CN113127441B - A method for dynamically selecting database components and a self-assembling database management system - Google Patents

Abstract

The embodiment of the application discloses a method for dynamically selecting components in a database management system, which can be applied to a self-assembled database management system comprising a plurality of micro-service components. The method comprises the steps of vectorizing information related to query to obtain vectorized information input, inputting the vectorized information into a reinforcement learning model to obtain a first target component, and then complementing the first target component into an execution path by using a tree search algorithm such as a Monte Carlo Tree Search (MCTS) algorithm, wherein the execution path comprises the first target component and a second target component, and the first target component and the second target component respectively realize different functions. According to the scheme provided by the embodiment of the application, the corresponding micro-service component can be automatically selected to respond to the query according to the query information and the state information of the database, so that different service scenes can be flexibly dealt with, and the execution efficiency is improved.

Description

A method for dynamically selecting database components and a self-assembling database management system

Technical Field

The embodiments of the present application relate to the field of databases, and in particular to a method for dynamically selecting components in a database management system and a self-assembling database management system.

Background Art

Database management system (DBMS), also commonly referred to as database, is the core of many application systems, used to provide data operations, permission control, data persistence, concurrency control, crash recovery and other capabilities for applications. The traditional database management system architecture is designed for general-purpose processors, and the core components include structured query language (SQL) parser, SQL optimizer, SQL executor and storage engine. Traditional database management systems divide the functions of components in the logical architecture, but they are tightly coupled physically. That is, the database management system is developed as a whole and deployed on specific hardware devices to implement data maintenance and management functions.

In the prior art, the connection relationship between various components in database management is set during development, so that the components passed by the database management system when executing the query instructions input by the user are fixed.

Since the combination of components is fixed, it cannot flexibly respond to different business scenarios.

Summary of the invention

The embodiments of the present application provide a method for dynamically selecting components in a database management system and a self-assembling database management system, which can automatically select corresponding components to run query information according to query information and the status of the database management system, can flexibly respond to different business scenarios, and improve execution efficiency.

The first aspect of the embodiments of the present application provides a method for dynamically selecting components in a database management system, which can be implemented by a database management system including multiple microservice components, or by a device including a database management system, such as a database server, or a virtual machine that allows implementation on a server. The method includes: obtaining query information and status information of the database management system, and the status information may include load information of multiple components in the database management system. The interaction parameters of the first target component and the first target component are obtained according to the reinforcement learning model and the vectorized information. The vectorized information can be obtained after the query information and the status information are vectorized. The interaction parameters may include parameters that need to be transmitted between the first target component and other components. The self-assembly manager can determine the first path based on the first target component using a search algorithm, and the self-assembly manager can respond to the query information based on the first path. The first path includes at least a first target component and a second target component, and the functions of the first target component and the second target component are different.

In an embodiment of the present application, a first target component can be obtained based on a reinforcement learning model and vectorized information, and then a search algorithm, such as a tree search algorithm, can be used to select one or more second target components from a database management system to obtain a first path, so that the database management system responds to query information according to the first path. The database management system can automatically select a corresponding component to run the query information based on the query information and the status information of the database management system, which can flexibly respond to different business scenarios and improve execution efficiency.

In a possible implementation, a reinforcement learning model is pre-trained, and the vectorized information is used as input of the reinforcement learning model to obtain the interaction parameters between the first target component and the second target component, thereby improving the feasibility of the solution.

In one possible implementation, the first target component and the interaction parameter of the first target component can be obtained according to the reinforcement learning model, the generative adversarial network GAN and the vectorized information. In an embodiment of the present application, the generative adversarial network can also be used to optimize the first target component selected according to the reinforcement learning model and the second target component selected according to the search algorithm.

In one possible implementation, an element in the reinforcement learning model can be replaced with a GAN generator to obtain a new reinforcement learning model, and then vectorized information is input into the new reinforcement learning model to obtain the first target component and the interaction parameter of the first target component. This improves the feasibility of the solution.

In a possible implementation, the search algorithm is a Monte Carlo tree search algorithm, which improves the feasibility of the solution.

In a possible implementation, after searching the database management system based on the search algorithm and selecting the second target component, the scoring information of the GAN discriminator on the first path can be obtained, and the scoring information is related to the performance parameters of the first path. If the score of the first target component does not meet the preset conditions, the score of the first target component can be fed back to the generator, so that the generator processes the first target component to obtain the processed first target component, and repeats the aforementioned search for the second target component and obtaining the scoring information based on the processed first target component. If the score meets the preset conditions, the optimized component is output, and the target state information is obtained. The optimized component is the first target component whose score meets the preset conditions. The target state information can be obtained according to the reinforcement learning model, the optimized component and the state information, and the target state information includes the determined component information and the query state information. If the number of the optimization components does not meet the preset number, the target state information is used as the state information to repeat the aforementioned steps of obtaining the first target component, searching for the second target component, and obtaining the scoring information. If the number of the optimization components meets the preset number, the second path is output, and the score of the first target component is obtained after processing the scoring information; the second path includes at least two optimization components, and each optimization component has a different function; according to the second path, at least two optimization components are triggered to respond to the query information.

In an embodiment of the present application, a first target component is obtained based on a reinforcement learning model and vectorized information, and then a second target component is selected using a search algorithm to obtain a first path, and an optimized component that meets a preset condition in the first target component is selected, and then the target state information is used as the state information to repeat the aforementioned steps of obtaining the first target component, searching for the second target component, and obtaining the scoring information until the number of optimized components meets a preset number, and a second path is output, and query information is responded to based on the optimized component in the second path. The optimized component that meets the preset condition can be automatically selected to execute the operation to respond to the query information based on the query information and the state information of the database, which can flexibly respond to different business scenarios and improve execution efficiency.

A second aspect of an embodiment of the present application provides a database management system, comprising multiple database components, each database component being used to implement a specific function. The database management system also includes one or more functional units for implementing the method of the aforementioned first aspect or any possible implementation method of the first aspect.

A second aspect of an embodiment of the present application provides a self-assembly manager, which includes multiple functional units and is used to implement the method of the aforementioned first aspect or any possible implementation manner of the first aspect.

A third aspect of an embodiment of the present application provides a self-assembly manager, which includes a processor and a memory, wherein the memory stores an executable program or instruction, and the processor allows the program or instruction to be executed to implement the method of the aforementioned first aspect or any possible implementation method of the first aspect.

A fourth aspect of an embodiment of the present application provides a computer storage medium, which stores instructions. When the instructions are executed on a computer, the computer executes the method of the first aspect or any possible implementation method of the first aspect.

