A database system data protection and disaster recovery intelligent operation and maintenance platform

Abstract

The application belongs to the technical field of database fault tolerance, and discloses a database system data protection and disaster recovery intelligent operation and maintenance platform. The platform is equipped with a distributed non-intrusive self-healing data collection probe, which can stably collect full-link time sequence data such as database kernel, business SQL, storage IO and network bandwidth. Through a Prophet, LSTM and attention mechanism fusion model, the business load fluctuation is accurately predicted, a multi-objective optimization system is constructed in combination with data grading, RTO / RPO compliance requirements, and differentiated strategies such as real-time synchronization, adaptive snapshot and full backup are customized for core, general and non-core businesses. During the load peak, the backup task is automatically degraded, and during the low peak, the protection is strengthened, and the resource scheduling is linked to dynamically allocate computing power. The platform realizes transaction-level incremental capture based on the database kernel redo / undo log native interface, accurately filters uncommitted and invalid transactions, and controls the synchronization delay within 10ms, thereby reducing the performance loss.

Description

Technical Field

[0001] This invention belongs to the field of database fault tolerance technology, specifically a smart operation and maintenance platform for database system data protection and disaster recovery. Background Technology

[0002] Data integrity and business continuity of a database system directly determine the stable operation of a digital system. Data protection and disaster recovery are core technologies for ensuring high availability of databases. Existing technologies based on scheduled snapshot backups, master-slave log synchronization disaster recovery, and threshold alarm fault monitoring have been widely applied, initially achieving basic data backup and failover capabilities. However, existing solutions still have the following technical problems: Data protection uses static, fixed configurations, which cannot be dynamically adjusted according to the full-link business load of the database, data characteristics, and business compliance requirements. This can easily lead to business disruptions due to backups competing for resources during peak periods, or insufficient protection during off-peak periods causing RPO to exceed the limit, making it difficult to achieve the optimal balance between protection effectiveness and resource consumption.

[0003] Incremental data capture and disaster recovery synchronization suffer from problems such as strong invasiveness, high latency, and insufficient anti-tampering capabilities. Trigger capture schemes degrade database performance, and general log capture cannot achieve precise filtering at the transaction level. Furthermore, logs are easily tampered with, deleted, or replayed, leading to inconsistencies between disaster recovery and production node data.

[0004] Fault monitoring relies on single indicator threshold alarms, resulting in high false alarm and false negative rates. It cannot integrate multi-dimensional indicators to mine fault characteristics. In complex scenarios, it is difficult to automatically locate the root cause and distinguish the fault type. Fault handling is highly dependent on operation and maintenance experience, which is inefficient and can easily lead to the spread of faults.

[0005] Disaster recovery relies on fixed manual processes, cannot automatically generate optimal recovery decisions based on fault scenarios, lacks automated effect verification, disaster recovery drills consume production resources, business switching interruptions are long, cannot meet the low RTO requirements of core businesses, and recovery success rates are difficult to guarantee. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent operation and maintenance platform for database system data protection and disaster recovery, so as to solve one or more problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent operation and maintenance platform for database system data protection and disaster recovery, comprising: Furthermore, the distributed non-intrusive data acquisition probe of the dynamic protection module has the ability to self-heal from faults. With optimized acquisition logic, it can collect real-time full-link time-series data of database kernel operation indicators, upper-layer business SQL execution load, business access peak characteristics, underlying storage IO performance, network link bandwidth status and business compliance level labels. A time-series prediction model integrating Prophet, LSTM and attention mechanism is adopted to focus on core business load fluctuations, predict business load fluctuation trends and peak periods. At the same time, combined with data classification results, business system preset RTO / RPO compliance requirements and compliance priorities, a multi-objective optimization model integrating protection effect, resource consumption and compliance compliance is constructed. The multi-objective optimization model takes compliance as a hard constraint and maximizes protection effectiveness while minimizing resource consumption as optimization objectives. It calculates the comprehensive score of different protection strategies and automatically selects the optimal strategy for output. After the strategy is generated, it directly sends instructions to the resource scheduling module to accurately allocate computing, storage, and network resources. Strategy adjustment and resource scheduling are executed synchronously.

[0008] Based on the multi-objective optimization model, differentiated and prioritized protection strategies are generated. Core business high-priority data adopts a protection strategy of real-time log synchronization and millisecond-level incremental verification. General business medium-priority data adopts adaptive incremental snapshots, and non-core business low-priority data adopts full backup and differential verification. At the same time, the backup task execution window is automatically adjusted. When the load increases suddenly, the system can temporarily downgrade or suspend the backup task and link the resource scheduling module to realize the dynamic allocation of backup resources.

[0009] Furthermore, the anti-tampering module implements real-time capture of transaction-level incremental data based on the database kernel-level redo / undo log native interface. After optimizing the capture logic, it has an uncommitted transaction caching mechanism to filter uncommitted transactions, rollback transactions and invalid operations, and only captures committed deterministic transaction data. For the captured transaction logs and incremental data, a dual-chain hash storage system is constructed using the national cryptographic SM3 algorithm and a custom check code. The hash value of each transaction log is associated with the hash value of the previous log, the globally unique block height, and the business identifier, forming a chain storage structure. The core transaction logs are additionally protected by SM4 encryption to enhance data security. Core transaction data is protected by the national standard SM4 symmetric encryption algorithm. The encryption key is stored securely and dynamically distributed by the platform's key management module. Encryption is performed only on committed transaction data of core business, while non-core transaction data is not encrypted to improve synchronization efficiency. The encrypted data exists in ciphertext form during both synchronization and storage. The disaster recovery node performs hash verification before decryption and restoration.

