AI-based data management methods for data centers

Abstract

This invention relates to an artificial intelligence-based data center data management method. It records data access methods, triggering entities, and time distribution within a preset period. Based on behavioral trajectories, interaction characteristics, and load changes, it determines the data state variation rate. When a behavioral transformation threshold is reached, it generates a data target processing path sequence. Combining the elasticity range, resource critical points, and real-time occupancy of the data center's current storage and computing resources, it assesses the resource stress of each path. By monitoring data behavior fluctuation trends, it sets delayed execution for tasks that are stabilizing and freezes tasks with continuous high fluctuations, triggering actual execution of the processing plan only when the fluctuation difference exceeds a threshold window. It sets a conflict detection window for executed actions, triggering reverse undo logic when concurrent conflicts, access anomalies, or resource overload are detected, rolling back the most recent non-mandatory task, and updating the state evaluation values ​​of tasks with paths related to the data.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence data management technology, specifically to a data center data management method based on artificial intelligence. Background Technology

[0002] Existing technologies, such as Chinese patent document CN112468597A, describe an AI-based data center resource allocation method and apparatus. This method focuses on perceiving and responding to device behavior within a target cache node range in mobile scenarios. It uses AI algorithms to identify the number of data requests sent by the device, request identifiers, and the device's dwell time within the cache node's coverage area, thereby determining the target data type and pushing cached content. While this solution can improve the user's data acquisition experience in mobile situations to some extent, it still has several key technological shortcomings and implementation drawbacks in terms of the overall intelligent management system for data centers. First, this patent focuses on analyzing data request behavior for terminal devices, and its processing logic is mainly based on local cache hit optimization, lacking the ability to manage global resources at the data center level. Data center operation involves numerous dimensions such as task scheduling, dynamic resource allocation, concurrent request processing, anomaly prediction and intervention. Relying solely on the frequency of requests and dwell behavior of devices within a certain time period to predict data type and cache target data to nodes cannot adapt to resource management and behavior scheduling scenarios for large-scale heterogeneous tasks, and cannot cover the processing logic of complex dependency chains under multi-source data interaction.

[0003] Secondly, this solution relies on static feature information from the device side to deduce target data requirements, and has not yet introduced a dynamic context feature judgment mechanism, such as the suddenness of data requests, the distribution pattern of access time, and the frequency of task switching, lacking intelligent recognition of data access behavior transformation trends. Without weighted judgment of behavioral trajectory trends, access load changes, and sudden behaviors, relying solely on the statistics of data request counts and type labels is prone to risks such as misjudgment, incorrect execution of low-frequency paths, or delayed response of high-priority tasks, limiting its generalization ability and the scope of applicability of its response mechanism. Thirdly, this method does not set up a fluctuation detection mechanism and a lazy decision buffer strategy, and has no adaptive intervention capability for sudden changes in resource fluctuations, rapid changes in access frequency, and task execution order under critical resource conditions. For example, when a certain type of data experiences explosive growth in a short period of time, its caching strategy may lag behind the change in access patterns, causing data to be unable to respond during hot periods, resulting in access congestion and redundant bandwidth resource overhead, thereby reducing the overall system efficiency.

[0004] Furthermore, this method lacks conflict identification and reverse rollback mechanisms. When access fails due to incorrect cache push decisions or abnormal resource allocation, it cannot roll back and repair the data operation path, nor does it involve the construction of a task dependency graph or a mechanism for identifying reversible nodes. This means that once a decision path fails, such as mistakenly pushing non-priority data or repeatedly caching highly saturated data, the system cannot achieve fine-grained reversal and replacement based on task dependencies. It can only passively wait for the next scheduling cycle to readjust, increasing the probability of resource waste and response latency. Finally, regarding resource evaluation, this method does not introduce a tension measurement system and cannot measure the impact of task paths on system bandwidth, CPU utilization, and storage ratio in real time. When multiple task paths enter the execution phase simultaneously and the resource redundancy range is limited, the system cannot prioritize critical path tasks or formulate dynamic decision-making logic based on resource sensitivity and dependency complexity. Therefore, it is not conducive to achieving globally optimal resource allocation in high-concurrency scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based data center data management method, thereby addressing some of the drawbacks and shortcomings pointed out in the background art.