A fifth aspect of an embodiment of the present application provides a computer software product. When the computer program product is executed on a computer, it enables the computer to execute the method of the first aspect or any possible implementation manner of the first aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1A is an architecture diagram of a database system provided in an embodiment of the present application;

FIG1B is a schematic diagram of the structure of a database system based on decoupling of microservice components in an embodiment of the present application;

FIG2 is a basic model of reinforcement learning in an embodiment of the present application;

3 is a schematic diagram of the structural relationship between the self-assembly manager and the database in an embodiment of the present application;

FIG4 is a flow chart of the RL+GAN+search algorithm in an embodiment of the present application;

FIG5 is a flow chart of a method for dynamically selecting a database component in an embodiment of the present application;

FIG6 is another flow chart of the method for dynamically selecting a database component in an embodiment of the present application;

FIG7 is an implementation scenario of the method for dynamically selecting a database component in an embodiment of the present application;

FIG8 is another schematic diagram of a flow chart of a method for dynamically selecting a database component in an embodiment of the present application;

FIG9 is another schematic diagram of a flow chart of a method for dynamically selecting a database component in an embodiment of the present application;

FIG10 is a schematic diagram of a structure of a self-assembly manager in an embodiment of the present application;

FIG11 is another schematic diagram of the structure of the self-assembly manager in an embodiment of the present application;

FIG12 is another schematic diagram of the structure of the self-assembly manager in an embodiment of the present application;

FIG13 is a schematic diagram of a database system in an embodiment of the present application;

FIG. 14 is another schematic diagram of a database system in an embodiment of the present application.

DETAILED DESCRIPTION

The embodiment of the present application provides a method and related equipment for dynamically selecting components in a database management system, which are used to automatically select components to generate a path sequence.

The method provided in the embodiment of the present application can be applied to a database system. FIG1A shows the architecture of a database system provided in the embodiment of the present application. According to FIG1A , the database system 10 includes a database management system (DBMS) 100 , a database 101 and a data storage device (DataStorage) 103 .

Database 101 is an organized data set stored in data storage 103, that is, a collection of associated data organized, stored and used according to a specific data model. According to the different data models used to organize data, data can be divided into multiple types, such as relational data, graph data, time series data, etc. Relational data is data modeled using a relational model, usually represented as a table, and the rows in the table represent a set of related values of an object or entity. Graph data, referred to as "graph", is used to represent the relationship between objects or entities, such as social relationships. Time series data, referred to as time series data, is a data column recorded and indexed in chronological order, used to describe the state change information of an object in the time dimension.

The database management system 100 is the core of the database system and is a system software for organizing, storing and maintaining data. The client 30 can establish a communication connection with the database management system 100 via the network 20 and can access the database 101 through the database management system 100. The database administrator (Database Administrator, DBA) also performs maintenance work on the database 101 through the database management system 100. The database management system 100 provides multiple functions for the client 30 to establish, modify and query the database 101, wherein the client 30 can be an application or a user device. The functions provided by the database management system 100 may include but are not limited to the following: (1) Data definition function. The database management system 100 provides a data definition language (DDL) to define the structure of the database 101. DDL is used to describe the database framework and can be saved in the data dictionary. (2) Data access function. The database management system 100 provides a data manipulation language (DML) to implement basic access operations on the database 101, such as retrieval, insertion, modification and deletion. (3) Database operation management function. The database management system 100 provides a data control function to effectively control and manage the operation of the database 101 to ensure that the data is correct and valid. (4) Database establishment and maintenance function, including loading of initial database data, database dumping, recovery, reorganization, system performance monitoring, analysis and other functions. (5) Database transmission. The database management system 100 provides transmission of processed data to achieve communication between the client 30 and the database management system 100, which is usually coordinated with the operating system.

The operation of the database management system 100 depends on the necessary software and hardware environment, including but not limited to the hardware layer 104 and the operating system 102. The hardware layer 104 includes the basic hardware units required for the operation of the operating system 102 and the database management system 100, such as processors, memory, input/output (I/O) devices, network interface controllers (NIC), etc. The operating system 102 is the system software that manages the hardware units and can provide functions such as memory management and thread scheduling.

The data storage device 103 can be a non-transient computer-readable storage medium such as a hard disk, a disk, a storage array, a storage server, a cloud storage, a storage area network (SAN), etc., which is communicatively connected to the computing node where the hardware layer 104 is located. Alternatively, the data storage device 103 can also be integrated in the computing node where the hardware layer 104 is located, and exchange data with the processor and I/O devices through a bus or other internal communication methods. It should be noted that the "computing node" in the embodiments of the present application refers to an entity that has the hardware resources required to perform data calculation and/or storage, such as a physical machine or a database server, or an entity that can call hardware resources for calculation and/or storage, such as a virtual machine (VM) or container deployed in a physical machine.

In one embodiment, the functions of the database management system 100 can be implemented by a processor executing executable codes stored in a memory. It should be understood that in various embodiments of the present invention, "executable program" should be broadly interpreted as including but not limited to: instructions, instruction sets, codes, code segments, subroutines, software modules, applications, software packages, threads, processes, functions, firmware, middleware, etc. In one embodiment, the database management system 100 can be software running on a server or a virtual machine, and the code or instructions corresponding to the software are stored in a memory, and when executed by at least one processor, the corresponding functions are implemented.

Those skilled in the art may understand that a database system may include fewer or more components than those shown in FIG. 1A , or include components different from those shown in FIG. 1A , and FIG. 1A merely shows components that are more relevant to the implementation method disclosed in the embodiment of the present invention.

FIG1B shows a logical architecture of a database management system 100 according to an embodiment of the present application. According to FIG1B , the database management system 100 includes:

Application program interface 110, job manager 120, component manager 130, message communication service 140, operating system 150, and memory 103. Further, database management system 100 also includes multiple database components, which are microservice components, and each microservice component is used to implement a specific function of the database management system. In one embodiment, the database components included in database management system 100 include: parser service 160, optimizer service 170, parser and optimizer combination service 165, executor service 180, storage engine service 182, metadata service 184, statistics service 186, self-monitoring service 188, clock service 190. The functions of each component of database management system 100 are further described below:

The application program interface 110 is an interface provided to the client to connect to and access the database management system 100, such as a Java Database Connectivity (JDBC) interface, an Open Database Connectivity (ODBC) interface, and the like.