[0010] The disaster recovery node verifies the integrity of the hash chain and the custom check code in real time. When the disaster recovery node detects that the log has been tampered with, deleted or replayed, it triggers a hierarchical alarm and automatically starts the tampering tracing mechanism. It traces the tampering node, tampering time and tampering operation in reverse through the hash chain, and automatically completes the compliant data based on the historical data of the evidence. After completion, the data consistency is verified again.

[0011] Furthermore, the resource scheduling module constructs a heterogeneous disaster recovery resource management system that is compatible with mainstream relational databases, non-relational databases, and domestically developed databases. It performs standardized and refined management of servers, storage, networks, and disaster recovery nodes, and has a resource load prediction function, which can predict future resource needs based on historical load data. Resource load prediction uses historical 7-day CPU, memory, storage, network, and concurrent synchronous task data as input data. It combines business load cycle patterns to predict resource usage trends for the next 24 hours and outputs peak, valley, and average resource usage for each dimension. The prediction results are directly used for pre-scheduled resource allocation, reserving redundant resources in advance during peak resource demand periods and automatically reclaiming idle resources during off-peak periods.

[0012] To address the differences between heterogeneous databases, the disaster recovery data synchronization interface is optimized to achieve cross-database synchronization adaptation between different types of databases. The disaster recovery node has self-healing capabilities, monitors its own operating status in real time, and automatically triggers backup node switching when a node failure is detected, while adjusting the resource allocation strategy. It has a resource usage threshold early warning mechanism. When the disaster recovery resource usage exceeds the preset threshold, the system will automatically expand the resources or adjust the protection strategy.

[0013] Furthermore, the fault diagnosis module collects time-series data in real time from six dimensions: database kernel indicators, system resource indicators, business access indicators, network link indicators, disaster recovery synchronization status indicators, and security audit indicators. The fault diagnosis module has an indicator anomaly weight allocation mechanism, with core indicators having higher weight than non-core indicators, and it constructs a dynamically updated fault feature library, integrating historical fault cases and real-time anomaly data to automatically update fault features. The weights of abnormal indicators are dynamically allocated according to the level of business impact. Database kernel indicators and disaster recovery synchronization status indicators are set to high weight, system resource indicators and business access indicators are set to medium weight, and network link indicators and security audit indicators are set to low weight. When a single indicator is abnormal, the degree of impact is determined according to its weight. When multiple indicators are abnormal, the weights are added together to calculate a comprehensive abnormal score. A fault warning is triggered only when the score reaches a threshold.

[0014] A dynamic fault propagation link graph is constructed based on graph neural networks and causal inference algorithms. It can capture the changes in causal relationships between indicators in real time, explore fault propagation paths, and realize fault early warning, fault type classification and fault severity grading through Transformer fusion model. Based on the fault propagation link graph, the root cause node of the fault is automatically located and the corresponding emergency response plan is matched. It supports a graded handling mode of automated pre-handling of minor faults and manual intervention for serious faults, and links the disaster recovery module and the dynamic protection module to achieve closed-loop handling of faults.

[0015] Furthermore, the disaster recovery module, based on the root cause, type, and severity of the fault output by the fault diagnosis module, matches a preset recovery scenario library, covering typical scenarios and extreme mixed scenarios including single table accidental deletion, data file corruption, instance downtime, cluster failure, and site-level disaster. It generates the optimal recovery path based on deep reinforcement learning and can dynamically adjust the recovery steps and order according to real-time resource status and business priority. This includes optimal recovery time point selection, recovery resource scheduling, recovery step orchestration, and business switching strategies, and has a recovery path feasibility pre-verification mechanism. Extreme mixed scenarios refer to complex scenarios where multiple types of failures occur simultaneously. The platform will first break down the mixed failures into multiple single failure units, sort the recovery order according to business priority and the degree of failure impact, repair the basic failures that support the operation of core businesses first, and then deal with the secondary failures. During the recovery process, the platform will monitor the failure repair status in real time and dynamically adjust the recovery steps to avoid interference between different failure recovery processes.

[0016] Build a disaster recovery exercise sandbox that is completely isolated from the production environment and synchronized in real time. It supports non-intrusive disaster recovery exercises and pre-verification of recovery effects. The recovery operation is completed in the sandbox environment first to verify data consistency, business availability and RTO / RPO compliance. Then, the recovery of the production environment and business switching are automatically executed. Business switching is achieved through virtual IP migration, traffic gray-scale switching, and rollback mechanisms, thereby reducing the RTO of core businesses.

[0017] Furthermore, the compliance audit module logs and stores the entire process of platform operations using the national cryptographic SM3 hash; it has an abnormal operation detection mechanism, builds a normal operation behavior model based on historical operation data, and when the system detects an abnormal operation, it triggers an alarm and links the anti-tampering module to block the operation, while recording the details of the abnormal operation for tracing the source; The audit log query and analysis functions have been optimized to support multi-dimensional queries by operation type, time, operator, and business module. It can automatically generate compliance audit reports and link with the fault diagnosis module to analyze the correlation between audit logs and fault data to help locate the root cause of the fault.

[0018] The compliance audit report automatically includes six core dimensions: operational behavior statistics, abnormal operation details, data protection compliance, disaster recovery drill records, fault handling tracing, and RTO / RPO compliance status. The report presents key data such as compliance rate, abnormality rate, and compliance rate in a visual format, and also includes complete operation log hash evidence information, which can be directly used for regulatory inspections and internal compliance verification.