[0006] The present invention adopts the following technical solution to solve the above-mentioned technical problems: a data center data management method based on artificial intelligence, including: recording the access method, triggering subject and time distribution of data within a preset period, and judging the data state variation rate based on behavior trajectory, interaction characteristics and load changes, and generating a data target processing path sequence when the variation rate reaches the behavior transformation threshold;

[0007] Based on the path sequence, combined with the elasticity range, resource critical point, and real-time occupancy of the current storage and computing resources in the data center, the stress level of each path on resources is assessed, and the execution, delay, or replacement of the path is determined by the stress trade-off rules to form an actual processing plan. The processing plan is temporarily stored in the inertial control pool. By monitoring the fluctuation trend of data behavior, the execution of tasks that are stabilizing is delayed, and tasks with continuous high fluctuations are frozen. The actual execution of the processing plan is only triggered when the fluctuation difference exceeds the threshold window.

[0008] Set a conflict detection window for executed actions. When a concurrent conflict, access anomaly, or resource overload is detected, trigger reverse undo logic to roll back the most recent non-mandatory task and update the status evaluation value of tasks with paths related to the data.

[0009] Furthermore, the determination of the data state variation rate includes a weighted calculation of the access time interval variance, request burstiness, and trigger subject switching frequency; the behavior transformation threshold is adjusted according to the stability of the access pattern of historical data within the same period; the generated data target processing path sequence includes at least one of the following: cache boosting path, storage migration path, access rate limiting path, or cleanup and recycling path.

[0010] Furthermore, the elastic range of the resource is expanded or contracted in real time according to the system load fluctuation rate; the resource critical point is dynamically recalibrated based on the frequency of historical triggering of the threshold; the path tension level is obtained by calculating the increase in concurrency ratio, the rate of increase in storage bandwidth, and the change in CPU usage caused by path execution.

[0011] Furthermore, the tension trade-off rules include the judgment logic of prioritizing high-sensitivity data paths, reducing high-load paths, and replacing low-priority paths; the lazy control pool constructs multiple sub-pools according to different types of tasks, which are respectively used to handle caching, migration, replication, or cleanup tasks; the data behavior fluctuation trend is obtained by comparing the access growth rate, access decline rate, and pause time changes in adjacent periods.

[0012] Furthermore, the condition that the fluctuation difference exceeds the threshold window is determined based on the absolute value of the difference between the fluctuation values ​​of the current window and the previous window; the access anomaly is achieved by identifying the access error rate, timeout rate, and abnormal access distribution; the reverse cancellation logic includes three rules: cancellation operation identification, cancellation impact area positioning, and cancellation range limitation.

[0013] Furthermore, by jointly weighting the access growth rate, access decay rate, and access interval change, the fluctuation difference value is calculated using a non-stationary integral enhancement function, expressed as:

[0014] ;

[0015] in:

[0016] For at any time The obtained access fluctuation differential enhancement amount; This represents the length of the fluctuation observation window. For time The rate of increase in access to the location; For time Access decay rate at the location; For time The change in the interval of access at a location; , , These are the weighting parameters corresponding to the growth rate, decay rate, and intermittent change, respectively. A time decay enhancement kernel function, used for near-time... The contribution of fluctuations is amplified by weight; For integration time;

[0017] The system employs an integral enhancement method to quantify nonlinear fluctuations in access behavior within a window, thereby improving the ability to identify critical changes in access patterns. The access error rate is obtained by statistically analyzing the ratio of erroneous requests to total requests per unit time. The access timeout rate is obtained by accumulating the number of access requests exceeding a preset time limit. Abnormal access distribution is determined by comparing the magnitude of deviation of access requests from the normal access curve on the time axis and the load deviation ratio between source nodes. Furthermore, in the event of anomalies or conflicts, based on the determination of the type, direction, and strength of dependencies between task nodes within the scheduling link, reversible execution nodes are identified, and nodes with strong dependencies are skipped, thus achieving a controllable reverse undo operation.

[0018] Furthermore, the acquisition of the access growth rate includes calculating the growth ratio based on the difference in access volume between two consecutive access cycles; the access interval change is obtained by statistically analyzing the time interval between two adjacent accesses and calculating its change magnitude; the statistics of the number of erroneous requests are determined based on a comprehensive analysis of HTTP status codes, protocol error types, and verification failure results.