The job manager 120 includes a scheduler 121, which selects the optimal execution path from the execution path provided by the component manager 231 according to the user request from the client, such as a query, and schedules the service component based on the optimal execution path for corresponding processing. When an error occurs in a microservice component in the execution path, the scheduler 121 schedules the backup execution path to continue to execute the current user request, thereby improving the availability of database services. The job manager 120 itself can also embed service components such as a parser 122, an optimizer 123, an executor 124, and a storage engine 125. The scheduler 121 can call the service components embedded in the job manager 120, or call service components outside the job manager 120, such as a parser and optimizer combined service 165, an executor service 180, etc. When calling external service components, a remote procedure call (RPC) method can be used.

The component manager 130 is mainly responsible for the management of microservice components, including but not limited to: component registration and deregistration, component startup and shutdown, component fault tolerance, component upgrade, component status monitoring, etc. (1) Component registration and deregistration: When adding a new service component, it is necessary to register with the component manager 130. When deleting a service component, it is also necessary to deregister with the component manager 130. The component manager 130 can maintain relevant metadata to record the status, deployment location and other information of each service component. This ensures that the component manager 130 understands the available service components in the system and the deployment location of each service component. (2) Component startup and shutdown: When the system starts, the component manager 130 starts service components such as metadata service, self-monitoring service, and statistics service according to the hardware configuration; in a distributed scenario, each service component can start multiple copies (the specific number is configurable) for backup. (3) Component fault tolerance: When a service component fails, the component manager 130 needs to schedule a backup service component to continue executing the current assembly plan. (4) Component upgrade: Components can be upgraded in rotation. For stateless components, they can be upgraded at any time; for stateful components, they must be upgraded when the current task is completed. (5) Component status monitoring: monitor the running status of service components during operation. When a node is added or exited in the database system, the running status needs to be updated. The running status includes component service status and component cluster status. Component service status refers to the number of tasks executed by each running service component and resource usage. Component cluster status refers to the service components running on each computing node, the available service components, and the resource usage of each computing node.

The message communication service 140 provides low-latency and high-bandwidth communication capabilities and can be used for communication between microservice components.

Parser service 160: the main function is to perform lexical analysis, syntax analysis, and semantic analysis on the input SQL statement and output the query parse tree;

Optimizer Service 170: Its main function is to process the input query parsing tree to generate an execution plan. Its processing logic includes query rewriting, path generation, cost model evaluation, optimal path selection, and execution plan tree generation.

Executer Service 180: Responsible for reading data from the storage engine, processing the data according to the execution plan, and returning it to the client;

Storage Engine Service 182: Its main function is to ensure persistent storage of data, provide efficient data access capabilities, and implement the atomicity, consistency, isolation, and durability (ACID) capabilities of the database.

Metadata Service 184: Mainly provides persistent storage of metadata and efficient metadata access capabilities;

Statistics Service 186: Mainly responsible for collecting and storing table statistics, including the number of pages, number of rows, and value distribution information of each column.

Instrumentation Service 188: monitors various status data of the database, such as hardware resource usage information (CPU, I/O, memory, network), key performance indicators (KPI) of each service component of the database, etc.

Clock Service 190: Provides unique and incremental timestamps with an error usually within 100ns.

The main functions of the self-assembly manager composed of the job manager 120 and the component manager 130 may be to be responsible for starting, scheduling, and managing component services, running the self-assembly algorithm, and realizing the self-assembly of database components.

In the embodiment of the present application, the component manager 130 can also determine one or more service component execution paths (hereinafter referred to as "execution paths") based on system resources, such as generating TOP-N optimal component execution paths. Each component execution path indicates multiple microservice components that are executed in order. These ordered microservice components constitute a complete database management system that can be executed end-to-end. Each microservice component on the execution path can be deployed on different computing nodes.

In the embodiment of the present application, the component manager 130 may include one component manager or multiple component manager clients, each of which is deployed in a single computing node (physical machine or virtual machine) to implement functions such as component registration, component deregistration, component start and stop, component resource management, and component upgrade on the node. In the entire database management system, there may be multiple component managers as backups, for example, three.

Microservice components can be started and run independently. For example, each microservice component can run as one or more instances, where the instances can be threads. Service components communicate with each other using lightweight communication mechanisms. In one embodiment, multiple microservice components can also be combined into a larger granularity combined service component, such as the parser and optimizer combined service 165 shown in FIG. 1B, which is a combined service component that includes both the parser service and the optimizer service functions.

Microservice components can be independently deployed to bare metal or containers and support online upgrades. Users or database kernels can flexibly match and assemble microservice components to build different types of database management systems based on current business scenarios.

In one embodiment, the microservice components of the database management system 100 can be distributed and deployed on computing nodes in a data center, or deployed across data centers. The computing nodes in the data center are interconnected through a high-speed network, for example, through an InfiniBand network. The computing nodes are physical machines, or other devices with data processing capabilities. Each computing node contains the basic hardware resources required for the operation of the operating system and application programs, such as processors, memory, input/output (I/O) devices, etc. One or more virtual machines (VMs) run on a computing node, and one or more microservice components can be deployed on each VM. Each microservice component can be deployed on a separate VM, but it should be understood that multiple microservice components can also be deployed on the same VM.

In one embodiment, the microservice components of the database management system 100 can also be deployed in the form of an integrated machine. One or more microservice components are deployed on each physical machine, and multiple physical machines can be integrated into one cabinet and communicate through a bus, such as a PCI-E bus.

In one embodiment, the microservice component can be dynamically loaded in the form of a dynamic link library. For example, during the execution of a microservice component, one or more other microservice components in the form of a dynamic link library can be dynamically loaded.

In the embodiment of the present application, the number of the above components is only schematically illustrated by taking the number in Figure 1 as an example. It can be understood that each component can be set to a different number as needed, and the number of each component is not specifically limited here. For example, the figure shows 2 parser services, 2 optimizer services, 3 executor services and 3 storage services, while the number of metadata services, statistics services and accelerator services is 1. The database management system 100 may also include multiple types of microservice components with the same functions. Taking the storage service 180 as an example, Figure 1B shows three types of storage engines: column storage engine, row storage engine and memory engine. For another example, the database management system 100 can simultaneously include multiple types of parser service components, such as a parser for parsing relational queries, a parser for parsing graph queries, a parser for parsing time series queries, etc. The number of each type of microservice components can be one or more. The database management system 100 can automatically select multiple microservice components to assemble into an ordered component sequence based on the information of the query submitted by the client, such as the query type, and the system load, such as the load of each microservice component. The ordered component is a path, and the microservice components indicated by the path are executed in order, which can realize the function of the database management system and improve the availability and maintainability of the entire system. Therefore, the core capability of the database management system 100 is to dynamically select the optimal path (component sequence) and call or trigger the component indicated by the path to perform a response operation to respond to the query. In one embodiment, the selection of the optimal path and the calling component execution operation can be implemented by the component manager 130 and the job manager 120 respectively. In one embodiment, the job manager 120 and the component manager 130 can also be integrated together to form a large functional unit, such as a self-assembly manager. The self-assembly manager can also be used as a service component of the database management system 100, independent of the microservice components involved in the execution process, and is the input to guide the execution process. By obtaining the result of the database execution as feedback, the self-model is updated by continuous reinforcement learning, so the self-assembly manager is placed outside as described in Figure 3.