[0019] Furthermore, the self-learning module, based on fault cases, recovery execution results, operation and maintenance data, and compliance audit data accumulated during platform operation, combined with publicly available fault handling experience in the industry, uses transfer learning algorithms to precipitate standardized operation and maintenance experience, forming operation and maintenance guidelines and emergency plans; it has a scenario adaptive adjustment function, which can automatically adjust operation and maintenance strategies and emergency plans according to the characteristics of different industries and different database types; It collects operational data from each module in real time, automatically optimizes the operational parameters of the platform's core modules through reinforcement learning algorithms, and has a learning effect verification mechanism to periodically verify the parameters and strategies optimized through self-learning.

[0020] The scenario-adaptive adjustment automatically obtains key configuration information such as the industry attributes of the business, database version type, compliance level requirements, business load cycle characteristics, and RTO / RPO core indicators through the built-in industry feature recognition engine and database type determination engine. It automatically loads exclusive parameter configuration templates and process rules for different industries: the financial industry automatically strengthens data anti-tampering, recovery timeliness, and audit evidence preservation; the government industry automatically improves compliance audit granularity and operation traceability capabilities; and the energy industry automatically strengthens node self-healing stability and resource redundancy guarantee. It also automatically adapts log parsing rules, synchronization interface parameters, and protection policy thresholds for different database types, completing parameter loading and policy adaptation automatically throughout the entire process.

[0021] The beneficial effects of this invention are as follows: 1. This invention platform is equipped with a distributed, non-intrusive, self-healing data acquisition probe, which can stably collect full-link time-series data such as database kernel, business SQL, storage IO, and network bandwidth, avoiding collection gaps and performance losses. It accurately predicts business load fluctuations through a fusion model of Prophet, LSTM, and attention mechanisms, and constructs a multi-objective optimization system in conjunction with data classification and RTO / RPO compliance requirements. It customizes differentiated strategies such as real-time synchronization, adaptive snapshots, and full backups for core, general, and non-core businesses. During peak load periods, backup tasks are automatically downgraded, and protection is strengthened during off-peak periods. It also dynamically allocates computing power through resource scheduling to prevent backup from monopolizing business resources and causing lag.

[0022] 2. This invention platform implements transaction-level incremental capture based on the database kernel's native redo / undo log interface, accurately filtering uncommitted and invalid transactions, controlling synchronization latency within 10ms, and reducing performance loss; it uses the national cryptographic SM3 hash and a custom check code to create a dual-chain evidence storage, with core transactions overlaid with SM4 encryption, and disaster recovery nodes performing real-time verification, automatic tracing of tampering, and data completion; at the same time, it builds a heterogeneous resource management system compatible with relational, non-relational, and domestically developed databases, optimizes cross-database synchronization interfaces, and enables disaster recovery nodes to support fault self-healing and resource threshold early warning.

[0023] 3. This invention's platform collects six dimensions of operational indicators, constructs a dynamic fault feature library and propagation link map, and uses graph neural networks and Transformer models to provide early warning of faults 15-30 minutes in advance and accurately locate the root cause. It adopts hierarchical handling to achieve fault closure. It generates the optimal recovery path through deep reinforcement learning, verifies the recovery effect without disturbance through an isolated disaster recovery sandbox, and compresses the RTO of core business by combining virtual IP drift and traffic gray-scale switching. With the compliance audit module and self-learning module of national cryptographic hash evidence storage, it continuously accumulates operation and maintenance experience, optimizes model parameters, and reduces the false alarm and false negative rates and operation and maintenance costs. Attached Figure Description

[0024] Fig. 1 This is a flowchart illustrating the entire process of intelligent operation and maintenance for data protection and disaster recovery in the database system of this invention. Fig. 2 This is a flowchart of the intelligent disaster recovery execution sub-process of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] like Figs. 1-2 As shown, this embodiment of the invention provides an intelligent operation and maintenance platform for database system data protection and disaster recovery, including: In this embodiment of the invention, the distributed non-intrusive data acquisition probe of the dynamic protection module has the ability to self-heal from failure. After optimizing the acquisition logic, it can avoid the data acquisition gap caused by the failure of a single probe. It can collect full-link time-series data in real time, including database kernel operation indicators, upper-layer business SQL execution load, business access peak characteristics, underlying storage IO performance, network link bandwidth status and business compliance level labels. The probes adopt a process-level side-mounted independent deployment mode, without modifying the database kernel source code, injecting business execution threads, or preempting the database's core computing and memory resources. This achieves non-intrusive data collection from the underlying architecture and will not affect the performance of normal database operations. The probe cluster has a built-in distributed heartbeat monitoring mechanism that exchanges status information at a frequency of seconds. When a single probe node goes offline abnormally, the network is interrupted, or the service crashes, the cluster will immediately trigger an automatic master-slave failover process. The backup probe will take over the collection task within milliseconds. At the same time, a local cache resending mechanism is configured so that the collected data during the failure period will be automatically replenished after recovery, ensuring that the entire time-series data collection is uninterrupted, without packet loss, and without omissions.

[0027] A time-series prediction model integrating Prophet, LSTM, and attention mechanisms is adopted to focus on core business load fluctuations, improve the accuracy of critical business load prediction, and predict the business load fluctuation trend and peak period in the next 24 hours. At the same time, it combines data classification results, business system preset RTO / RPO compliance requirements and compliance priorities. The data classification results are divided based on three dimensions: access frequency, update frequency, and business importance. A multi-objective optimization model integrating protection effect, resource consumption, and compliance compliance is constructed. The time-series prediction model consists of a Prophet trend cycle decomposition layer, a bidirectional LSTM long-time-series feature extraction layer, a multi-head adaptive attention weight allocation layer, and a fully connected regression output layer cascaded together. Residual connections are used between layers to avoid gradient vanishing. The specific training steps for the model are as follows: ①Use full-link time-series data such as database kernel, business SQL, storage IO, and network bandwidth as input features, with the input window being the sampled data of the past 72 hours and the prediction window being the next 24 hours, to construct a supervised learning training set; ② Using the mean absolute error (MAE) as the loss function, set batch_size=32 and initial learning rate=1×10⁻⁶. -3 200 iterations, 8 attention heads; ③ The attention layer assigns higher weights to core business indicators, strengthens the extraction of key load features, and ultimately outputs the business load forecast, peak period and fluctuation range for the next 24 hours.

