A method and system for object storage service based on big data component cluster

Abstract

The application discloses a kind of object storage service method and system based on big data component cluster, it is related to distributed storage technical field;The application forms real-time multidimensional characteristics by extracting data object static characteristics, cluster dynamic environment characteristics and historical access mode characteristics, splicing and preprocessing;Based on DDPG deep reinforcement learning model, continuous action space is processed through Actor-Critic network architecture, and a storage strategy model is obtained by training using historical storage data;Real-time characteristics are input into the model to generate an optimization strategy that includes storage location, number of copies, data shards, and life cycle management, and are converted into executable instructions for the cluster;The application overcomes the shortcomings of existing static strategies, enabling dynamic optimization of storage resources, effectively balancing performance, cost, and reliability, and improving the adaptive management capabilities and resource utilization efficiency of massive unstructured data storage in a heterogeneous cluster environment.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of distributed storage, and particularly relates to an object storage service method and system based on a big data component cluster. BACKGROUND

[0002] With the advent of the digital economy era, massive unstructured data is growing. In typical scenarios such as e-commerce promotion, social media, and Internet of Things monitoring, billions of pictures, videos, and log files are generated daily. These data have the characteristics of large total volume and dynamic changes in access patterns. Although object storage has become a standard solution for storing such data, it still faces severe challenges in practical applications.

[0003] Existing storage systems use static strategies to manage data, which cannot adapt to dynamic business needs. For example, in an e-commerce platform, a promotion event can cause the access volume of certain product pictures to increase by hundreds of times instantaneously. However, a static sharding strategy cannot perceive such changes in real time, resulting in severe uneven load distribution among cluster nodes, with some nodes overloaded and others idle. Meanwhile, the decay law of social media content is complex, but existing life cycle management still uses fixed time thresholds, which prevents high-value data from being stored with high performance in a timely manner, and invalid data occupies high-quality resources for a long time.

[0004] In terms of reliability, current solutions only recover data after a failure occurs. This passive response method seriously affects normal services in large-scale clusters. More importantly, the performance, cost, and reliability goals of a storage system often conflict with each other. The existing architecture lacks a unified intelligent decision-making mechanism, forcing administrators to manually allocate resources among different management tools, which makes it difficult to achieve overall optimization.

[0005] Therefore, the existing technology has the problems of rigid strategy, reactive management, and lack of overall optimization. SUMMARY

[0006] (I) Technical problems solved

[0007] To solve the problems in the related art, the present application provides an object storage service method based on a big data component cluster to overcome the above technical problems existing in the prior art.

[0008] (II) Technical solutions

[0009] To solve the above technical problems, the present application is implemented by the following technical solutions:

[0010] S1, obtain data object static features, big data component cluster dynamic environment features, and historical access pattern features; the historical access pattern features are obtained by mapping the data object static features to a metadata database;

[0011] S2, splice and preprocess the data object static features, the big data component cluster dynamic environment features and the historical access mode features to obtain real-time multi-dimensional spliced features;

[0012] S3, construct a DDPG model based on the continuous action space characteristics of the object storage data and the Actor-Critic architecture, collect historical storage service data, and train the DDPG model using the historical storage service data combined with a reward function to obtain a DDPG storage strategy model;

[0013] S4, input the real-time multi-dimensional spliced features into the DDPG storage strategy model to obtain an object storage strategy instruction;

[0014] S5, convert the object storage strategy instruction into an instruction recognizable by the big data component cluster to obtain an object storage control instruction, and the big data component cluster performs distributed storage based on the object storage control instruction;

[0015] Preferably, the S1 comprises the following steps:

[0016] S11, the data access layer receives a data object write request and performs verification, and after verification, agrees to the data object write request to obtain a data object;

[0017] extract a static feature vector of the data object to obtain a data object static feature; the data object static feature includes data size, data type and user or application label;

[0018] S12, collect system state data of the big data component cluster storage layer to obtain real-time big data component cluster data, and extract features of the real-time big data component cluster data to obtain big data component cluster dynamic environment features;

[0019] S13, query a meta database to retrieve an access mode of historical data having the same static features as the real-time data object to obtain a historical access mode feature;

[0020] Preferably, the S12 comprises the following steps:

[0021] S121, extract CPU utilization, memory usage, disk I / O throughput and network bandwidth occupancy of each storage node in the real-time big data component cluster data to obtain cluster load features;

[0022] S122, extract the remaining capacity and current IOPS of each storage medium in the real-time big data component cluster data to obtain storage resource features;

[0023] S123, extract the network topology structure and the health status of the nodes in the real-time big data component cluster data to obtain topology and health status;

[0024] The cluster load feature, storage resource feature, and topology and health status jointly constitute the big data component cluster dynamic environment feature;

[0025] Preferably, the S2 comprises the following steps:

[0026] S21, horizontally splicing the data object static feature, the big data component cluster dynamic environment feature, and the historical access mode feature to obtain an initial real-time multi-dimensional spliced feature;

[0027] S22, performing data cleaning and missing processing on the initial real-time multi-dimensional spliced feature to obtain an initial processed real-time multi-dimensional spliced feature;

[0028] Performing continuous numerical feature normalization, classification type normalization, and parameter normalization on the initial processed real-time multi-dimensional spliced feature to obtain a real-time multi-dimensional spliced feature;

[0029] Preferably, the S3 comprises the following steps:

[0030] S31, constructing a DDPG model; the DDPG model comprises an Actor-Critic architecture; the Actor-Critic architecture comprises an Actor network and a Critic network; the Actor network comprises an online Actor network and a target Actor network, and the Critic network comprises an online Critic network and a target Critic network;

[0031] S32, collecting historical cluster state logs, data access logs, and metadata and operation logs to obtain initial historical storage service data; and associating and aligning the data in the initial historical storage service data to obtain data in the historical storage service data;

[0032] S33, constructing a historical state-action-reward sequence based on the historical storage service data;

[0033] S34, training the DDPG model using the historical state-action-reward sequence in combination with a reward function to obtain a DDPG storage strategy model;

[0034] Preferably, the S33 comprises the following steps:

[0035] S331, extracting features based on the historical storage service data and splicing to obtain a historical multi-dimensional spliced feature;

[0036] S332, extracting actual executed actions from the operation logs in the historical storage service data to obtain historical action data; and taking the historical action data as actions;

[0037] S333, set a reward function, calculate the historical reward data obtained after performing the historical action data in the state of the historical multi-dimensional splicing feature;

[0038] use the historical multi-dimensional splicing feature as the state data, the historical action data as the action data, and the historical reward data as the reward data to obtain the historical state-action-reward sequence;

[0039] Preferably, the S34 comprises the following steps:

[0040] S341, set the convergence condition and the training termination condition of the DDPG model; specifically, the convergence condition comprises an average reward value threshold, and the training termination condition is a maximum training number; when any condition is met, the training is terminated;

[0041] S342, randomly sample a small batch of experience data from the experience replay buffer of the historical state-action-reward sequence to obtain first state data;

[0042] input the first state data into the online Actor network of the DDPG model to obtain first action data of its decision; execute the first action data in the simulation environment, and the environment will feedback second state data, and the first reward value is calculated and obtained through the reward function;