After receiving the query instruction, the database management system 100 specifically determines how to match and assemble the various service components in the database management system 100, that is, which components are selected in what order (path) to respond to the query instruction. The optimal path (component sequence consisting of multiple specific components that respond to the query instruction) to respond to the query instruction can be selected through reinforcement learning (RL), search algorithms, and generative adversarial networks (GAN). The following first describes the machine learning model used in the dynamic selection of database components.

FIG2 shows an example of a reinforcement learning model for selecting an optimal path according to an embodiment of the present application:

Reinforcement learning is an important branch of machine learning and a product of the intersection of multiple disciplines. It emphasizes how to act based on the environment to maximize expected benefits.

The basic models of reinforcement learning include:

Agent, current state, action, environment, and reward.

Among them, the agent is the core of the basic model. It can perceive the state and select a suitable action through learning based on the reinforcement signals provided by the environment to maximize the long-term reward.

State is the environment information of the agent, which contains all the information used by the agent to select actions.

The Environment receives the actions executed by the agent, evaluates the quality of the actions, and converts them into a quantifiable reward and feeds them back to the agent.

Reward is a quantifiable scalar feedback signal from the Environment to the agent, which is used to evaluate the quality of the action taken by the agent.

The agent executes action under the state and interacts with the environment. The environment gives the next state state+1 and reward to the agent for learning. Through continuous contact with the environment, the agent learns the rules from the environment and thus learns the optimal solution.

The above briefly introduces the database management system and reinforcement learning model of the embodiment of the present application. The following describes the self-assembly algorithm (RL+GAN+search algorithm) proposed in the embodiment of the present application:

Please refer to FIG4 , the RL+GAN+search algorithm process in the embodiment of the present application:

The execution path of the database is divided into multiple states. With the iterative idea of reinforcement learning as the core, the generator of the generative adversarial network (GAN) is used as the intelligent agent to generate an execution action for the current state state to reach the next state state'. The search algorithm is used to randomly complete each action into an execution path. The discriminator of GAN is then used to score each path and feed it back to the generator to promote its iteration and generate better actions until the state ends. Finally, a complete execution path is generated for the database. The actual running system of the database is used as the environment, and the execution results are fed back to the generator and the discriminator to promote network optimization.

In the embodiment of the present application, the search algorithm is schematically illustrated as a Monte Carlo tree search (MCTS) algorithm. It can be understood that in practical applications, the search algorithm can also be other algorithms. For example, the search algorithm can be an upper confidence bound apply to tree (UCT) algorithm or an upper confidence bound (UCB) algorithm, which is not limited here.

In the embodiment of the present application, with the iterative idea of reinforcement learning as the core, the GAN generator replaces an element in the reinforcement learning model, generates a first target component for the current query information to reach the next target state information, uses a search algorithm to search the components in the database and selects the second target component to obtain the first path, and then uses the GAN discriminator to score each first path, and feeds back to the generator to promote the iteration of the generator to obtain the optimized component until the second path is generated, and finally generates a complete path for the database to respond to the query information.

Please refer to FIG5 . Based on the database management system shown in FIG1B , the method for dynamically selecting a database component in the embodiment of the present application includes:

501. The self-assembly manager obtains query information and status information of the database management system;

The self-assembly manager in the embodiment of the present application is the self-assembly manager described in the embodiment corresponding to FIG. 1B .

In the embodiment of the present application, the self-assembly manager is only schematically illustrated as being integrated by the component manager and the job manager. It is understandable that the operations performed by the self-assembly manager can also be performed by the component manager and the job manager in cooperation with each other.

In the embodiment of the present application, the self-assembly manager receives query information through an application program interface, and the query information can be a query statement described in a specific language, such as an SQL statement or instruction, which is used to implement operations on data, such as adding, modifying, deleting, querying, etc. It can be understood that the query information can also be a configuration file or other file, which is not specifically limited here.

In the embodiments of the present application, the query information may be input by a user through a client or other device, may be configured by a user through a file, or may be generated by other systems, which are not specifically limited here.

For the convenience of description, it is illustrated in conjunction with Figure 6: Figure 6 only uses the database status information including database status, hardware information, network information and GUC parameters as an example for schematic illustration. It can be understood that in actual applications, the database status information may include at least one of the database status, hardware information, network information and GUC parameters.

The status information of the database management system may include the load information of multiple components in the database management system, and may also include the load information of multiple components in the database management system and the performance parameters of the database management system (which may include memory, CPU and other loads or SQL execution efficiency). The number of components is set according to actual needs and may be all components in the database management system, which is not limited here.

In the embodiment of the present application, the target state information may include four query states: SQL, syntax tree, execution plan, and data, wherein the SQL query state indicates that the self-assembly manager is obtaining the query instruction input by the user. The syntax tree query state indicates that the self-assembly manager is parsing the query instruction to obtain the natural word order. The execution plan query state indicates that the self-assembly manager is executing the natural word order. The data query state indicates that the database obtains the data in response to the query instruction through the natural word order.

502. The self-assembly manager replaces an element in the reinforcement learning model with the generator of the GAN to obtain a new reinforcement learning model;

Utilizing the scoring-feedback-optimization between the generator and the discriminator in GAN, as shown in Figure 6: the self-assembly manager replaces the agent in the reinforcement learning model with the generator of GAN to obtain a new reinforcement learning model.

503. The self-assembly manager inputs the vectorized information into the new reinforcement learning model to obtain the first target component and the interaction parameter of the first target component;

Since the assembly of components is sequential, the query information and the database status information need to be vectorized to obtain vectorized information. The self-assembly manager inputs the vectorized information into the new reinforcement learning model to obtain the interaction parameters between the first target component and the first target component, as shown in Figure 6: The query information input by the user is combined with the database status and load information, processed through vectorization and passed to the generator. The generator generates an action (selecting the interaction parameters between the first target component and the first target component) according to the current state and query request.