[0028] Data classification follows a pre-defined quantitative scoring standard. Access frequency is divided into high-frequency, medium-frequency, and low-frequency categories based on the number of daily accesses. Update frequency is divided into high-frequency, medium-frequency, and low-frequency update categories based on the number of daily data modifications. Business importance is divided into core, important, and general categories based on the business's domain, scope of impact, and compliance level. The system calculates a comprehensive level through weighted scoring, automatically classifying data into core, general, and non-core levels. Different levels of data correspond to different protection priorities and resource allocation weights.

[0029] Based on the multi-objective optimization model, differentiated and prioritized protection strategies are generated. Core business high-priority data adopts a protection strategy of real-time log synchronization and millisecond-level incremental verification. General business medium-priority data adopts adaptive incremental snapshots, with the snapshot period dynamically adjusted according to load fluctuations. The period is extended during peak load and shortened during off-peak load. Non-core business low-priority data adopts full backup and differential verification. At the same time, the backup task execution window is automatically adjusted. When the load suddenly increases, the system can temporarily downgrade or suspend the backup task and link the resource scheduling module to realize the dynamic allocation of backup resources.

[0030] The system determines peak loads based on core metrics such as CPU utilization, database connection count, and SQL execution time. Once these metrics exceed preset peak thresholds, the system automatically reduces the priority of non-core business backup tasks, extends snapshot cycles, and suspends full backup tasks, retaining only real-time synchronization and incremental verification functions for core businesses. During off-peak periods, once the metrics fall below the thresholds, the system automatically resumes all backup tasks, strengthens the protection of non-core business data, and simultaneously optimizes resource allocation to prioritize data protection tasks.

[0031] The adaptive incremental snapshot feature sets multiple period tiers. During off-peak hours, high-frequency snapshots are executed every 10 minutes; during off-peak hours, regular snapshots are executed every 30 minutes; and during peak hours, the period is automatically extended to every 60 minutes for low-frequency snapshots. The period adjustment is synchronized with three core metrics: SQL response time, database connection count, and storage I / O load. The closer the metrics are to the peak threshold, the longer the snapshot period, minimizing the resource consumption of snapshot tasks on business operations.

[0032] In this embodiment of the invention, the anti-tampering module implements real-time capture of transaction-level incremental data based on the database kernel-level redo / undo log native interface. The capture logic has been optimized and has an uncommitted transaction caching mechanism to avoid invalid data occupying synchronization resources. It filters uncommitted transactions, rollback transactions and invalid operations, and only captures committed deterministic transaction data, keeping the incremental data synchronization delay within 10ms. The uncommitted transaction caching mechanism uses a high-performance memory circular queue to temporarily store uncommitted transaction data, continuously monitoring database transaction commit and rollback status signals. Only after receiving a successful transaction commit confirmation instruction will valid data be added to the incremental synchronization queue. Uncommitted transaction data that has timed out will be automatically cleaned up and its memory space released according to preset rules. The synchronization process directly connects to the database kernel log buffer, using zero-copy technology to skip application layer parsing and data forwarding overhead. Combined with incremental data lightweight compression and streaming transmission strategies, it maximizes synchronization efficiency and stably controls the transaction-level incremental data synchronization latency to within 10 milliseconds.

[0033] For the captured transaction logs and incremental data, a dual-chain hash storage system is constructed using the national cryptographic SM3 algorithm and a custom check code. The hash value of each transaction log is associated with the hash value of the previous log, the globally unique block height, and the business identifier, forming a chain storage structure. The core transaction logs are additionally protected by SM4 encryption to enhance data security. Chain-based evidence storage strictly follows the order of database transaction commits to generate hash values ​​one by one. The hash value of each transaction log is correlated with the hash result of the previous log, the globally unique block height, and the business-specific identifier to form a complete chain structure that is bound to the previous and cannot be reversed. If any data is tampered with, deleted, or replayed, the entire hash chain verification will fail. The disaster recovery node simultaneously starts a dual mechanism of full-cycle verification and incremental real-time verification to compare the hash chain data of the production node and the disaster recovery node in real time and quickly locate the abnormal data location.

[0034] The disaster recovery node verifies the integrity of the hash chain and the custom check code in real time. When the disaster recovery node detects that the log has been tampered with, deleted or replayed, it triggers a hierarchical alarm and automatically starts the tampering tracing mechanism. It traces the tampering node, tampering time and tampering operation in reverse through the hash chain, and automatically completes the compliant data based on the historical data of the evidence. After completion, the data consistency is verified again.

[0035] The tampering tracing starts from the position where the hash chain breaks, and traverses backwards through the block height, business identifier and the data associated with the previous hash to locate the specific transaction entry that was tampered with, the corresponding disaster recovery node number, the data synchronization timestamp and the operation source information. At the same time, a complete tampering tracing report is generated, recording the tampering stage, the range of data changes and abnormal characteristics.

[0036] Automatic data completion prioritizes pulling the original log data of the corresponding transaction from the production node in real time. If there is no real-time data, it retrieves the matching complete data from the local historical evidence repository and the off-site backup repository, and rewrites it according to the original transaction commit order and chain hash rules. The completion process follows the original data format and business logic, without changing the data content and business attributes. After completion, the hash chain structure of the corresponding segment is automatically reconstructed.