[0019] Furthermore, the degree of deviation of the abnormal access distribution is determined by comparing the magnitude of the periodic deviation of the access request from the normal access curve on the time axis; the judgment of the degree of deviation of the source node includes comparing the load deviation ratio of the access request on different nodes; the identification of reversible execution nodes in the scheduling link includes determining the type, direction and strength of the dependency relationship between nodes.

[0020] Furthermore, the normal access curve is obtained by smoothing and fitting historical multi-period access data to provide a time offset benchmark; the determination of the magnitude of the periodic deviation includes magnitude weighting of the difference between the abnormal access peak and the normal curve peak; the load deviation ratio of the node is obtained by comparing the node access volume with the average node access volume, and different deviation tolerances are set according to the service type. When the load deviation ratios of multiple nodes exceed the tolerance range at the same time, the access mode is identified as node-level unbalanced access anomaly.

[0021] Furthermore, the determination of the type of dependency relationship includes identifying three types of relationships: data dependency, execution order dependency, and resource binding dependency; the determination of the dependency direction is based on the sequential triggering order between tasks and is used to generate a dependency graph sequence table of reversible nodes.

[0022] The beneficial effects of this invention are as follows: By introducing an AI-based data behavior cognition mechanism, this invention achieves real-time perception and fine-grained differentiation of massive data center access patterns. Through the construction of a series of intelligent logics, including data state variation rate, behavior transition threshold, adaptive path sequence generation, resource tension assessment, and lazy control execution, this invention can predict task trends in advance when data access behavior changes, avoiding the lag problem of post-processing in traditional rule-based systems. Simultaneously, by automatically adjusting the elastic range and resource critical points according to system load fluctuations, resource scheduling no longer relies on fixed thresholds, thus maintaining stable operation under high concurrency, sudden traffic surges, or resource-constrained environments. Compared with existing static scheduling strategies, this invention can automatically delay processing when data behavior stabilizes, automatically freeze tasks during periods of sustained high volatility, and only trigger execution when the fluctuation difference threshold is truly reached, effectively reducing redundant migration, ineffective scheduling, and resource oscillations.

[0023] Furthermore, this invention elevates anomaly detection from single-indicator identification to a multi-feature, high-precision comprehensive judgment system through the collaborative judgment of access error rate, access timeout rate, and abnormal access distribution. In particular, the joint analysis of time axis offset and node load deviation can accurately capture potential node-level imbalances, malicious access source proliferation, or link congestion issues, fundamentally improving the security and robustness of data center operations. In addition, the reversible node identification and reverse rollback mechanism designed in this invention can roll back the task link with minimal impact when conflicts or overloads are detected, strictly skipping nodes with strong dependencies, thereby avoiding damage to data consistency and task integrity. Through comprehensive analysis of dependency type, dependency direction, and dependency strength, this invention can construct a stable, self-healing data processing flow, achieving closed-loop intelligent management of the entire process from behavior identification, resource assessment, execution control to anomaly rollback, significantly improving the automation level, system stability, and overall resource utilization efficiency of the data center. Attached Figure Description

[0024] Figure 1 This is a simplified logic judgment diagram for intelligent data center management in this invention.

[0025] Figure 2 This is a diagram illustrating the intelligent scheduling function driven by data access behavior in this invention.

[0026] Figure 3 This is a logic diagram for access anomaly identification and reversible rollback determination in this invention.

[0027] Figure 4 This is a flowchart of the artificial intelligence data scheduling process for peak sales periods in Embodiment 1 of the present invention.

[0028] Figure 5 This is a diagram showing the relationship between the order log data anomaly detection and intelligent dependency rollback function in Embodiment 2 of the present invention. Detailed Implementation