In one embodiment, when the query information is an SQL statement, it can be converted into a natural language sequence, and the interaction parameters include parameters that need to be transmitted between components. The interaction parameters can also include a service address, which is not limited here. The service address can be used to implement a call through a remote procedure call (RPC) at a remote end point, and the interaction parameters can include a function call interface between components. The components communicate through a message communication service component.

504. The self-assembly manager determines a first path using a search algorithm according to the first target component;

After the self-assembly manager determines the first target component, the path is not yet complete. Assume that the full path is 0-10. At this time, there is only an incomplete sequence 0-5, and many optional components {6.0, 6.1, 6.2, ...}. Temporarily, each of these components is used as the current state {0-5-6.x}, and then the following components {0-5-6.x-10} are randomly added to the first target component based on empirical data.

The self-assembly manager searches the components in the database based on the search algorithm and selects the second target component to determine the first path. During testing or actual runtime, the database can respond to query information according to the first path, and the first path includes the first target component and the interaction parameters of the first target component and at least one second target component other than the first target component.

Optionally, after obtaining the first path, the first target component and the second target component in the first path may be formed into a plan list and returned to the database so that the first target component and the second target component can perform operations in response to the query information.

In one embodiment, as shown in FIG6 , the MCTS algorithm is used to generate a first path by combining the currently generated action (the first target component and the interaction parameter of the first target component) with the random completion path of the historical state, and the historical state includes the components currently determined in the database.

505. The self-assembly manager obtains the scoring information of the first path by the discriminator of the GAN;

By utilizing the scoring-feedback-optimization between the generator and the discriminator in the GAN, the discriminator scores the first path according to the performance parameters of the first path, and the self-assembly manager obtains the scoring information of the discriminator on the first path. The performance parameters of the first path may be the efficiency of the first path in executing SQL statements.

506. If the score of the first target component does not meet the preset conditions, the self-assembly manager feeds back the score of the first target component to the generator so that the generator processes the first target component and uses the processed first target component to execute steps 504 to 506. If the score of the first target component meets the preset conditions, the optimized component is output and step 507 is executed.

The user can set preset conditions as needed. For example, the preset condition can be that the time to execute an SQL statement is 5 microseconds. If the time for the first target component to execute the same SQL statement is 6 microseconds, the self-assembly manager will feed back the score of the first target component to the generator. The generator will optimize according to the score of the discriminator until the time for the first target component to execute the same SQL statement is less than 5 microseconds. Then, the optimized component is output and step 507 is executed. The optimized component is the first target component whose score meets the preset condition. The score of the first target component is obtained by processing the scoring information of the first path.

In the embodiment of the present application, the score of the first target component can be obtained by taking the average of the score information of multiple first paths, or taking the maximum value, which is not limited here.

In one embodiment, as shown in FIG6 : after the discriminator scores the first path, the average value is taken as the score of the action (target action), and if the score meets the preset conditions, the action (optimization component) is output.

507. The self-assembly manager obtains target state information according to the reinforcement learning model, the optimization component and the state information;

It can be seen from the reinforcement learning model that when the first target component is selected (action), the state information of the database will change (determined components) and become the target state information. The target state information includes the currently determined components, and the target state information can change according to the number of determined components.

The query status information in the database status information may include the number of currently selected components and information of the current data flow.

508. If the number of the optimized components does not meet the preset number, the self-assembly manager uses the target state information as the state information to execute steps 502 to 506. If the number of the optimized components meets the preset number, the second path is output.

If the number of the optimized components does not meet the preset number, the self-assembly manager updates the state information, that is, executes steps 502 to 506 using the target state information as the state information, and continues to select the optimized components in the next layer of components. If the number of the optimized components meets the preset number, outputs a second path, which includes at least two optimized components, and each optimized component has a different function.

The preset number in the embodiment of the present application is related to the number of components required to respond to the query information in the database management system, and is not specifically limited here. For example, the preset number can be the number of components required at least to respond to the query information to obtain the query results, and these components are necessary to respond to the query information.

Exemplarily, there are 10 components that respond to query information in the database management system, and the preset number is 10, that is, the full path is 0-10 (10 layers). Each time steps 502 to 506 are looped, an optimization component of one layer is selected. If the optimization components in each of the 10 layers are selected, the second path is output. The second path includes 10 optimization components, and the function of each optimization component is different.

509. The self-assembly manager triggers at least two preferred components to perform operations in response to the query information according to the second path.

After outputting the second path, the self-assembly manager triggers at least two optimization components in the second path to perform operations in response to the query information. Exemplarily, if the second path includes 10 optimization components, the self-assembly manager triggers the 10 optimization components to perform operations in response to the query information. The specific operations performed will be described later in conjunction with FIG. 8 and FIG. 9 for actual situations.

The following describes different scenarios in the embodiment of the present application in conjunction with FIG. 7:

1. Offline learning during testing:

During the database testing phase, according to various typical scenarios, it is necessary to vectorize queries, loads, and data so that the assembleable sequences converge into TOP-10 paths (10 second paths), and use optimal control theory to achieve initial self-assembly. That is, for each scenario (such as finance, cloud), select the TOP-10 assembly solutions and recommend them to customers.

2. Deployment and learning at the beginning:

When the database is launched, the top-3 paths (3 second paths) are learned and converged again in the top-10 assembly solution sequence based on the user's real data and scenarios. The main difference is that it is constrained learning, and there is no need to consider solutions outside the top-10, which greatly improves the speed of learning convergence.

3. Online learning at runtime:

When the database is running, the Top-3 assembly schemes are rotated online according to the real-time changes in user load and data, and fine-tuned in the Top-3 assembly schemes; after clarifying the assembly process and basis, the existing execution plan is continuously fine-tuned to achieve the optimal one (1 second path).

There may be many implementation scenarios for the method of dynamically selecting a database component in the embodiment of the present application. The above three are only used as examples for schematic illustration and are not specifically limited here.

For ease of understanding, the method for dynamically selecting a database component in an embodiment of the present application is described below in conjunction with FIG. 8 and FIG. 9 :

Please refer to FIG8 , which is an embodiment of a method for dynamically selecting a database component in an embodiment of the present application:

Customer services have traffic peaks and valleys, and the load sizes of peaks and valleys are different. Using the method of dynamically selecting database components in the embodiment of the present application, microservice components can be automatically assembled and adjusted according to the business load to meet the business load of peak events.

The embodiment of the present application can perform a more fine-grained division for the execution service module, including row execution and vector execution, and the storage service can be divided into row storage, column storage and memory engine.