[0037] In this embodiment of the invention, the resource scheduling module constructs a heterogeneous disaster recovery resource management system adapted to mainstream relational databases, non-relational databases, and domestically developed databases. It performs standardized and refined management of servers, storage, networks, and disaster recovery nodes, and has resource load prediction capabilities. It can predict future resource demands based on historical load data, and realize dynamic allocation and load balancing of disaster recovery resources. The mainstream relational databases include MySQL, Oracle, SQL Server, etc., and the non-relational databases include MongoDB, Redis, etc. Resource balancing and adaptation are based on node load rate, synchronization task volume, and network link quality. The resource carrying capacity of each disaster recovery node is calculated in real time. High-priority synchronization tasks and core business data are scheduled to low-load, high-performance nodes, while non-core tasks are allocated to nodes with sufficient resources to avoid single-node resource overload. At the same time, the number of threads and buffer size are dynamically adjusted according to database type and synchronization pressure to keep the resource utilization of each node within a reasonable range.

[0038] To address the differences between heterogeneous databases, the disaster recovery data synchronization interface is optimized to achieve cross-database synchronization adaptation between different types of databases. The disaster recovery node has self-healing capabilities, monitors its own operating status in real time, and automatically triggers backup node switching when a node failure is detected, while adjusting the resource allocation strategy. The platform has a built-in modular and scalable standardized data adaptation plugin system. The plugins have built-in log parsing rules, data structure definitions and communication protocol adaptation logic for mainstream relational databases, non-relational databases and domestically developed databases. It can automatically identify the source and target database types, automatically complete data type conversion, field mapping matching, protocol format conversion and instruction translation, and is compatible with multi-type database mixed deployment scenarios, reducing the difficulty of disaster recovery deployment and operation and maintenance in heterogeneous environments.

[0039] The node self-healing mechanism adopts a multi-dimensional status polling monitoring mode, which collects operational indicators such as node CPU, memory, storage, network, service process, and data synchronization status in real time, and confirms the validity of node failure through multiple fault judgment rules. After triggering the switch, the faulty node's routing access and task scheduling permissions are immediately cut off, the ongoing synchronization tasks and protection tasks are migrated to the backup node, the global resource allocation routing table and task execution queue are updated synchronously, and the node status is automatically verified after the switch is completed.

[0040] It has a resource usage threshold early warning mechanism. When the disaster recovery resource usage exceeds the preset threshold, the system will automatically expand the resources or adjust the protection strategy to avoid disaster recovery failure due to insufficient resources.

[0041] Resource thresholds are configured with multi-level early warning lines and alarm rules independently based on CPU utilization, memory usage, remaining storage capacity, network bandwidth utilization, and concurrent synchronous tasks. After an early warning is triggered, flexible protection strategies are implemented first to reduce resource consumption by reducing the backup frequency of non-core business and compressing the interval of non-critical data collection. If the threshold is still exceeded after flexible adjustment, the system will automatically initiate an elastic resource expansion application to quickly expand computing, storage, and network resources to ensure sufficient resource supply for core tasks such as disaster recovery synchronization, fault diagnosis, and data recovery.

[0042] In this embodiment of the invention, the fault diagnosis module collects time-series data in real time from six dimensions: database kernel indicators, system resource indicators, business access indicators, network link indicators, disaster recovery synchronization status indicators, and security audit indicators. The fault diagnosis module has an abnormal indicator weight allocation mechanism, with core indicators having higher weights than non-core indicators to avoid false alarms, and it constructs a dynamically updated fault feature library, integrating historical fault cases and real-time abnormal data to automatically update fault features. The fault feature database adopts a mechanism that combines real-time incremental updates with periodic full updates. After each fault occurs and is handled, the system automatically extracts fault indicator features, propagation paths, and handling results, and adds them to the feature database. The feature database is deduplicated, classified, and optimized daily, eliminating invalid features and strengthening high-frequency fault features, while synchronizing the latest fault case data from the industry.

[0043] Database kernel metrics include transaction count, lock wait, log write speed, and cache hit rate; system resource metrics include CPU utilization, memory usage, storage IOPS, and remaining disk space; business access metrics include connection count, request volume, response time, and failure rate; network link metrics include bandwidth utilization, packet loss rate, latency, and jitter; disaster recovery synchronization status metrics include synchronization latency, data backlog, verification success rate, and node connectivity; security audit metrics include the number of abnormal operations, login failures, permission changes, and alarms.

[0044] A dynamic fault propagation link graph is constructed based on graph neural networks and causal inference algorithms. It can capture the changes in causal relationships between indicators in real time, explore fault propagation paths, and realize fault early warning 15-30 minutes in advance, fault type classification and fault severity classification through Transformer fusion model. The fault type classification distinguishes faults into hardware faults, software faults, configuration errors, human error, malicious attacks, etc. The fault propagation path graph is constructed using six operational metrics as graph nodes and the strength of causal relationships between metrics as graph edge weights. The graph neural network employs two layers of graph convolutional GCN and one layer of global average graph pooling. Causal inference uses the PC stabilization algorithm to determine the causal direction between metrics. The specific steps of model training are as follows: input historical fault time series dataset and fault labels, learn node feature representations and edge association weights; set the causal association threshold to 0.7, filter weakly associated metrics, and retain strong causal propagation paths; aggregate neighborhood node features through graph convolution, and output the fault propagation path topology, root cause node association degree, and fault propagation probability.

[0045] The Transformer fusion model employs a 6-layer Transformer encoder structure and a 2-layer fully connected classification output layer. Input features are fused with six dimensions of index data and fault propagation map embedding features. The specific training steps are as follows: using fault type and fault severity level as classification labels, a cross-entropy loss function is used; batch_size=64 and initial learning rate=5×10 are set. -4 The system has 150 iteration rounds and 512 hidden layer dimensions. The encoder captures temporal dependencies and feature associations through a self-attention mechanism, and outputs fault warning confidence, fault type classification results, and severity grading results.