[0029] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] Combined with appendix Figure 1 This invention presents an artificial intelligence-based data center data management method that utilizes an access behavior perception mechanism and adaptive path planning to achieve dynamic management and precise control of the data lifecycle. An access monitoring module is deployed in the data center operating environment to continuously record the distribution of access methods, triggering entities, and access times for various types of data within a preset period. The access methods include operations such as read, write, delete, and migrate; the triggering entities include end users, upper-layer applications, automatic schedulers, or third-party interface requests; and the access time distribution includes time-dimensional indicators such as daily access concentration, periodic fluctuation characteristics, and the frequency of sudden access events. By continuously collecting the above access behavior data and constructing behavior trajectories, a behavior state recognition unit is further introduced to model and analyze the data usage patterns on the time axis. This modeling process combines interaction characteristics and load change indicators for comprehensive judgment. Interaction characteristics refer to the strength of the coupling relationship between data and other tasks or applications, while load change indicators reflect the resource occupancy trend and change magnitude corresponding to the data. Based on the acquired behavior sequence, this invention defines the data state variation rate as a quantitative parameter to measure the drastic change in the current data usage state. This variation rate is calculated by weighting three indicators: the variance of the access time interval, the burstiness of requests, and the frequency of trigger subject switching. When this variation rate reaches a preset behavior transition threshold within a certain period, the system determines that the data usage characteristics are shifting from a stable period to a sudden change period or from regular access to hotspot access, and accordingly activates the path prediction module to generate a target processing path sequence. The path sequence is the optimal operation path given based on the prediction results of the current data state and future trends. Specifically, it may include various optional schemes such as cache capacity improvement path, storage media migration path, access rate limiting path, and data cleanup or recycling path. This path will be used to drive subsequent resource scheduling and execution processes.

[0031] The system calls the resource awareness module to obtain real-time resource availability parameters, including storage capacity, I / O channel bandwidth, processor computing power, and memory usage. It then compares the current load of each type of resource with its historical load trends to construct a resource elasticity range model. This elasticity range describes the tolerance range of various resources under different load fluctuation conditions. If the current resource usage state approaches or exceeds the boundary of this elasticity range, it indicates that the current resource is in a high-tension state. Simultaneously, the system references resource critical point setting logic. This critical point is a threshold dynamically calibrated by statistically analyzing the frequency of resource overload events in historical periods to identify key nodes. Next, the system maps the potential impact of each path on resources to path tension levels. These tension levels are quantified based on factors such as the increase in resource concurrency, the rate of increase in bandwidth usage, and the increase in processor load during path execution, and serve as important input indicators for path determination.

[0032] After obtaining the tension intensity of all candidate paths, the system introduces a tension trade-off rule base to determine the feasibility of each path. This rule base includes multiple decision strategies, such as prioritizing high-sensitivity task paths, dynamically reducing the execution frequency of paths under high tension, and skipping or replacing low-priority paths. Based on the decision results, a final actual processing plan is formed. This plan records the current state of each task path, including whether it should be executed immediately, delayed, or marked as replaceable. To avoid system resource oscillations due to frequent task execution within a short period, this invention introduces a lazy control mechanism, temporarily storing the processing plan in a lazy control pool before execution. The lazy control pool caches all tasks to be executed and monitors the behavioral fluctuation trends of the data corresponding to each task. These fluctuation trends are obtained by comparing the continuous periodic access growth rate, access decline rate, and access interval variation magnitude. If the data of a task exhibits a stable behavior, the system marks it as a delayed execution task and sets a timed reassessment point. If a task exhibits a continuously high-fluctuation behavior, the system freezes it and suspends its entry into the scheduling process. The system will only trigger the actual scheduling and execution of a task's processing plan when the absolute difference between the behavior fluctuation difference value of a task and the previous period exceeds a preset threshold window.

[0033] To further enhance the system's responsiveness to abnormal states and prevent task execution results from causing continuous system resource overload or link-level access conflicts, a conflict detection window is set up after each processing task enters the execution phase. This conflict detection window is a behavior observation period of a limited duration, designed to monitor the impact of task execution on the system's operational status, including resource load, data path response status, and other concurrent task interactions. During conflict detection, the system continuously acquires multiple operational metrics, including sudden increases in access queue length, abnormal increases in access error rate, increased processing task response latency, and data node access distribution offsets. Multi-dimensional feature fusion is used to determine whether concurrent conflicts, access anomalies, or resource overload exist.

[0034] When the system's operating status is detected to meet the set conflict judgment conditions, the system will immediately initiate reverse rollback logic. This logic first retrieves the execution path and related operation links of the current task through a dependency tracking mechanism, identifying the most recently rollbackable non-mandatory task. A non-mandatory task refers to an operation node that does not have a unique dependency structure in the execution strategy and is tolerable for not interrupting the business process. After identifying the task, the system rolls it back to its pre-execution state. During the rollback process, other task paths with direct or indirect data coupling relationships with the original task are also located. The system re-evaluates the scheduling status, resource requests, and behavioral trends of tasks in these related paths and updates their status evaluation values ​​accordingly. The status evaluation values ​​reflect the task's execution priority, execution risk level, and system load adaptability. The updated evaluation results are fed back to the scheduling control center for subsequent scheduling decision reference.