The public cloud/private cloud is installed with a deployment database service corresponding to the deployment learning shown in FIG. 7 , and the algorithm in the component manager is trained to converge to the TOP-3 path as described in the application scenario of the second path above.

The job manager receives a large number of query requests.

Since processing may be limited by computing power, network transmission capacity, etc. during peak traffic, it is necessary to determine which factor has become the bottleneck and limited performance. Assumption: Based on the service monitoring information, the component manager determines that the current cluster computing capacity is the bottleneck point.

The component manager generates a path using idle nodes based on the algorithm shown in FIG4 and the current database component status (including the utilization of each node and the hardware configuration of each node).

The component manager schedules idle nodes according to the algorithm shown in FIG4 and starts the SQL execution service and the GPU acceleration service.

That is, the service components are distributedly deployed on multiple computing nodes, such as computing nodes 801-806 in FIG. 8 , and one or more service components are deployed on each computing node, and a component manager client 832 is deployed to manage the service components on the node.

For example, computing node 801 includes component manager client 832 and SQL parser and optimizer service 862. Computing node 802 includes component manager client 832 and SQL execution service 880, wherein SQL execution service 880 includes row execution and vector execution. Computing node 804 includes component manager client 832 and SQL execution service 880 and self-monitoring service 888. Computing node 805 includes component manager client 832 and storage service 882.

The component manager client 832 and the SQL executor service 880 may also constitute a computing node 803. The component manager client 832 and the storage service 882 may also constitute a computing node 806. The services included in the computing node are not specifically limited here.

The SQL executor service 880 may be a row execution or a vector execution, and the storage service 882 may be a column storage engine or a row storage engine.

The job manager 820 receives multiple query requests, and according to the execution path determined by the component manager 831, calls the corresponding service components on each computing node through the component manager client 832 to perform business processing. When the business peak arrives, if the component manager 831 determines that the current system calculation is a bottleneck point based on the service monitoring information, then based on the hardware configuration of the idle computing node 104 (here it is assumed that the idle node includes a GPU), it determines the microservice components that need to be enabled on the idle node, including the SQL executor service 880 and the GPU acceleration service 898. The component manager 831 requests the component manager client 832 on the idle node 104 to start the SQL executor service and the GPU acceleration service. The idle node 104 starts the SQL executor service 880 and the GPU acceleration service 898 to cope with the business peak. The idle computing node 804 can be a node whose load is lower than the preset value, or it can be a newly added node.

In this embodiment, the database can automatically respond to business load bursts and the like by installing the method of dynamically selecting database components in the embodiment of the present application without manual intervention.

Please refer to FIG. 9 , another embodiment of the method for dynamically selecting a database component in an embodiment of the present application:

The database shown in FIG. 9 includes a query parsing service, component 90 , a query optimization service, component 91 , an execution engine service, component 92 , a message communication service, component 93 , and a storage engine service, component 94 .

Among them, the query parsing service and component 90 include parsers 901 to 906; the query optimization service and component 91 include optimizers 911 to 916; the execution engine service and component 92 include executors 921 to 926; the message communication service and component 93 include transmission control protocol (TCP) 931, stream control transmission protocol (SCTP) 932, remote direct memory access (RDMA) 933; the storage engine service and component 94 include memory (row memory) 941 and 944, memory (column memory) 942 and 945, memory 943 and 946.

The user sends a request 1 to the database, which requires data interaction between multiple nodes of the distributed database and analysis after aggregation.

The component manager receives request 1, runs the algorithm of Figure 4 to generate the optimal execution path (first path) based on the current database component status and request characteristics, selects the TCP service in the message communication service component, and establishes a long connection for request 1.

After request 1 is responded to by other nodes, data aggregation and migration are required, and the database generates internal message communication request 2 to request data transmission from other nodes.

The component manager receives request 2, runs an algorithm to generate the optimal execution path (the second path) based on the current database component status and request characteristics, selects the RDMA service in the message communication service component (RDMA is a technology for high-speed communication between nodes that can provide high-speed and high-bandwidth communication services), and performs large-scale data transmission.

In the embodiment of the present application, TCP or RDMA, row memory or column memory can be determined by request characteristics.

In this embodiment, the communication service is flexibly selected according to the different characteristics of the request, and the user request interaction is maintained by establishing a long connection. Data is transmitted through a high-speed network, thereby avoiding possible input and output IO congestion of the communication network.

The above describes the method for dynamically selecting a database component in the embodiment of the present application. The following describes the self-assembly manager in the embodiment of the present application. Please refer to FIG. 10. An embodiment of the self-assembly manager in the embodiment of the present application includes:

The acquisition unit 1001 is used to acquire query information and status information of the database management system, where the status information includes load information of each component in the database management system.

The first processing unit 1002 is used to obtain the interaction parameters between the first target component and the second target component according to the reinforcement learning model and the vectorized information. The vectorized information is obtained by vectorizing the query information and the state information. The interaction parameters include the parameters and service addresses that need to be transmitted between the first target component and the second target component.

The determining unit 1003 is used to determine a first path using a search algorithm according to the first target component, wherein the first path includes at least a first target component and a second target component, and the first target component and the second target component have different functions.

In this embodiment, the operations performed by each unit in the self-assembly manager are similar to those described in the embodiment shown in FIG. 5 , and will not be described in detail here.

In this embodiment, the first processing unit 1002 obtains the first target component according to the reinforcement learning model and the vectorized information, and the determination unit 1003 then uses the search algorithm to determine the first path. It can automatically select the corresponding component to run the query information according to the query information obtained by the acquisition unit 1001, and can flexibly respond to different business scenarios and improve execution efficiency.

Please refer to FIG. 11 , another embodiment of the self-assembly manager in the embodiment of the present application includes:

The acquisition unit 1101 is used to acquire query information and database status information, where the status information includes load information of each component in the database.

The first processing unit 1102 is used to obtain interaction parameters between at least one first target component and the second target component according to the reinforcement learning model and vectorized information, the vectorized information is obtained by vectorizing the query information and the state information, and the interaction parameters include parameters and service addresses that need to be transmitted between the first target component and the second target component.

The determination unit 1103 is used to search the components in the database based on the search algorithm and select the second target component to obtain the first path, so that the database responds to the query information according to the first path. The first path includes at least a first target component and a second target component, and the functions of the first target component and the second target component are different.

The first processing unit 1102 is specifically used to replace an element in the reinforcement learning model with the generator of the GAN to obtain a new reinforcement learning model; input vectorized information into the new reinforcement learning model to obtain the first target component and the interaction parameter of the first target component.