[0043] input the first state data and the first action data into the online Critic network to obtain an evaluation value of the current decision of the online Critic network, denoted as a current evaluation value;

[0044] input the second state data into the target Actor network to obtain second action data; then input the second action data and the second state data into the target Critic network to obtain a target evaluation value, denoted as a target evaluation value;

[0045] reduce the gap between the current evaluation value and the target evaluation value by minimizing the loss function to obtain a loss value; update the parameters of the online Critic network according to the loss value;

[0046] the online Actor network adjusts the network parameters by calculating the policy gradient of the current evaluation value; the adjustment direction is to increase the evaluation value given by the online Critic network, that is, to guide the Actor network to make decisions in similar states in the future to obtain higher rewards;

[0047] use a soft update formula to slowly synchronize the parameters of the online Actor network and the online Critic network to their corresponding target networks to maintain the stability of the training process;

[0048] S343, repeating step S342, when the average reward value in the period ≥ average reward value threshold, or the training cycle number reaches the set maximum training number, the training is terminated, and the DDPG storage strategy model is obtained;

[0049] Preferably, the S4 comprises the following steps:

[0050] S41, inputting the real-time multi-dimensional splicing features into the DDPG storage strategy model, and outputting the object storage strategy instruction through the Actor network according to the action features in the real-time multi-dimensional splicing features;

[0051] Preferably, the S5 comprises the following steps:

[0052] S51, the policy executor receives the object storage strategy instruction and translates it into a specific control command that can be recognized by the big data component cluster storage layer, to obtain an object storage control instruction;

[0053] S52, the big data component cluster storage layer divides the data object according to the received object storage control instruction, to obtain a data copy after the division;

[0054] The data copy after the division is routed to the physical node and storage medium specified by the policy, and the final storage location, policy information and life cycle label of the data object are updated in the meta database.

[0055] An object storage service system based on a big data component cluster, used to implement the above-mentioned object storage service method based on a big data component cluster, the system comprising:

[0056] A feature extraction module for extracting key features from multi-source data, including data object static features, big data component cluster dynamic environment features and historical access pattern features;

[0057] A feature preprocessing module for horizontally splicing the data object static features, cluster dynamic environment features and historical access pattern features to obtain real-time multi-dimensional splicing features;

[0058] A model training module for constructing and training a storage strategy model based on the DDPG algorithm, learning the optimal decision from historical storage service data; designing an Actor-Critic network architecture, wherein the Actor network processes continuous action space and the Critic network evaluates decision quality; then constructing a state-action-reward sequence from historical logs and calculating the reward value through a reward function; iteratively training through experience replay and network parameter update until the convergence condition is met, and outputting the trained DDPG storage strategy model;

[0059] A strategy generation module inputs the multi-dimensional splicing features into the trained DDPG storage strategy model, and generates fine-grained object storage strategy instructions through forward propagation calculation of the Actor network.

[0060] A strategy execution module is configured to convert the object storage strategy instructions output by the strategy generation module into specific control commands recognizable by the big data component cluster, and execute distributed storage operations.

[0061] (Three) beneficial effects

[0062] The present application has the following beneficial effects:

[0063] The present application realizes dynamic optimization and adaptive decision of object storage strategy through the DDPG model. The system can comprehensively analyze the static characteristics of data objects, real-time load of the cluster and historical access patterns, and generate fine-grained storage strategies that take into account performance, cost and reliability. This intelligent decision mechanism based on the DDPG model effectively overcomes the rigidity problem of traditional static strategies, and significantly improves the adaptability of the storage system in a dynamic heterogeneous environment.

[0064] The present system realizes multi-objective collaborative optimization of storage resources. Through the designed reward function, performance indicators (such as access delay), cost factors (such as storage medium selection) and reliability requirements (such as the number of replicas) are unified into the optimization target, so that the system can intelligently balance multiple key indicators while ensuring service quality, avoiding the common resource waste or single-objective optimization limitations in traditional solutions, thereby significantly improving the overall efficiency of the cluster.

[0065] The present application constructs a complete intelligent management closed loop from feature extraction to strategy execution. The system can not only make accurate initial placement decisions based on real-time state, but also predict data life cycle through historical access patterns, realize forward-looking data migration and layering. This active management mechanism can effectively prevent hot spot imbalance and identify potential fault risks in advance, thereby reducing service window period risks and improving the reliability and operation efficiency of the storage system.

[0066] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0068] Figure 1A flowchart of a method for object storage service based on a big data component cluster according to the present application;

[0069] Figure 2 A module diagram of a system for object storage service based on a big data component cluster according to the present application. DETAILED DESCRIPTION

[0070] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only some of the embodiments of the application, but not all the embodiments of the application. Based on the embodiments of the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0071] In the prior art, object storage service usually adopts a strategy based on static rules for data management. The following is the processing mode of the prior art:

[0072] After completing authentication and analysis, the video is mechanically classified as a "medium" file according to the video size, and a fixed storage process is triggered, and the preset parameters are strictly followed: regardless of the real-time state of the cluster, the default 3-copy strategy, fixed 10MB chunk size, and simple "active data storage SSD" rule are adopted, and all data blocks are written to the SSD storage pool; the final result is that the metadata query delay is stable within 15ms, the delay is large, the file read-write throughput is 400MB / s, the throughput is low, and the storage cost is high because the fixed 3-copy SSD scheme is adopted, and the cost is high;

[0073] Based on the above prior art, the main problems are that the static rules do not match the dynamic load, the resource utilization is low, and the fault recovery is slow;

[0074] To solve the above problems, please refer to Figure 1 The embodiments of the application disclose a method for object storage service based on a big data component cluster, comprising the following steps:

[0075] S1, obtain data object static features, big data component cluster dynamic environment features, and historical access mode features; the historical access mode features are obtained by mapping the data object static features to a metadata database;

[0076] S2, splice and preprocess the data object static features, big data component cluster dynamic environment features, and historical access mode features to obtain real-time multi-dimensional spliced features;

[0077] S3, based on the object storage data continuous action space characteristics and Actor-Critic architecture, build DDPG model; collect historical storage service data; use the historical storage service data to combine the reward function to train the DDPG model, and obtain the DDPG storage strategy model;

[0078] S4, input the real-time multi-dimensional splicing features to the DDPG storage strategy model, and obtain the object storage strategy instruction;

[0079] S5, convert the object storage strategy instruction into an instruction that can be recognized by the big data component cluster, and obtain the object storage control instruction; the big data component cluster performs distributed storage based on the object storage control instruction;

[0080] The above embodiment constructs an object storage service system integrating perception, decision and execution. The system forms comprehensive environmental perception by fusing data static attributes, cluster real-time state and historical access mode. The system automatically masters the optimal decision strategy through learning historical data by using a deep reinforcement learning model, and converts intelligent decision into specific distributed storage operation. The system effectively solves the problems of rigid strategy and single optimization target of traditional storage systems, can dynamically adjust the storage strategy according to real-time work load, realizes precise balance between performance, cost and reliability, and significantly improves the resource utilization rate, operation and maintenance automation level and overall system performance of large-scale heterogeneous storage clusters;

[0081] The above step S1 includes the following steps:

[0082] S11, receiving a data object write request from a client through a data access layer and verifying, after verification, agreeing to the data object write request and obtaining a data object; specifically, the data object is unstructured data such as video, picture, document and log file;

[0083] Extracting a static feature vector of the data object to obtain a data object static feature; specifically, the data object static feature includes data size, data type and user or application label; at the same time, according to file size specification, the data object is classified into three categories: normal (less than 100K), medium (100K to HDFS Block size, default 128MB) and max (greater than 128MB) to adapt to differentiated storage strategies;

[0084] S12, collecting system state data of a big data component cluster storage layer to obtain real-time big data component cluster data;

[0085] Extract features from real-time big data component cluster data to obtain dynamic environment features of the big data component cluster; the dynamic environment features of the big data component cluster are used to perceive the health status and resource load of the cluster in real time, ensuring that the storage strategy can adapt to the real-time status of the cluster and achieve load balancing and resource optimization.