[0046] Based on the fault propagation link graph, the root cause node of the fault is automatically located and the corresponding emergency response plan is matched. It supports a graded handling mode of automated pre-handling of minor faults and manual intervention for serious faults. It links the disaster recovery module and the dynamic protection module to realize closed-loop handling of faults, reduce the reliance on manual fault handling, and prevent the spread of faults.

[0047] Fault classification is divided into three levels: minor, moderate, and severe, based on the scope of impact, business interruption risk, and data security risk. Minor faults are characterized by single-indicator anomalies, no business impact, and no data risk. The system automatically performs pre-emptive actions such as restarting the service, clearing the cache, and adjusting parameters. Moderate faults are characterized by multiple-indicator anomalies, localized business delays, and minor data risks. The system automatically performs fault isolation and initial repair. Severe faults are characterized by node downtime, business interruption, and data corruption risks. The system immediately triggers an alarm and locks manual intervention permissions, waiting for confirmation from operations and maintenance personnel before performing subsequent recovery operations. The classification and handling rules can be adaptively adjusted according to the business scenario.

[0048] Root cause localization first locates the source node of abnormal indicators through fault propagation map, then filters core influencing indicators by combining weight rules, eliminates secondary abnormal interference, then confirms the root cause type by comparing with fault feature library, and finally outputs three key pieces of information: root cause node location, abnormal indicator details, and fault triggering cause. The localization process is fully automated, and the localization results are directly used to match the response plan.

[0049] In this embodiment of the invention, the disaster recovery module, based on the root cause, type, and severity of the fault output by the fault diagnosis module, matches a preset recovery scenario library, covering typical scenarios and extreme mixed scenarios such as single table accidental deletion, data file corruption, instance downtime, cluster failure, and site-level disaster. It generates the optimal recovery path based on deep reinforcement learning and can dynamically adjust the recovery steps and order according to real-time resource status and business priority. It includes optimal recovery time point selection, recovery resource scheduling, recovery step orchestration, and business switching strategy. It has a recovery path feasibility pre-verification mechanism to avoid invalid recovery operations. A Deep Q-Network (DQN) reinforcement learning model is adopted. The state space is defined as fault scenario type, RTO / RPO compliance indicators, real-time resource utilization, and business priority level. The action space is defined as recovery step selection, recovery resource scheduling, and business switching strategy. The reward function is set as the compliance achievement rate minus the weighted value of resource occupancy and recovery time. The specific steps of model training are as follows: a fault scenario library is built in the disaster recovery exercise sandbox for simulation training, with a total of 300 training iterations; the initial exploration rate ε=0.1 is set, and an experience replay mechanism is adopted to improve training stability; the model learns the optimal decision through continuous trial and error, and outputs the optimal recovery path that meets RTO / RPO requirements, has the lowest resource occupancy, and the shortest recovery time.

[0050] The recovery path feasibility pre-verification includes several core checks. First, it verifies the integrity and availability of backup data, whether the disaster recovery node resources meet the recovery requirements, and whether the network link is unobstructed. Then, it verifies whether the recovery steps match the business priorities, whether the recovery time meets the RTO requirements, and whether the data after recovery meets the consistency standards. The recovery process can only be executed after all checks are passed. If any check fails, the recovery path will be automatically regenerated until all checks are passed.

[0051] Build a disaster recovery exercise sandbox that is completely isolated from the production environment and synchronized in real time. It supports non-intrusive disaster recovery exercises and pre-verification of recovery effects. The recovery operation is completed in the sandbox environment first to verify data consistency, business availability and RTO / RPO compliance. Then, the recovery of the production environment and business switching are automatically executed. The disaster recovery exercise sandbox is deployed using independent virtualized containers and physically isolated resources, and is completely isolated from the network links, storage media and computing nodes of the production environment. It synchronizes production environment data in real time through incremental logs to maintain a high degree of consistency between sandbox data and production data. All disaster recovery exercises, recovery simulations and policy testing operations are executed in a closed loop within the sandbox, without writing any data to the production environment, occupying production business resources, or affecting the normal operation of the production system.

[0052] Service switching is achieved through virtual IP migration, traffic canary switching, and rollback mechanisms, avoiding service interruptions caused by switching failures and reducing the RTO of core services.

[0053] Before switching services, data consistency verification and service availability checks are performed on the disaster recovery node. Once confirmed to be correct, virtual IP binding and migration are executed to switch the service access point to the disaster recovery node. Then, following the principle of prioritizing core business, business traffic is gradually switched from the production node to the disaster recovery node in a small-scale, gradual, and gray-scale manner, while simultaneously monitoring availability indicators such as business response speed, success rate, and number of connections in real time. If business anomalies, data inconsistencies, or service lag occur during the switchover process, a one-click rollback mechanism for all traffic is immediately triggered to quickly switch all traffic back to the original production node, maximizing the continuity of business operations and minimizing the interruption time of core business operations.

[0054] In this embodiment of the invention, the compliance audit module logs and stores the entire process of platform operations using the national cryptographic SM3 hash. These operations include adjusting protection strategies, executing backup tasks, synchronizing incremental data, conducting disaster recovery drills, fault recovery, and resource scheduling. The module also features an abnormal operation detection mechanism. Based on historical operation data, it constructs a normal operation behavior model. When the system detects abnormal operations such as illegally adjusting protection strategies or maliciously deleting backup data, it triggers an alarm and activates the anti-tampering module to block the operation. Simultaneously, it records the details of the abnormal operation for source tracing. The abnormal operation detection model is trained based on the platform's long-term accumulated historical normal operation data to generate a standard behavior baseline. It extracts features from multiple dimensions such as operation time, operator permissions, compliance of operation content, operation frequency, operation object, and operation terminal location for comparison and judgment. Any operation that deviates from the normal behavior baseline will be identified in real time. The system will trigger alarms according to the abnormal risk level and simultaneously link the anti-tampering module to immediately intercept the illegal operation, and retain the details and feature data of the abnormal operation throughout the process.