The acquisition unit 1101 is further used to obtain scoring information of the GAN discriminator on the first path, where the scoring information is related to the performance parameters of the first path.

The self-assembly manager in this embodiment also includes:

The first trigger unit 1105 is used to feed back the score of the first target component to the generator if the score of the first target component does not meet the preset conditions, so that the generator processes the first target component to obtain a processed first target component, and triggers the first processing unit 1102 and the first trigger unit 1105 to perform corresponding operations based on the processed first target component. If the score meets the preset conditions, the optimization component is output and the second processing unit 1104 is triggered to perform corresponding operations. The optimization component is the first target component whose score meets the preset conditions, and the score of the first target component is obtained after processing the scoring information.

The second processing unit 1104 is used to obtain target state information according to the reinforcement learning model, the optimization components and the state information, where the target state information includes the determined component information and the query state information.

The second trigger unit 1106 is used to trigger the first processing unit 1102, the determination unit 1103 and the first trigger unit 1105 to perform corresponding operations using the target state information as state information if the number of the optimization components does not meet the preset number, and output a second path if the number of the optimization components meets the preset number. The second path includes at least two optimization components, and each optimization component has a different function.

The response triggering unit 1107 is used to trigger at least two optimization components to perform operations in response to the query information according to the second path.

In this embodiment, the operations performed by each unit in the self-assembly manager are similar to those described in the embodiment shown in FIG. 5 , and will not be described in detail here.

In this embodiment, the first processing unit 1102 obtains the first target component according to the reinforcement learning model and the vectorized information, and the determination unit 1103 selects the second target component by using the search algorithm to obtain the first path, and selects the optimization component that meets the preset conditions in the first target component according to the first trigger unit 1105. If the number of the optimization components does not meet the preset number, the second trigger unit 1106 uses the target state information as the state information to trigger the first processing unit 1102, the determination unit 1103 and the first trigger unit 1105 to perform corresponding operations until the number of optimization components meets the preset number, and outputs the second path, so that the response trigger unit 1107 triggers at least two optimization components to perform operations in response to the query information according to the second path. The optimization component that meets the preset conditions can be automatically selected to perform operations in response to the query information according to the query information obtained by the acquisition unit 1101 and the state information of the database, which can flexibly respond to different business scenarios and improve execution efficiency.

The following describes the self-assembly manager in the embodiment of the present application. Please refer to FIG. 12. An embodiment of the self-assembly manager in the embodiment of the present application includes:

The self-assembly manager 1200 may include one or more processors (central processing units, CPU) 1201 and a memory 1205 , wherein the memory 1205 stores one or more application programs or data.

The memory 1205 may be a volatile storage or a persistent storage. The program stored in the memory 1205 may include one or more modules, each of which may operate on a series of instructions in the self-assembly manager. Furthermore, the processor 1201 may be configured to communicate with the memory 1205 and execute a series of instruction operations in the memory 1205 on the self-assembly manager 1200.

The self-assembly manager 1200 may also include one or more power supplies 1202, one or more wired or wireless network interfaces 1203, one or more input and output interfaces 1204, and/or, one or more operating systems, such as Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

The processor 1201 can execute the operations executed by the self-assembly manager in the embodiment shown in FIG. 5 , and the details are not repeated here.

FIG13 shows a schematic diagram of a database system provided by an embodiment of the present application, which is a cluster database system using a shared-storage architecture, including multiple computing nodes (such as nodes 1-N in FIG13), each node is deployed with one or more database microservice components, such as the parser service, optimizer service, executor service, storage engine service, metadata service, statistics service, self-monitoring service and clock service in the aforementioned embodiment, etc. A plurality of microservice components that are scheduled for execution in a specific order constitute an execution path, which can realize the function of a database management system to provide users with services such as query and modification of data. The component manager is used to manage all microservice components in the system. Specifically, the component manager is responsible for the registration and deregistration of microservice components, and maintains metadata of microservice components, including but not limited to: information about microservice components currently available in the database system, and the deployment location of each microservice component (such as the ID of the computing node where it is located).

Furthermore, the component manager can generate one or more execution paths based on the microservice components available in the system. For example, the component manager can generate an optimal execution path and a suboptimal execution path, or generate a TOP-N execution path. The suboptimal execution path is used as an alternative path. Accordingly, the job manager receives the query submitted by the client, calls multiple database microservice components to process the query according to the optimal execution path generated by the component manager to obtain the query result, and then returns the query result to the client. When an error occurs in a microservice component in the optimal execution path, multiple microservice components are called to process the query according to the alternative execution path.

In one embodiment, when the load of a microservice component in the system is high, for example, when the number of tasks executed by the microservice component exceeds a set threshold, the component manager can start one or more backups of the microservice component on an idle computing node, wherein the backup of the microservice component can be regarded as an instance of the microservice component, which has the same function as the microservice component. The job manager shares the load of the microservice component by calling the backup of the microservice component to improve the availability of the system. In other words, the component manager can dynamically adjust the microservice component based on the load of the system and adjust the execution path.

In one embodiment, computing nodes communicate with each other via an infinite bandwidth InfiniBand network. A shared data memory stores shared data that can be accessed by multiple computing nodes. The computing nodes perform read and write operations on the data in the shared data memory through a switch. The shared data memory can be a shared disk array or cloud storage, etc. The computing node can be a physical machine, such as a database server, or a virtual machine running on abstract hardware resources. If the node is a physical machine, the switch is a storage area network (SAN) switch, an Ethernet switch, a fiber switch or other physical switching device. If the node is a virtual machine, the switch is a virtual switch.

FIG. 14 is a schematic diagram of a cluster database system using a shared-nothing architecture. According to FIG. 14 , each computing node has its own exclusive hardware resources (such as data storage), operating system and database, and the nodes communicate through a high-speed network. Under this architecture, data will be allocated to each node according to the database model and application characteristics, and the query task will be divided into several parts, executed in parallel on all nodes, and calculated in coordination with each other to provide database services as a whole, and all communication functions are implemented on a high-bandwidth network interconnection system. One or more database microservice components are deployed on each node, such as the parser service, optimizer service, executor service, storage engine service, metadata service, statistics service, self-monitoring service and clock service in the aforementioned embodiment. When processing a query task, each node can call the microservice components located in other computing nodes through the job manager to assist in query processing. The data storage device (Data Store) includes but is not limited to a solid-state hard disk (SSD), a disk array or other types of non-transient computer-readable media.

Just like the cluster database system with a shared disk architecture described in FIG. 13 , the computing nodes here can be either physical machines or virtual machines.