[0086] S13. Query the metadata database to retrieve access patterns of historical data with the same static characteristics as the real-time data objects, thus obtaining historical access pattern features. Specifically, the metadata database is a dedicated database for storing system metadata, recording the attribute information, storage location, access history, and policy configuration of all data objects, providing data support and contextual information for intelligent decision-making. The metadata database is built on an HBase distributed storage cluster and adopts a dynamic table partitioning design. It generates uniformly distributed RowKeys through a hash algorithm to avoid data hotspots and ensure that the metadata query latency is ≤10ms, providing performance guarantees for fast retrieval of historical access patterns. Similar features include being from the same user or of the same type; access patterns include initial access frequency and heat decay period.

[0087] In specific implementation, step S11 above is as follows: the data type is determined by the file extension or MIME type, and the user or application tag is used to infer the access mode; specifically, the API gateway of the data access layer receives HTTP / HTTPS requests, parses the request header and data body, verifies the client's access key and signature, confirms that it has the right to perform write operations, temporarily writes the data stream to the high-speed buffer for subsequent processing, generates a unique identifier (such as UUID) for the data object, and records the basic information received, including timestamp and client IP.

[0088] Step S12 above includes the following steps:

[0089] S121. Extract the CPU utilization, memory usage, disk I / O throughput, and network bandwidth utilization of each storage node in the real-time big data component cluster data to obtain the cluster load characteristics.

[0090] S122. Extract the remaining capacity and current IOPS of each storage medium (such as SSD, HDD) in the real-time big data component cluster data to obtain the storage resource characteristics;

[0091] S123. Extract the network topology and node health status (such as disk SMART warning information) from the real-time big data component cluster data to obtain the topology and health status.

[0092] The cluster load characteristics, storage resource characteristics, and topology and health status together constitute the dynamic environment characteristics of the big data component cluster.

[0093] In a specific implementation, the step S12 is specifically: the cluster load feature is obtained by aggregating (e.g., calculating the average value, maximum value, quantile) the indicators of all nodes in the cluster to represent the overall load; the health score is calculated according to the medium type (SSD / HDD) and the usage time to calculate the wear degree; the network topology state is obtained from the cluster management tool (e.g., Kubernetes, YARN) to obtain the rack awareness information, the network delay matrix between nodes, and the node health state is extracted from the disk SMART data to extract the key warning indicators (e.g., reallocation sector count, CRC error) to confirm whether the DataNode, Object Storage Gateway and other service processes are normal;

[0094] The above embodiment is applied to a large-scale heterogeneous object storage cluster environment, and an intelligent decision basis is constructed through multi-dimensional feature extraction; the system performs security verification and buffer processing on the unstructured data written in the data access layer, extracts static features such as data size, type and user label, and simultaneously collects load indicators such as CPU, memory, disk I / O and network bandwidth of each node in the cluster, and combines the storage medium capacity and health state to form dynamic environment features; then, by querying the meta database, the access rule of the historical data with similar features is obtained, and a complete environment portrait is constructed; the problem of strategy rigidity caused by single feature perception in traditional storage systems is effectively solved, and through the collaborative analysis of static attributes, real-time load and historical patterns, a comprehensive and accurate environment perception basis is provided for subsequent intelligent decision-making, thereby significantly improving the matching accuracy of storage strategies and dynamic business needs.

[0095] In a specific implementation, the step S13 is specifically: according to the static features of the current data object, a query condition is established in the meta database to filter a set of historical data objects with the same static features; a query request is initiated to the meta database to obtain the historical access records of the matching data objects, including access timestamp, access frequency, read-write ratio and other original log data; the query results are statistically analyzed to collect key pattern indicators including access frequency, hotness half-life and typical storage period;

[0096] The key pattern indicators are normalized and combined into a structured historical access pattern feature vector for subsequent feature splicing;

[0097] The step S2 includes the following steps:

[0098] S21, horizontally splicing the data object static features, the dynamic environment features of the big data component cluster, and the historical access pattern features to obtain initial real-time multi-dimensional spliced features;

[0099] S22, data cleaning and missing processing are performed on the initial real-time multi-dimensional spliced features to obtain initial processed real-time multi-dimensional spliced features;

[0100] The initial processing real-time multi-dimensional splicing features are subjected to continuous numerical feature normalization, classification type normalization and parameter normalization to obtain real-time multi-dimensional splicing features; specifically, the real-time multi-dimensional splicing features are a fixed dimension, numerical range unified, high-quality floating point number vector; the continuous numerical features include data size, CPU load; the parameters include mean (μ), standard deviation (σ), minimum value (min) and maximum value (max);

[0101] In specific implementation, the continuous numerical feature normalization is performed by Z-Score standardization or Min-Max normalization, the classification type normalization is converted into a numerical value by one-hot encoding, and the parameter normalization is calculated from the historical data distribution;

[0102] After the multi-source feature extraction is completed, the above embodiment realizes the unified processing of heterogeneous data through a feature engineering module; the data object static features, cluster dynamic environment features and historical access mode features are horizontally spliced to form an initial feature vector; the data integrity is ensured through data cleaning and missing value processing, and the continuous numerical features are subjected to Z-Score standardization processing and the classification features are subjected to one-hot encoding processing, so as to finally generate a fixed dimension, numerical unified floating point number vector; the problem of unstable model training caused by the diverse feature sources and different dimensions is effectively solved, a high-quality standardized input is provided for the subsequent DDPG model, and the accuracy and reliability of intelligent decision-making are significantly improved;

[0103] The above step S3 includes the following steps:

[0104] S31, constructing a DDPG model; the DDPG model includes an Actor-Critic architecture; the Actor-Critic architecture includes an Actor network and a Critic network; the Actor network includes an online Actor network and a target Actor network, and the Critic network includes an online Critic network and a target Critic network;

[0105] S32, collect a historical cluster state log, a data access log, and metadata and an operation log to obtain initial historical storage service data; and associate and align data in the initial historical storage service data to obtain data in the historical storage service data; the historical cluster state log is from a cluster monitoring system and includes time-sequenced CPU, memory, disk I / O, network bandwidth, and storage capacity; the data access log is from an object storage gateway and an access log of an HDFS NameNode; and includes a unique identifier of each data object, an access timestamp, an operation type, an access delay, and a data size; the metadata and the operation log are from a metadata database and a cluster management tool and include a storage location of a data object, a number of copies, a storage strategy, and a record of a management operation such as data migration, copy repair, and deletion that is automatically or manually performed by the system;