[0035] Combined with appendix Figure 2 To achieve intelligent judgment and path guidance on changes in data usage status, the system performs structured analysis of data access behavior and introduces a data status variation rate evaluation mechanism. This variation rate characterizes the intensity of changes in data usage patterns within the current period relative to previous periods. Its calculation involves a weighted fusion of three indicators: access time interval variance, request burstiness, and trigger subject switching frequency. Access time interval variance reflects the temporal density changes in data access, burstiness describes the degree to which requests arrive in a concentrated manner within a short period, and trigger subject switching frequency reflects the frequency of data alternation between different callers. The system assigns adjustable weight parameters to each of the three indicators and calculates the weighted composite value as the data's status variation rate within the current time period.

[0036] To ensure dynamic adaptability and contextual relevance in the judgment, this invention establishes a behavior transition threshold adjustment mechanism. The system periodically adjusts the transition threshold based on the stability of access patterns for this type of data within the same period in historical data. If a type of data exhibits high stability historically, its threshold will be relatively increased to prevent misjudgment of the transition status due to minor fluctuations. Conversely, if historical behavior shows significant fluctuations, the system will automatically lower the transition judgment threshold for that data to improve response sensitivity.

[0037] After determining the state mutation rate and confirming that the data behavior has entered a transition state, the system will activate the path generation module to generate a target processing path sequence based on the current data usage trend, resource status, and scheduling strategy. This path sequence serves as a recommendation for subsequent management operations related to the data. It includes a caching capability enhancement path, a caching optimization strategy to improve the efficiency of high-frequency data access, a storage migration path to transfer data from high-load storage media to low-load or high-performance devices, an access rate limiting path to set access rate control rules for bursty traffic data, and a cleanup and recycling path to identify and eliminate low-frequency data to release system resources.

[0038] The system monitors the usage of various critical resources in real time, including storage capacity, bandwidth channels, CPU cores, and memory blocks. For each type of resource, the system establishes a resilience range, representing the load capacity the resource can handle under normal operating conditions. This range dynamically expands or contracts based on real-time load fluctuations. When resource load fluctuations are drastic and exceed a set threshold, the system automatically compresses the resilience range to improve anomaly detection sensitivity. When load changes stabilize, the system widens the range boundaries to increase scheduling flexibility and resource utilization, thereby achieving adaptive control of resource resilience management.

[0039] Building upon the elasticity range, the system also introduces a dynamic calibration mechanism for resource critical points. This mechanism dynamically adjusts the critical value setting for a resource by statistically analyzing the number of times a critical threshold for a specific resource category has been triggered in historical scheduling cycles, combined with the current system's operational characteristics. If a threshold is frequently triggered recently, the system will lower its upper limit to provide early warning of load risks. If it has not been triggered for an extended period, the system will raise the threshold level to reduce unnecessary resource scheduling fluctuations.

[0040] Path tension is quantified using three metrics: the increase in concurrency percentage caused by path execution, the rate of increase in storage bandwidth, and the change in CPU usage. The increase in concurrency percentage measures whether path execution significantly improves the system's task concurrency density; the rate of increase in storage bandwidth assesses the pressure on storage channels during path access; and the change in CPU usage reflects the fluctuation in computing resources during path execution. The system normalizes these three metrics and then performs a weighted combination to arrive at a path tension value. This tension value serves as a crucial basis for the scheduler to evaluate whether to execute the path. High-tension paths will be delayed or converted to alternative paths, while low-tension paths will be prioritized for execution.

[0041] After calculating path tension, the system introduces tension trade-off rules to guide the dynamic adjustment of path execution strategies. These rules primarily include three scheduling decision logics. First, prioritizing high-sensitivity data paths. These paths typically involve critical business processes, high access frequency, or low fault tolerance scenarios. The system assigns higher scheduling priority to these paths based on a preset sensitivity rating mechanism, ensuring timely processing even under resource-critical conditions. Second, reducing the execution priority of high-load paths. These paths cause additional tension amplification under the current resource conditions; therefore, the system adjusts their execution time or scheduling nodes to reduce the overall system impact. Third, replacing low-priority paths. When resources are scarce or the risk of path conflicts is high, the system will attempt to functionally replace them with backup paths that consume fewer resources or have lower execution risks, thereby achieving task objectives while reducing resource consumption.

