.3.The method according to claim 1 or 2,wherein the step (b) comprises automatically installing and configuring the plurality of machine learning execution modules and workflow management modules sequentially without user intervention according to priorities set based on interdependencies among the modules.4.The method according to any one of claims 1 to 3,wherein the workflow management module is configured to:(i) support a preprocessing process of input data, a training and evaluation process of an artificial intelligence model, and an inference process using a trained artificial intelligence model; and(ii) support definition of data storage resources, allocation of the data storage resources in response to requests, and deployment of the allocated resources in a form of pods.5.The method according to any one of claims 1 to 4,wherein the plurality of machine learning execution modules includes an object storage,and the object storage is configured such that interfaces between internal resources of the container orchestration cluster and external storage are automatically linked and activated in order to store and manage at least a portion of training data, model files, and model artifacts of an artificial intelligence model.6.The method according to claim 5,wherein the step (b) comprises:generating execution configuration information including at least a portion of service port information, user account information, and storage volume information in order to construct an object storage environment; andcreating and activating the object storage in a container form based on the generated execution configuration information.7.The method according to any one of claims 1 to 6,wherein the plurality of machine learning execution modules includes an interactive code execution module,and the step (b) comprises:(b1) dynamically generating a container build file including base image information and library configuration information for a target execution environment based on the detected available resources of the system;(b2) generating a user-customized container image based on the generated container build file; and(b3) registering the generated container image in a repository and activating the container image to be usable in association with the workflow management module.8.The method according to any one of claims 1 to 7,wherein the plurality of workflow management modules includes a pipeline management module,and the step (b) comprises:(b3) generating a pipeline configuration file defining each stage of machine learning execution as an independent container; and(b4) configuring management logic to monitor system resources allocated at a completion time of pipeline tasks to be executed in the future and dynamically release unnecessary resources.9.The method according to any one of claims 1 to 8,wherein the plurality of machine learning execution modules includes a model optimization module configured to automatically optimize hyperparameters for improving performance of an artificial intelligence model,and the step (b) comprises:(b5) installing a plurality of optimization modules for searching for an optimal combination of hyperparameters based on at least one of a plurality of optimization algorithms; and(b6) configuring the model optimization module to be executed in a distributed search environment in association with the workflow management module.10.The method according to any one of claims 1 to 9,wherein the plurality of machine learning execution modules includes a lifecycle experiment and tracking management module for managing experiments and tracking across an entire lifecycle of a model,and the step (b) comprises:(b7) constructing a base environment having reproducible and consistent characteristics based on library configuration information and version information for lifecycle experiment and tracking management of the model; and(b8) configuring an API interface between a specific artificial intelligence model and the base environment so that experiment-related data of the specific artificial intelligence model is automatically transmitted to and recorded in the base environment.11.The method according to any one of claims 1 to 10,wherein the plurality of machine learning execution modules includes a model deployment and serving module for deploying and servicing a trained model,and the step (b) comprises:(b9) configuring a service configuration module for defining an application service specification including internal dependencies and execution paths of the trained model;(b10) configuring an execution environment configuration module for defining an infrastructure execution environment for operating the model for which the service specification is defined in a containerized form; and(b11) configuring, in association with the execution environment, a service module including interface logic for providing the containerized model as an online-based service or a local-based service.12.The method according to claim 11,further comprising:(b12) configuring a mirroring environment of an object storage based on real-time detected available storage capacity information so that data processing in response to inference requests is performed without interruption, thereby ensuring data availability.13.The method according to any one of claims 1 to 12,wherein the method is performed by executing a single file including a script for performing all processes including the step (a) and the step (b), such that the entire MLOps system environment is automatically constructed without user intervention.14.An apparatus for constructing an MLOps (Machine Learning Operations) system environment, comprising:an interface unit configured to support transmitting an environment construction command to a system; anda processor configured to detect available resources of a target system or support detection thereof, dynamically configure a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications, and automatically install and configure at least one of a plurality of machine learning execution modules and workflow management modules on the configured container orchestration cluster in association with the resource configuration of the cluster.15.A non-transitory computer-readable recording medium storing computer-executable instructions for constructing an MLOps (Machine Learning Operations) system environment,wherein the instructions cause a computer to perform:(a) detecting available resources of a target system, and dynamically configuring a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications; and(b) automatically installing and configuring at least one of a plurality of machine learning execution modules and workflow modules on the configured container orchestration cluster in association with the resource configuration of the cluster.16.An MLOps (Machine Learning Operations) system, comprising:an interface unit configured to receive an environment construction command; anda processor configured to detect available resources of the system or support detection thereof, dynamically configure a container orchestration cluster in which a plurality of nodes are logically grouped based on the detected available resources so as to be optimized for system specifications, and automatically install and configure at least one of a plurality of machine learning execution modules and workflow management modules on the configured container orchestration cluster in association with the resource configuration of the cluster.