[0106] S33, construct a historical state-action-reward sequence based on the historical storage service data;

[0107] S34, train the DDPG model using the historical state-action-reward sequence in combination with a reward function to obtain a DDPG storage strategy model;

[0108] In a specific implementation, the step S31 is specifically: the reason for selecting DDPG is that, in object storage cluster management, a decision (action) has a continuous action space characteristic, such as selecting a target node (from a list of hundreds of nodes), determining a number of copies (which can be 2, 3, 4, and the like), and suggesting a data shard size; DDPG is a deep reinforcement learning algorithm specially designed for solving a continuous action space problem, which can output a continuous and deterministic action vector; secondly, DDPG adopts an Actor-Critic (double network structure), the Actor is used for decision-making, and the Critic is used for guiding the improvement of a strategy; and finally, DDPG adopts offline training, which is suitable for offline simulation of object storage cluster management;

[0109] The Actor network includes an input layer, a hidden layer 1, a hidden layer 2, a hidden layer 3, and an output layer; the input layer has the same dimension as a state feature vector (usually 150-300 neurons) and adopts a linear activation function; the hidden layer 1 contains 512 neurons, uses a ReLU activation function, adopts a He Normal method for weight initialization, and enables Dropout (a dropout rate is 0.1) in a training stage; the hidden layer 2 contains 256 neurons, uses a ReLU activation function, adopts a He Normal weight initialization, and adds a batch normalization process; the hidden layer 3 contains 128 neurons, uses a ReLU activation function, and adopts a He Normal weight initialization; and the output layer has a dimension matched with an action space and contains five special output heads;

[0110] The Critic network comprises a state branch, an action branch, a merging layer and an output layer; the state branch comprises an input layer (with the same dimension as the state feature vector), a hidden layer 1 (with 256 neurons and using a ReLU activation function), and a hidden layer 2 (with 128 neurons and using a ReLU activation function); the action branch comprises an input layer (with the same dimension as the action space), a hidden layer 1 (with 128 neurons and using a ReLU activation function); the merging layer is connected in such a manner that the 128-dimensional output of the state branch and the 128-dimensional output of the action branch are spliced into a 256-dimensional vector, comprising a hidden layer 1 (with 256 neurons and using a ReLU activation function) and a hidden layer 2 (with 128 neurons and using a ReLU activation function); the output layer comprises one neuron, adopts a linear activation function for direct output, and uses the Xavier Uniform method for weight initialization;

[0111] In a specific implementation, the step S32 is specifically: the data in the initial historical storage service data are associated and aligned, logs of different sources are associated and aligned according to timestamps, and a unified time series data set is formed, and each record represents a complete state St of the system at a certain time point t and a subsequent action At (such as a data being written and placed in a certain position);

[0112] The step S33 comprises the following steps:

[0113] S331, features are extracted and spliced based on the historical storage service data to obtain historical multi-dimensional spliced features; the step of extracting and splicing features based on the historical storage service data is the same as the steps S1 and S2;

[0114] S332, actual executed actions are extracted from operation logs in the historical storage service data to obtain historical action data; the historical action data are taken as actions (such as “data X is placed in nodes [A, B, C] in the form of 3 copies”);

[0115] S333, a reward function is set, and historical reward data obtained after an action in the historical multi-dimensional spliced action state is executed is calculated according to the reward function; for example, an access delay of data X and a load balancing degree of nodes A, B and C within a period of time are calculated, and a reward value Rt is obtained by comprehensive consideration; the reward function is as follows:

[0116] The reward value Rt = α*performance reward + β*cost reward + γ*reliability reward - δ*punishment term; wherein, Rt represents a reward value of the current action, α, β, γ and δ respectively represent a performance reward weight, a cost reward weight, a reliability reward weight and a punishment term weight;

[0117] The performance reward is obtained by adding a delay reward and a throughput reward, the delay reward = -log(actual delay / target delay), the actual delay is the time from receiving a data read request to returning the first data byte, and the target delay is an average value based on historical normal delay data, the metadata query response target delay can be set to ≤10 ms, and the file read-write throughput target can be set to 600 MB / s in a gigabit network environment, the throughput reward = tanh(actual throughput / target throughput), the actual throughput is the actual amount of data successfully transmitted per unit time, and the target throughput is an average value based on historical normal throughput data;

[0118] The cost reward is obtained according to a storage cost reward and a network cost reward, the storage cost is calculated according to the unit capacity price of different storage media (such as SSD yuan / GB / month, HDD yuan / GB / month), the storage cost reward = -(data size*storage unit price), and the network cost is used when data is placed across availability zones or data centers, the network cost reward = -(data size*network unit price); the cost reward herein is used to preferentially select inexpensive resources under the premise of meeting performance;

[0119] The reliability reward is derived from cluster health monitoring and load balancer statistics, and is obtained by adding a load balancing degree reward, a fault domain separation reward, and a health degree reward, the load balancing degree is obtained by calculating the standard deviation of CPU / disk utilization of all storage nodes, the smaller the standard deviation, the higher the reward, and the load balancing degree reward = -log(1+utilization standard deviation), the fault domain separation checks whether data replicas are distributed in different fault domains (such as different racks, different power supplies), if completely separated, a fixed positive reward value is given, otherwise it is zero, and the health degree reward is obtained by normalizing the overall health score of the node (based on SMART data, memory errors, etc.);

[0120] The calculation of the reliability reward needs to fully consider the fault tolerance and load balancing mechanism of the system. The reward action can effectively utilize the HDFS Block copy strategy to realize data redundancy, and guarantee the availability of data through intra-zone / cross-zone replication strategy, to ensure the overall availability of the system. Specifically, the fault tolerance and load balancing mechanism guarantees the reliability of data through data redundancy and intra-zone / cross-zone replication, and realizes the load balancing of multiple service nodes through reverse proxy; through the copy strategy of Hadoop cluster for data block, the data redundancy of meta information and data file is realized; object intra-zone / cross-zone replication is to automatically and asynchronously (near real-time) copy the creation, update and deletion of objects in the Bucket of a certain region of the same or different account to the target Bucket in another region, to ensure security and availability; through reverse proxy of the object storage server, the load balancing of intra-zone or cross-zone service nodes is realized, to guarantee the high availability of object storage service;

[0121] The penalty term is obtained by adding the SLA violation penalty, node overload penalty, and storage strategy violation penalty; the SLA violation penalty imposes a fixed value Pbpenalty when the access delay or error rate exceeds the threshold value specified in the SLA agreement; the fixed value Pbis based on the loss evaluation caused by SLA violation, and is set to 5-10 times the maximum reward value that can be obtained in a single positive reward period (such as 1 hour) through historical data analysis; for example, if the system can obtain a maximum reward of +20 in 1 hour, the SLA violation penalty fixed value is set to -100 to -200; the node overload penalty triggers when the CPU, memory or disk I / O usage of any node exceeds the safety threshold (such as 90%), and the penalty value increases linearly with the degree of over-limit; the linear increase formula is penalty value = Pb+k (actual utilization rate-safety threshold); where k represents the slope in the linear increase formula; the safety threshold is set according to the cluster SLA commitment and historical operation and maintenance experience, and is determined by analyzing the system performance inflection point and hardware stability to determine the critical value of triggering alarm, and is set before the critical point at which resource utilization begins to significantly affect service quality; the storage strategy violation penalty, such as not storing specific data in the specified region according to the compliance requirements, will be punished by a fixed value; the penalty term particularly includes the penalty for violating the data lifecycle management strategy, such as failing to perform the corresponding operation within the set lifecycle (such as "automatically archived after 30 days of creation"), which will trigger the strategy violation penalty to guide the model to follow the automated lifecycle management principle;