[0055] The entire process operation log uses the national cryptographic SM3 algorithm to generate hash values ​​for each log entry. Each log hash value is associated with and bound to the operation time, operator, operation content, and business module to generate tamper-proof evidence data. Log data is archived and stored in a hierarchical manner according to daily, weekly, and monthly standards. After archiving, a hash verification value is automatically generated to prevent the archived logs from being tampered with. During retrieval, log data can be quickly located through multi-dimensional conditions, while the integrity of the hash value is automatically verified.

[0056] The audit log query and analysis functions have been optimized, supporting multi-dimensional queries by operation type, time, operator, business module, etc. It can automatically generate compliance audit reports and link with the fault diagnosis module to analyze the correlation between audit logs and fault data, helping to locate the root cause of faults such as human error and malicious attacks.

[0057] In this embodiment of the invention, the self-learning module is based on fault cases, recovery execution results, operation and maintenance data, and compliance audit data accumulated during platform operation, combined with publicly available fault handling experience in the industry, and uses transfer learning algorithms to precipitate standardized operation and maintenance experience, forming operation and maintenance guidelines and emergency plans; it has a scenario adaptive adjustment function, which can automatically adjust operation and maintenance strategies and emergency plans according to the characteristics of different industries such as government affairs, finance, and energy, and different database types; Transfer learning employs a pre-training-fine-tuning paradigm, based on a general fault handling pre-trained model. The underlying feature extraction layer is frozen, and the top-level industry adaptation layer is fine-tuned. The learning rate is set to 1 × 10⁻⁶. -4 It can quickly adapt to the business characteristics and database types of different industries such as government affairs, finance, and energy, and transform general operation and maintenance experience into scenario-specific strategies.

[0058] Scenario adaptation identifies four key pieces of information—business industry attributes, database type, compliance level, and load characteristics—and automatically matches corresponding industry operation and maintenance parameters and emergency procedures. For the financial industry, it strengthens data security and rapid recovery strategies; for the government industry, it strengthens compliance audit and evidence storage requirements; and for the energy industry, it strengthens node self-healing and resource stability assurance.

[0059] The system collects real-time operational data from each module, such as prediction accuracy, fault diagnosis accuracy, and recovery success rate. It automatically optimizes the operational parameters of the platform's core modules through reinforcement learning algorithms, such as the attention weights of the time-series prediction model and the association thresholds of the GNN fault propagation graph. It also has a learning effect verification mechanism, which periodically verifies the parameters and strategies optimized by self-learning to ensure that the optimization effect meets the standards.

[0060] Reinforcement learning uses model prediction accuracy, fault diagnosis accuracy, and recovery success rate as positive rewards, and false alarm rate and resource consumption as negative rewards. It iteratively optimizes core hyperparameters such as attention weights of time series prediction models, GNN fault propagation graph association thresholds, and Transformer encoder parameters.

[0061] The self-learning effect verification adopts a combination of comparative testing and scenario simulation. The optimized model parameters, protection strategies, and emergency plans are regularly compared with the previous version to test core indicators such as load prediction accuracy, fault warning accuracy, recovery success rate, and resource utilization. At the same time, various fault scenarios are simulated in the disaster recovery exercise sandbox to verify the execution effect of the optimized strategy. If the indicators improve and there are no abnormal issues, the optimization is confirmed to be effective. If the effect does not meet the standards, the parameters of the previous version are rolled back and the learning is iterated again.

[0062] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0063] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart operation and maintenance platform for database system data protection and disaster recovery, characterized in that, include: The dynamic protection module is used to collect runtime sequence data of the entire database chain, predict business load fluctuations, generate differentiated protection strategies by combining data classification results and compliance requirements, and link with the resource scheduling module to realize dynamic allocation of backup resources. The anti-tampering module is used to achieve real-time capture and effective filtering of transaction-level incremental data, build a dual-chain hash evidence storage system, and ensure the consistency, integrity and immutability of disaster recovery data and production data through real-time verification of disaster recovery nodes, tamper tracing and data completion mechanisms. The resource scheduling module is used to build a heterogeneous disaster recovery resource management system that is compatible with various types of databases, realize resource load prediction, dynamic allocation and balanced adaptation, has disaster recovery node self-healing and resource threshold early warning functions, and optimizes the data synchronization interface. The fault diagnosis module is used to collect multi-dimensional operating indicators from the database in real time, build a dynamically updated fault feature library and dynamic fault propagation link map, realize early warning, classification and root cause location of faults through intelligent prediction model, and realize closed-loop fault handling by linking related modules through hierarchical handling mode. The disaster recovery module is used to match various recovery scenarios based on fault diagnosis results, generate the optimal recovery path and verify its feasibility, and complete the recovery verification using a disaster recovery exercise sandbox isolated from the production environment. The compliance audit module is used to log and hash-store the entire process of platform operations. It has functions such as abnormal operation detection, multi-dimensional log query and compliance report generation, and links with related modules to achieve collaborative protection and fault tracing. The self-learning module is used to accumulate operational experience and optimize operating parameters based on platform operation data and industry experience through machine learning algorithms. It has the ability to adapt to different scenarios and regularly verify the optimization effect.

2. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 1, characterized in that, The dynamic protection module uses a distributed, non-intrusive data acquisition probe with fault self-healing capabilities to collect real-time time-series data across the entire link, including database kernel operation metrics, upper-layer business SQL execution load, business access peak characteristics, underlying storage IO performance, network link bandwidth status, and business compliance level labels. The dynamic protection module uses a time-series prediction model that integrates Prophet, LSTM, and attention mechanisms to predict business load fluctuation trends and peak periods. Combined with data classification results, the business system's preset RTO / RPO compliance requirements and compliance priorities, it constructs a multi-objective optimization model that integrates protection effectiveness, resource consumption, and compliance compliance. Based on the multi-objective optimization model, it generates differentiated and prioritized protection strategies, automatically adjusts the backup task execution window, and links with the resource scheduling module to achieve dynamic allocation of backup resources.

3. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 2, characterized in that, The anti-tampering module uses the database kernel-level redo / undo log native interface to achieve real-time capture of transaction-level incremental data, configures an uncommitted transaction caching mechanism, filters uncommitted transactions, rollback transactions and invalid operations, and only captures committed deterministic transaction data. For the captured transaction logs and incremental data, the anti-tampering module uses the national cryptographic SM3 algorithm and a custom verification code to construct a dual-chain hash evidence storage system. The hash value of each transaction log is associated with the hash value of the previous log, the globally unique block height, and the business identifier, forming a chain evidence storage structure. The disaster recovery node verifies the integrity of the hash chain and the custom verification code in real time. When the verification detects data anomalies, it triggers a hierarchical alarm, simultaneously starts the tampering tracing mechanism to trace the abnormal information, and automatically completes the compliant data based on the historical data of the evidence storage. After completion, the data consistency is verified again.

4. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 3, characterized in that, The resource scheduling module constructs a heterogeneous disaster recovery resource management system that is compatible with mainstream relational databases, non-relational databases, and domestically developed databases. It performs standardized and refined management of servers, storage, networks, and disaster recovery nodes, and predicts future resource needs based on historical load data. The resource scheduling module optimizes the disaster recovery data synchronization interface for the differences between heterogeneous databases, enabling cross-database synchronization and adaptation between different types of databases; the disaster recovery node has the ability to self-heal from failure, monitors its own operating status in real time, and automatically triggers the switchover of the backup node and adjusts the resource allocation strategy when a node failure is detected; a resource occupancy threshold early warning mechanism is configured, and when the disaster recovery resource occupancy exceeds the preset threshold, the resource is automatically expanded or the protection strategy is adjusted.

5. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 4, characterized in that, The fault diagnosis module collects time-series data in real time from six dimensions: database kernel indicators, system resource indicators, business access indicators, network link indicators, disaster recovery synchronization status indicators, and security audit indicators. It also configures an abnormal weight allocation mechanism for indicators, with core indicators having higher weights than non-core indicators. The fault diagnosis module constructs a dynamically updated fault feature library, integrates historical fault cases and real-time abnormal data, automatically updates fault features, and constructs a dynamic fault propagation link graph based on graph neural networks and causal inference algorithms to capture changes in causal relationships between indicators in real time and explore fault propagation paths.

6. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 5, characterized in that, The fault diagnosis module uses the Transformer fusion model to achieve early warning of faults, classification of fault types, and severity grading of faults. Based on the fault propagation link graph, it automatically locates the root cause node of the fault and matches the corresponding emergency response plan. It supports a graded handling mode of automated pre-handling of minor faults and manual intervention for severe faults, and links with the disaster recovery module and dynamic protection module to achieve closed-loop fault handling.

7. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 6, characterized in that, The disaster recovery module is based on the root cause, type and severity of the fault output by the fault diagnosis module, and matches it with a preset recovery scenario library, covering typical scenarios and extreme mixed scenarios including single table accidental deletion, data file corruption, instance downtime, cluster failure, and site-level disaster. The disaster recovery module generates the optimal recovery path based on deep reinforcement learning. It can dynamically adjust the recovery steps and order according to real-time resource status and business priority, and configure a recovery path feasibility pre-verification mechanism. It builds a disaster recovery exercise sandbox that is completely isolated from the production environment and synchronized in real time, supports non-intrusive disaster recovery exercises and recovery effect pre-verification, and achieves smooth business switching through virtual IP drift, traffic gray-scale switching and rollback mechanisms.

8. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 7, characterized in that, The compliance audit module logs and stores the entire process of platform operations using the national cryptographic SM3 hash, configures an abnormal operation detection mechanism, builds a normal operation behavior model based on historical operation data, triggers an alarm when an abnormal operation is detected and links the anti-tampering module to intercept the illegal operation, and records the details of the abnormal operation for subsequent tracing. The compliance audit module supports multi-dimensional log queries based on operation type, operation time, operator, and business module. It can automatically generate compliance audit reports and link the audit logs with the fault diagnosis module to analyze the correlation between the audit logs and fault data, helping to locate the root cause of the fault.

9. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 8, characterized in that, The self-learning module is based on the fault cases, recovery execution results, operation and maintenance data, and compliance audit data accumulated during the platform operation process. Combined with publicly available fault handling experience in the industry, it uses transfer learning algorithms to precipitate standardized operation and maintenance experience and form standardized operation and maintenance guidelines and emergency plans. The self-learning module has a scenario-adaptive adjustment function, which can automatically adjust operation and maintenance strategies and emergency plans according to the business characteristics of different industries and different database types.

10. The intelligent operation and maintenance platform for database system data protection and disaster recovery according to claim 9, characterized in that, The self-learning module collects real-time operating data from each module of the platform, automatically optimizes the operating parameters of the platform's core modules through reinforcement learning algorithms, configures a learning effect verification mechanism, and periodically verifies the effectiveness of the self-learned optimized operating parameters and protection strategies.

Images