It should be noted that in the database system shown in Figures 13 and 14, the specific functions and workflows of the microservice components, job manager and component manager can refer to the aforementioned related embodiments and will not be repeated here.

Those skilled in the art will appreciate that a database system may include fewer or more components than those shown in FIG. 13 and FIG. 14, or may include components different from those shown in FIG. 13 or FIG. 14, and FIG. 13 and FIG. 14 merely illustrate components that are more relevant to the implementation disclosed in the embodiments of the present application. For example, although four computing nodes are described in FIG. 13 and FIG. 14, those skilled in the art will appreciate that a cluster database system may include any number of nodes. The database management system functions of each node may be implemented by an appropriate combination of software, hardware, and/or firmware running on each node.

In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, read-only memory), random access memory (RAM, random access memory), disk or optical disk and other media that can store program code.

Claims (11)

1. A method for dynamically selecting components in a database management system, comprising: Step 1: Acquire query information and status information of a database management system, wherein the query information includes a data query statement, and the status information includes load information of multiple components in the database management system; Step 2: Obtain interaction parameters between the first target component and the first target component according to the reinforcement learning model and the vectorized information, wherein the vectorized information is obtained by vectorizing the query information and the state information, and the interaction parameters include parameters that need to be transmitted between the first target component and other components; Step 3: Determine a first path using a search algorithm according to the first target component, wherein the first path indicates a component sequence including the first target component and at least one second target component, and the first target component and the second target component have different functions; wherein determining the first path using a search algorithm according to the first target component comprises: randomly supplementing the first target component with the following components according to empirical data; searching for components in a database based on the search algorithm to obtain the second target component, so as to determine the first path; Step 4: Obtain scoring information of the first path by a discriminator of a generative adversarial network (GAN), wherein the scoring information is related to a performance parameter of the first path; Step 5: If the score of the first target component does not meet the preset condition, the score of the first target component is fed back to the generator of the GAN, so that the generator processes the first target component to obtain a processed first target component, and performs steps 3 to 5 based on the processed first target component; if the score meets the preset condition, an optimized component is output, and step 6 is performed, where the optimized component is the first target component whose score meets the preset condition or the processed first target component; the score of the first target component is obtained by processing the scoring information; Step 6: Obtain target state information according to the reinforcement learning model, the optimization component and the state information, wherein the target state information includes information of the optimization component and query state information; Step 7: If the number of the optimization components does not meet the preset number, repeat steps 2 to 5 using the target state information as the state information; if the number of the optimization components meets the preset number, output a second path, where the second path includes at least two optimization components, each of which has a different function; According to the second path, at least two of the optimization components are triggered to respond to the query information. 2. The method according to claim 1, characterized in that the step of obtaining the interaction parameters between the first target component and the first target component according to the reinforcement learning model and the vectorized information comprises: The interaction parameters of the first target component and the first target component are obtained according to the reinforcement learning model, the generative adversarial network GAN and the vectorized information. 3. The method according to claim 2, characterized in that obtaining the interaction parameters of the first target component and the first target component according to the reinforcement learning model, the generative adversarial network GAN and the vectorized information comprises: Replacing an element in the reinforcement learning model with the generator of the GAN to obtain a new reinforcement learning model; The vectorized information is input into the new reinforcement learning model to obtain the interaction parameters between the first target component and the first target component. 4 . The method according to claim 1 , wherein the search algorithm is a Monte Carlo tree search algorithm. 5. A self-assembly manager, comprising: An acquisition unit, used to acquire query information and status information of a database management system, wherein the query information includes a data query statement, and the status information includes load information of multiple components in the database management system; A first processing unit, configured to obtain interaction parameters between a first target component and the first target component according to a reinforcement learning model and vectorized information, wherein the vectorized information is obtained by vectorizing the query information and the state information, and the interaction parameters include parameters that need to be transmitted between the first target component and other components; a determining unit, configured to determine a first path using a search algorithm according to the first target component, wherein the first path indicates a component sequence including the first target component and at least one second target component, and the first target component and the second target component have different functions; The determining unit is specifically configured to randomly add the following components to the first target component according to empirical data; search the components in the database based on the search algorithm to obtain the second target component, so as to determine the first path; The acquisition unit is further used to obtain scoring information of the first path by a discriminator of a generative adversarial network (GAN), wherein the scoring information is related to a performance parameter of the first path; The self-assembly manager also includes: A first triggering unit is configured to feed back the score of the first target component to the generator of the GAN if the score of the first target component does not meet the preset condition, so that the generator processes the first target component to obtain a processed first target component, trigger the first processing unit and the first triggering unit to perform corresponding operations based on the processed first target component, and output an optimization component if the score meets the preset condition, and trigger the second processing unit to perform a corresponding operation, wherein the optimization component is the first target component whose score meets the preset condition or the processed first target component, and the score of the first target component is obtained by processing the scoring information; A second processing unit, configured to obtain target state information according to the reinforcement learning model, the optimization component and the state information, wherein the target state information includes information of the optimization component and query state information; a second triggering unit, configured to, if the number of the optimization components does not meet the preset number, use the target state information as the state information to trigger the first processing unit, the determination unit, and the first triggering unit to perform corresponding operations, and, if the number of the optimization components meets the preset number, output a second path, wherein the second path includes at least two of the optimization components, and each of the optimization components has a different function; A response triggering unit is used to trigger at least two of the optimization components to respond to the query information according to the second path. 6. The self-assembly manager according to claim 5 is characterized in that the first processing unit is specifically used to obtain the interaction parameters between the first target component and the first target component according to the reinforcement learning model, the generative adversarial network GAN and the vectorized information. 7. The self-assembly manager according to claim 6 is characterized in that the first processing unit is specifically used to replace an element in the reinforcement learning model with the generator of the GAN to obtain a new reinforcement learning model; and input the vectorized information into the new reinforcement learning model to obtain the first target component and the interaction parameters of the first target component. 8. The self-assembly manager according to any one of claims 5 to 7, wherein the search algorithm is a Monte Carlo tree search algorithm. 9. A self-assembly manager, comprising: Processor, memory, bus, input and output devices; The processor is connected to the memory and the input and output device; The bus is respectively connected to the processor, the memory and the input and output devices; The processor executes the method according to any one of claims 1 to 4. 10. A computer storage medium, characterized in that instructions are stored in the computer storage medium, and when the instructions are executed on a computer, the computer executes the method according to any one of claims 1 to 4. 11. A computer program product, characterized in that when the computer program product is executed on a computer, the computer is caused to execute the method according to any one of claims 1 to 4.