[0122] The determination of the weight coefficients a, b, g, d is a process of converting business objectives into a mathematical model, and the weights are different in different scenarios; in the initial scenario, according to historical experience, the fixed weights are set as a = 0.3, b = 0.3, g = 0.3, d = 0.1; in the online service scenario, performance is crucial, so a (performance weight) should be set to the highest, b (cost weight) is relatively low, and the reduced weight of b is added to the performance weight, g and d remain unchanged; in the data archiving scenario, cost is the primary consideration, so b should be set to the highest, a is relatively low, and the reduced amount of a and b is added to the cost weight b, g and d remain unchanged; in the financial or critical data scenario, reliability is the bottom line, g (reliability weight) should be set to the highest, and the weight of d (penalty term) is also correspondingly increased, and the reduced amount of a and b is added to g and d;

[0123] Taking the online service scenario as an example, a = 0.5, b = 0.1, g = 0.3, d = 0.1; performance: actual delay = 5ms (target = 10ms), throughput = 980MB / s (target = 1000MB / s), delay reward = -log(5 / 10) = 0.3; throughput reward = tanh(980 / 1000) = 0.75, performance reward = 0.3 + 0.75 = 1.05; the cost of using SSD storage 1GB data (unit price = 0.1 yuan / GB), no network cost; cost reward = -(1x0.1) = -0.1; reliability: node utilization standard deviation = 0.15, fault domain is completely separated, node health = 0.9, load balancing reward = -log(1+0.15) = -0.14, fault domain reward = +0.5 (fixed value), health reward = 0.9, reliability reward = (-0.14) + 0.5 + 0.9 = 1.26; penalty term: no SLA violation and policy violation, but there is a node CPU utilization = 92% (over limit 2%); node overload penalty = 10 + 2x(92-90) = 14 (set Pb = 10, k = 2);

[0124] At this time, Rt = 0.5*1.05 + 0.1*(-0.1) + 0.3*1.26 - 0.1*14 = 0.525 - 0.01 + 0.378 - 1.4 = -0.507;

[0125] The negative reward result will guide the model to avoid similar placement strategies that cause node overload in subsequent decision-making;

[0126] The historical multi-dimensional splicing features are taken as state data, the historical action data is taken as state data, and the historical reward data is taken as reward data, to obtain a historical state-action-reward sequence;

[0127] The above steps provide a high-quality supervision signal for model training by constructing a historical state-action-reward sequence; the system extracts multi-dimensional features as states based on historical storage data, parses actual execution actions from operation logs, and uses a multi-objective reward function Rt=α*performance reward+β*cost reward+γ*reliability reward-δ*punishment term for comprehensive evaluation; effectively solving the problems of single optimization target and lack of systematic evaluation system in traditional methods, and providing a solid foundation for the DDPG model to learn the optimal strategy considering performance, cost and reliability by quantitatively evaluating the comprehensive effect of historical decisions, significantly improving the generalization ability and practicality of the strategy;

[0128] The above step S34 comprises the following steps:

[0129] S341, set the convergence condition and training termination condition of the DDPG model; specifically, the convergence condition includes an average reward value threshold, and the training termination condition is a maximum number of training times; when any condition is met, the training is terminated;

[0130] S342, randomly sample a small batch of experience data from the experience replay buffer of the historical state-action-reward sequence to obtain first state data;

[0131] The first state data is input into the online Actor network of the DDPG model to obtain the first action data of its decision; the first action data is executed in the simulation environment, and the environment will feedback the second state data, and the first reward value is calculated and obtained through the reward function;

[0132] The first state data and the first action data are input into the online Critic network to obtain the evaluation value of the online Critic network for the current decision, denoted as the current evaluation value;

[0133] The second state data is input into the target Actor network to obtain the second action data; the second action data and the second state data are input into the target Critic network to obtain the target evaluation value, denoted as the target evaluation value;

[0134] The difference between the current evaluation value and the target evaluation value is reduced by minimizing the loss function (such as mean square error loss) to obtain a loss value; the parameters of the online Critic network are updated according to the loss value;

[0135] The online Actor network adjusts the network parameters by calculating the policy gradient of the current evaluation value; the adjustment direction is to increase the evaluation value given by the online Critic network, that is, to guide the Actor network to make decisions in similar states in the future that can obtain higher rewards;

[0136] The parameters of the online Actor network and the online Critic network are slowly synchronized to their corresponding target networks using a soft update formula to maintain stability of the training process.

[0137] S343, repeating step S342, when the average reward value in the cycle is greater than or equal to the average reward value threshold, or the number of training cycles reaches the set maximum number of training, the training is terminated, and a DDPG storage policy model is obtained;

[0138] The average reward value threshold is set according to benchmark testing and empirical analysis in an offline simulation environment. In the early stage of model training, a performance benchmark is established by running a rule-based static policy (such as round robin allocation, fixed replica policy) in the simulation environment, and the average reward value obtained by the performance benchmark is calculated. The average reward value threshold is set to be 15-30% higher than the average reward value obtained by the performance benchmark, which is used to ensure that the trained model has significantly better performance than the static policy, and the average reward value threshold is also used to ensure that the model meeting the threshold can meet the minimum business income expectation according to the quantitative requirements of performance and cost targets on the business side.

[0139] The above embodiment uses the DDPG deep reinforcement learning algorithm based on historical state-action-reward sequences in the model training stage, and optimizes the intelligent policy through a carefully designed training mechanism. The convergence and termination conditions are set to ensure the training efficiency. The experience replay buffer is used for small batch sampling to improve data utilization efficiency. The online and target double network architecture is constructed, and the Critic network is used to evaluate the decision value and the Actor network is used to optimize the policy output. The soft update mechanism is used to ensure the training stability. This training scheme effectively solves the policy solidification problem caused by the lack of continuous learning ability of traditional storage systems, enables the model to learn the optimal decision under multi-objective trade-off from historical experience, significantly improves the adaptive ability and intelligent level of the storage policy, and lays a solid foundation for building an intelligent storage system with continuous optimization capability.

[0140] The above step S4 includes the following steps:

[0141] S41, inputting the real-time multi-dimensional splicing features into the DDPG storage policy model, and the DDPG storage policy model calculates the object storage policy instruction through the Actor network according to the action features in the real-time multi-dimensional splicing features.

[0142] In a specific implementation, the step S41 is specifically: the object storage policy instruction includes a target storage component, a storage medium, a number of copies, a data shard size and a placement position, a data encoding strategy, and an initial lifecycle label for a fine-grained storage and management policy of a current data object; meanwhile, a differentiated storage strategy is adopted based on file size classification (normal / medium / max): for the normal class (<100K), meta information and data are directly stored to HBase; for the medium class (100K to 128MB), MOB technology of HBase is used for block storage; for the max class (>128MB), actual data is written to HDFS, and meta information is recorded in HBase; in addition, a security mechanism is integrated into the instruction set, such as permission control and lifecycle management based on RBAC, to realize access control and automatic lifecycle management of the Object object; in the process of formulating the policy instruction, HDFS+HBase is used as a basic component, a five-layer architecture of "client layer-API adaptation layer-metadata management layer-data storage layer-fault tolerance and load balancing layer" is adopted, and the decision coverage is ensured from API access, metadata management, data storage to service high availability of the whole link, forming a complete and collaborative intelligent storage solution.

[0143] The target storage component is selected from HDFS, an object storage bucket, or other specific storage services; the storage medium is specified as SSD or HDD; the number of copies is dynamically determined according to data importance and current cluster state (such as 2 copies or 3 copies); the data shard size and placement position are precisely specified as which several physical nodes in the cluster store the data shard; the data encoding strategy is selected as which compression algorithm or erasure code strategy; and the initial lifecycle label is a predicted initial hotness, and an initial migration or degradation time threshold is set for it.

[0144] The above embodiment generates a fine-grained storage instruction including a storage location, a number of copies, an encoding strategy, and the like by inputting real-time multi-dimensional features into a trained DDPG model; this intelligent decision-making process effectively solves the problem that a traditional static strategy cannot adapt to a dynamic environment, realizes accurate resource allocation according to real-time cluster state and data characteristics, and significantly improves storage efficiency and resource utilization.

[0145] The step S5 includes the following steps:

[0146] S51, the policy executor receives the object storage policy instruction and translates it into specific control commands that the big data component cluster storage layer can recognize, obtaining an object storage control instruction; the translation process adopts an "client layer-API adaptation layer-metadata management layer-data storage layer-fault tolerance and load balancing layer" architecture, realizes S3 / Swift standard API parsing and AKSK authentication through the API adaptation layer, and adopts AES-256 transmission encryption and SM4 storage encryption to ensure the security and compatibility of the instruction;

[0147] S52, the big data component cluster storage layer divides the data object according to the received object storage control instruction, and obtains a data copy after the division;

[0148] The data copy after the division is routed to the physical node and storage medium specified by the policy, and the final storage location, policy information and life cycle label of the data object are updated in the metadata database; at the same time, the fault tolerance and load balancing layer mechanism is used to guarantee data reliability through data redundancy (such as HDFS Block copy strategy) and intra / inter-area replication, and load balancing of multiple service nodes is realized through reverse proxy to ensure system availability ≥ 99.99%; for the delete operation, soft deletion is adopted and combined with MapReduce-based garbage collection service for periodic cleaning to prevent data residue and ensure effective recovery of storage space;

[0149] In specific implementation, the step S51 is specifically: the object storage control instruction is a highly abstract policy description, and the executor needs to first parse and verify it to ensure the integrity and rationality of the instruction; by calling the cluster management tool or directly generating the bottom layer control script, the abstract policy is converted into specific commands executable by the storage layer; this process includes determining the target storage component, the number of copies, the node position list and other parameters, and encapsulating into a standard job request to prepare for subsequent distributed storage operation; for the Kubernetes environment, the abstract policy is converted into specific commands executable by the storage layer, specifically: creating a persistent volume declaration containing a specified storage class (such as ssd-tier), resource quota (such as storage capacity) and node selector (explicit physical node); for the HDFS cluster, specific Balancer instructions are generated, including the number of target copies of data blocks, the accurate block placement node list (such as specifying racks and disk types), and the accurate byte number of data shard size; these specific commands directly drive the bottom layer system to perform accurate resource allocation and data distribution operation;

[0150] In a specific implementation, the step S52 is specifically: cutting the original data object into a plurality of data shards of fixed or variable sizes according to a shard size specified in the policy; generating a specified number of replicas for each data shard; in this process, the system applies a data encoding method specified in the policy, such as a compression algorithm or an erasure code, to optimize storage efficiency and reliability; all shards and their replicas are assigned unique global identifiers to ensure accurate positioning and retrieval in a distributed environment;

[0151] The above embodiment converts intelligent decisions into specific control instructions through the policy executor, and completes the whole process of data sharding, encoding, routing and metadata updating; realizes an automatic closed loop from decision to execution, avoids the inefficiency and errors of manual operation in the traditional scheme, and significantly improves the accuracy and efficiency of storage policy execution.

[0152] The following is based on the embodiment 1 of the present application, as follows,

[0153] Through the hardware and basic software layer, a Hadoop cluster (2 NameNode masters, 8 DataNode) and a HBase cluster (1 HMaster, 8 RegionServer) containing 10 nodes are deployed. The HDFS Block size is set to 128MB, and the HBase Region threshold is set to 10GB;

[0154] The intelligent decision layer deploys the DDPG storage strategy model of the present application and its related feature extraction and preprocessing modules;

[0155] The service access layer deploys 3 API adaptation nodes, uses Nginx as a reverse proxy, and configures AK / SK authentication, AES-256 transmission encryption and SM4 storage encryption

[0156] The client layer receives a video processing application through the AWS S3 CLI client, initiates a PUT request, and uploads a video file video 001.mp4 with a size of 80MB; the API adaptation layer receives the request, completes S3 API parsing, AK / SK identity authentication and transmission decryption; after authentication, the request and its payload are forwarded to the intelligent decision system; the system extracts the file size (80MB), data type (video / MP4), and user label ("video processing application"); according to the differentiated storage policy, this file is classified as medium (100K ~ 128MB);

[0157] The cluster state is collected in real time; it is found that the IOPS of the SSD disks of nodes 3 and 5 are high (load characteristics), the remaining capacity of the HDD disk of node 7 is sufficient (resource characteristics), and the health status of all nodes is good;

[0158] The meta database based on HBase is queried and optimized for RowKey distribution and query delay <10ms, and the same type of video files from the same application are found, which are usually accessed frequently within the first 7 days after creation, and then the heat significantly decays;

[0159] The above three types of features are horizontally spliced to form an initial multi-dimensional feature vector; the vector is data cleaned, missing value processed, and continuous features (such as file size, CPU utilization) are Z-Score standardized, and classification features are one-hot encoded, to finally generate a standardized, fixed-dimensional real-time multi-dimensional spliced feature vector;

[0160] The real-time multi-dimensional spliced feature vector is input into the trained DDPG storage strategy model; the model infers based on the current state and outputs a fine-grained object storage strategy instruction, which not only includes the basic differentiated strategy based on file size, but also is intelligently optimized; specifically including target storage components and strategies, storage media, copy number, and life cycle labels; the target storage components and strategies are based on the file being medium, and the basic strategy is to use HBaseMOB block storage; the model further decides to set the block size to 8MB (instead of the default 10MB) to better match the current cluster I / O mode; the storage medium specifically balances performance and cost according to the model and decides to place the first two data blocks on the SSD of node 4 (to ensure initial access speed), and the remaining data blocks on the HDD of node 7 (to take advantage of its sufficient capacity to reduce costs); the copy number is dynamically decided by the model according to the importance of the file and the health of the cluster to create 2 copies (instead of the fixed 3 copies), which guarantees reliability while saving storage space and network bandwidth; the life cycle label is based on the historical access pattern, and the model sets the initial life cycle rule for it: "automatically converted to archive storage after 30 days";

[0161] The policy executor receives the policy instruction and translates it into a control command recognizable by the underlying big data component cluster, and this process ensures the safety and standard compatibility of the instruction through the API adaptation layer; the data storage layer executes the instruction, divides an 80MB file into blocks of 8MB size, routes the block data and its copies to the specified nodes (SSD of node 4 and HDD of node 7), uses the HBase MOB technology to store the block data into the actual data table, and establishes an index in the meta information table; the metadata management layer uses the uniformly distributed RowKey to quickly record the metadata of the file, the block index, the storage location and the life cycle rule in the optimized HBase metadata table; the fault tolerance and load balancing layer guarantees data redundancy through the replication strategy of HDFS, and balances the subsequent access requests for the file to different API service nodes through Nginx reverse proxy, ensuring service high availability (≥99.99%);

[0162] The above embodiment realizes dynamic optimization by introducing an intelligent decision-making model on the basis of inheriting the advantages of the five-layer architecture; for an 80MB video file, the system intelligently decides on a 2-copy, 8MB block and SSD / HDD hybrid storage strategy;

[0163] Finally, the metadata query delay is stabilized within 8ms, the file read / write throughput reaches 650MB / s, the storage cost is reduced by 35% compared with the fixed 3-copy SSD scheme, and through predictive life cycle management, the automation operation and maintenance efficiency is improved by more than 50%;

[0164] The embodiment shows an innovative path of combining the stable capability of traditional big data components with the leading artificial intelligence decision-making capability; by introducing an intelligent decision-making layer, the existing and mature object storage architecture can break through the limitations of static configuration and evolve into a dynamic system that can perceive the environment, make autonomous decisions and continuously optimize, thereby better coping with the increasingly complex and variable mass data storage challenges.

[0165] Further, please refer to Figure 2 An object storage service system based on a big data component cluster, for implementing the above-mentioned object storage service method based on a big data component cluster, comprising a feature extraction module, a feature preprocessing module, a model training module, a policy generation module and a policy execution module;

[0166] The feature extraction module is used to extract key features from multi-source data, including data object static features, big data component cluster dynamic environment features and historical access pattern features; the data access layer analyzes the client write request to obtain static attributes such as data size, type and user label; at the same time, the real-time system state data of the cluster is collected, and the meta database is queried to retrieve the access pattern of similar historical data, providing comprehensive input features for subsequent decision-making;

[0167] The feature preprocessing module integrates and standardizes multiple features output by the feature extraction module to generate high-quality real-time multi-dimensional splicing features; it horizontally splices data object static features, cluster dynamic environment features, and historical access pattern features, then ensures data integrity through data cleaning and missing value processing, normalizes continuous numerical features, one-hot encodes classification features, and unifies parameter ranges to output a fixed-dimensional, numerically standardized floating-point number vector, preparing for model input;

[0168] The model training module constructs and trains a storage strategy model based on the DDPG algorithm, learns optimal decisions using historical storage service data; it designs an Actor-Critic network architecture, with the Actor network processing continuous action space and the Critic network evaluating decision quality; it then constructs state-action-reward sequences from historical logs and calculates reward values using a reward function; it iteratively trains through experience replay and network parameter updates until convergence conditions are met, outputting a trained DDPG storage strategy model;

[0169] The strategy generation module inputs preprocessed real-time multi-dimensional splicing features into the trained DDPG storage strategy model, calculates and generates fine-grained object storage strategy instructions through forward propagation of the Actor network; these instructions include target storage components, storage media, replica quantities, data shard size and placement location, data encoding strategies, and initial lifecycle labels, dynamically optimizing data storage efficiency, cost, and reliability;

[0170] The strategy execution module is used to convert abstract instructions output by the strategy generation module into specific control commands recognizable by the big data component cluster, and to execute distributed storage operations. It parses and verifies instructions through a strategy executor, calls cluster management tools to generate control scripts, shards, encodes, and routes data objects to designated nodes and storage media, while updating storage location and lifecycle information in the metadata database to ensure reliable storage and efficient management of data according to the strategy.

[0171] The object storage service system based on the big data component cluster solves the core problems of policy rigidity, single optimization target and dependence on artificial experience in the traditional storage system by constructing an intelligent object storage service architecture; the system realizes a complete closed loop from data perception to policy execution: through multi-dimensional feature extraction and fusion, an accurate environment state perception is constructed; a deep reinforcement learning model is used to realize autonomous learning and dynamic optimization of the storage strategy; finally, the decision is converted into specific cluster operation through intelligent policy execution; this design enables the system to automatically realize the optimal allocation of storage resources in complex and variable big data scenarios, while achieving a precise balance between performance, cost and reliability and other multiple targets, significantly improving the storage efficiency, operation and maintenance automation level and overall system performance of large-scale data clusters.

[0172] In the description of the present specification, the description referring to the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the invention. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0173] The above disclosed preferred embodiments of the invention are only used to help explain the invention. The preferred embodiments do not describe all the details and limit the invention to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the invention, so that those skilled in the art can well understand and utilize the invention.

Claims

1. A method for object storage services based on big data component clusters, characterized in that, Includes the following steps: S1. Obtain static characteristics of data objects, dynamic environment characteristics of big data component clusters, and historical access pattern characteristics; the historical access pattern characteristics are obtained by mapping the static characteristics of data objects to a metadata database; S1 includes the following steps: S11. The data access layer receives the data object write request and verifies it. If the verification is successful, the data object write request is approved and the data object is obtained. Extract the static feature vector of the data object to obtain the static features of the data object; the static features of the data object include data size, data type, and user or application tags; S12. Collect system status data of the big data component cluster storage layer to obtain real-time big data component cluster data; extract features of the real-time big data component cluster data to obtain dynamic environment features of the big data component cluster. S13. Query the metadata database to retrieve the access patterns of historical data that have the same static characteristics as the real-time data objects, and obtain the historical access pattern characteristics. S2. The static features of the data object, the dynamic environment features of the big data component cluster, and the historical access pattern features are spliced ​​together and preprocessed to obtain real-time multi-dimensional spliced ​​features. S3. Based on the continuous action space characteristics of object storage data and the Actor-Critic architecture, construct the DDPG model; collect historical storage service data; use the historical storage service data combined with the reward function to train the DDPG model and obtain the DDPG storage strategy model. S4. Input the real-time multi-dimensional stitching features into the DDPG storage strategy model to obtain the object storage strategy instruction; S5. The object storage strategy instructions are converted into instructions that the big data component cluster can recognize, resulting in object storage control instructions; the big data component cluster performs distributed storage based on the object storage control instructions. S3 includes the following steps: S31. Construct the DDPG model; the DDPG model includes an Actor-Critic architecture; the Actor-Critic architecture includes an Actor network and a Critic network; the Actor network includes an online Actor network and a target Actor network, and the Critic network includes an online Critic network and a target Critic network. S32. Collect historical cluster status logs, data access logs, and metadata and operation logs to obtain initial historical storage service data; correlate and align the data in the initial historical storage service data to obtain the data in the historical storage service data. S33. Construct a historical state-action-reward sequence based on historical storage service data; S34. Train the DDPG model using the historical state-action-reward sequence combined with the reward function to obtain the DDPG storage policy model; step S33 includes: S331. Extract features from historical storage service data and concatenate them to obtain historical multidimensional concatenated features; the steps of extracting features from historical storage service data and concatenating them are the same as those in S1 and S2. S332. Extract the actual actions executed from the operation logs in the historical storage service data to obtain historical action data; use the historical action data as actions. S333. Define a reward function, and calculate the historical reward data obtained after executing the historical action data under the historical multi-dimensional splicing action state; calculate the access latency of data X and the load balancing of nodes A, B, and C over a period of time, and comprehensively derive a reward value Rt; the reward function is as follows: Reward value Rt = α Performance Bonus + β Cost incentive + γ Reliability Bonus - δ Penalty item; where Rt represents the reward value for performing the current action, and α, β, γ, and δ represent the performance reward weight, cost reward weight, reliability reward weight, and penalty item weight, respectively. The performance reward is obtained by adding the latency reward and the throughput reward; the latency reward = -log(actual latency / target latency), where the actual latency is the time from receiving the data read request to returning the first data byte, and the target latency is the average value based on historical normal latency data; the target latency for metadata query response is set to ≤10ms, and the target file read / write throughput can be set to 600MB / s in a 10 Gigabit network environment; the throughput reward = tanh(actual throughput / target throughput), where the actual throughput is the amount of data successfully transmitted per unit time, and the target throughput is the average value based on historical normal throughput data.

2. The object storage service method based on a big data component cluster according to claim 1, characterized in that, S12 includes the following steps: S121. Extract the CPU utilization, memory usage, disk I / O throughput, and network bandwidth utilization of each storage node in the real-time big data component cluster data to obtain the cluster load characteristics. S122. Extract the remaining capacity and current IOPS of each storage medium in the real-time big data component cluster data to obtain the storage resource characteristics; S123. Extract the network topology and node health status from the real-time big data component cluster data to obtain the topology and health status. The cluster load characteristics, storage resource characteristics, topology, and health status together constitute the dynamic environment characteristics of the big data component cluster.

3. The object storage service method based on a big data component cluster according to claim 1, characterized in that, S2 includes the following steps: S21. Horizontally stitch together the static features of the data object, the dynamic environment features of the big data component cluster, and the historical access pattern features to obtain the initial real-time multidimensional stitched features. S22. Perform data cleaning and missing data processing on the initial real-time multidimensional splicing features to obtain the initial processed real-time multidimensional splicing features. The initial processed real-time multidimensional splicing features are normalized by continuous numerical features, classification type, and parameters to obtain real-time multidimensional splicing features.

4. The object storage service method based on a big data component cluster according to claim 3, characterized in that, S33 includes the following steps: S331. Extract and concatenate features based on historical storage service data to obtain historical multidimensional concatenated features; S332. Extract the actual actions executed from the operation logs in the historical storage service data to obtain historical action data; use the historical action data as actions. S333. Set a reward function, and calculate the historical reward data obtained after executing the historical action data under the state of historical multidimensional splicing features based on the reward function; Using the historical multidimensional splicing features as state data, historical action data as action data, and historical reward data as reward data, a historical state-action-reward sequence is obtained.

5. The object storage service method based on a big data component cluster according to claim 4, characterized in that, S34 includes the following steps: S341. Set the convergence condition and training termination condition for the DDPG model; specifically, the convergence condition includes an average reward value threshold, and the training termination condition is the maximum number of training iterations; training terminates when either condition is met. S342. Randomly sample a small batch of experience data from the experience replay buffer of the historical state-action-reward sequence to obtain the first state data; The first state data is input into the online Actor network of the DDPG model to obtain the first action data for its decision; the first action data is executed in the simulation environment, and the environment will feed back the second state data. At the same time, the first reward value is calculated and obtained through the reward function. The first state data and the first action data are input into the online Critic network to obtain the evaluation value of the online Critic network for the current decision, which is recorded as the current evaluation value. The second state data is input into the target Actor network to obtain the second action data; then the second action data and the second state data are input together into the target Critic network to obtain the target evaluation value, which is recorded as the target evaluation value. The loss value is obtained by minimizing the loss function to reduce the gap between the current evaluation value and the target evaluation value; the parameters of the online Critic network are then updated based on the loss value. The online Actor network adjusts its parameters by calculating the policy gradient of the current evaluation value. The adjustment is aimed at increasing the evaluation value given by the online Critic network, that is, guiding the Actor network to make decisions that will yield higher rewards in similar situations in the future. The soft update formula is used to slowly synchronize the parameters of the online Actor network and the online Critic network to their corresponding target network in order to maintain the stability of the training process. S343. Repeat step S342. When the average reward value within the period is greater than or equal to the average reward value threshold, or when the number of training cycles reaches the set maximum number of training cycles, the training terminates and the DDPG storage strategy model is obtained.

6. The object storage service method based on a big data component cluster according to claim 1, characterized in that, S4 includes the following steps: S41. Input the real-time multidimensional stitching features into the DDPG storage strategy model. The DDPG storage strategy model performs forward propagation calculation through the Actor network based on the action features in the real-time multidimensional stitching features, and outputs object storage strategy instructions.

7. The object storage service method based on a big data component cluster according to claim 1, characterized in that, S5 includes the following steps: S51. The policy executor receives the object storage policy instruction and translates it into a specific control command that the big data component cluster storage layer can recognize, thus obtaining the object storage control instruction. S52. The big data component cluster storage layer divides the data object into fragments according to the strategy based on the received object storage control instructions, and obtains the fragmented data copies. The fragmented data copies are routed to the physical nodes and storage media specified by the policy, and the final storage location, policy information and lifecycle tags of the data objects are updated in the metadata database.

8. An object storage service system based on a big data component cluster, characterized in that, A method for implementing an object storage service based on a big data component cluster as described in any one of claims 1-7, wherein the system comprises: The feature extraction module is used to extract key features from multi-source data, including static features of data objects, dynamic environment features of big data component clusters, and historical access pattern features. The feature preprocessing module is used to horizontally stitch together the static features of data objects, the dynamic environment features of the cluster, and the historical access pattern features to obtain real-time multi-dimensional stitched features. The model training module constructs and trains a storage strategy model based on the DDPG algorithm, learning the optimal decision using historical storage service data. It designs an Actor-Critic network architecture, where the Actor network processes the continuous action space and the Critic network evaluates the decision quality. Then, it constructs a state-action-reward sequence from historical logs and calculates the reward value using a reward function. It iterative training is performed through experience replay and network parameter updates until the convergence condition is met, and the trained DDPG storage strategy model is output. The strategy generation module inputs multi-dimensional spliced ​​features into the trained DDPG storage strategy model and generates fine-grained object storage strategy instructions through forward propagation of the Actor network. The strategy execution module is used to convert the object storage strategy instructions output by the strategy generation module into specific control commands that can be recognized by the big data component cluster, and execute distributed storage.

Images