[0042] To facilitate the execution of the aforementioned scheduling strategy, this invention establishes a lazy control pool mechanism for temporarily storing and managing all pending tasks. Considering the varying resource and scheduling requirements of different task types, the system further divides the lazy control pool into multiple sub-pools, each corresponding to a specific task category: caching tasks, migration tasks, replication tasks, and cleanup tasks. Caching tasks typically involve the pre-loading of frequently accessed data; migration tasks are used to transfer data between storage devices to alleviate load; replication tasks are used to generate data copies to ensure disaster recovery capabilities; and cleanup tasks are used to evict cold data and release resources. Each sub-pool operates independently and is dynamically managed; the system can determine the scheduling timing and priority of each task type based on its behavioral trends and resource consumption.

[0043] In determining whether a task has entered the execution state, the system primarily considers the fluctuation trend of data behavior. This trend is obtained by comparing three indicators: the access growth rate, the access decline rate, and the pause time change of the data corresponding to the task within adjacent periods. The access growth rate describes the rate at which access volume increases, the access decline rate reflects the magnitude of the decline in access volume, and the pause time change measures whether there are significant intervals or interruptions in access. These three indicators are continuously observed and weighted through a sliding time window to determine whether the current state of the data is stable, whether there are sudden behaviors, or whether it is in a cooling-off state. Based on this, the system adjusts the storage strategy of tasks in the lazy control pool, including delaying, freezing, or activating scheduling requests.

[0044] The system incorporates a fluctuation differential judgment mechanism before and after scheduling execution to prevent misallocation of system resources or misjudgment of tasks due to short-cycle anomalies. It calculates the absolute value of the difference between the fluctuation values ​​of the current and previous windows by comparing the behavioral observation results of two consecutive time windows. This fluctuation value is a composite evaluation metric constructed based on multiple behavioral indicators such as access growth rate, decline rate, and pause time changes. If the aforementioned difference value exceeds a set threshold window, the system determines that the current data behavior has deviated significantly and triggers a state change operation for the corresponding task, including task activation, scheduling execution, or unfreezing. This approach significantly reduces misjudgments caused by periodic minor fluctuations or noise, improving the timing accuracy of task execution.

[0045] To further enhance the judgment and response to system access anomalies, this invention designs a multi-dimensional access anomaly identification mechanism. This mechanism comprehensively evaluates system access quality by jointly judging the access error rate, timeout rate, and abnormal distribution of access behavior. The access error rate is obtained by statistically analyzing the proportion of erroneous requests to the total number of requests per unit time, suitable for identifying problems such as interface failures, permission verification failures, and data structure anomalies. The timeout rate is obtained by comparing the cumulative number of access requests exceeding a preset response time limit with the total number of requests, reflecting the system's processing capacity bottleneck under current load conditions. Abnormal access distribution is identified by analyzing the non-linear aggregation behavior of access requests on the time axis and the degree of load distribution offset between source nodes, revealing hidden problems such as access bursts, node skew, or network congestion. The system normalizes the three indicators and determines whether access is abnormal by setting a comprehensive threshold. If an anomaly is determined, the task status is recorded and the relevant response mechanism is triggered.

[0046] After the system detects an anomaly, this invention further introduces reverse undo logic to roll back task paths that adversely affect system resources, maximizing resource recovery and limiting the spread of the anomaly. This reverse undo logic comprises three core rules. First, undo operation identification: the system identifies reversibly executable task nodes in the current scheduling chain through execution path records and dependency analysis, excluding strongly bound tasks that cannot be safely rolled back. Second, undo impact area location: the system determines the path range affected by the current undo operation based on data relationships between tasks and resource sharing models, identifying the set of tasks requiring coordinated rollback. Third, undo scope limitation: the system sets the execution boundary for undo based on the maximum rollback depth, the maximum number of affected nodes, and the current stability state of the system, ensuring that the reverse operation is completed within a controlled range and does not trigger secondary chain conflicts.

[0047] To achieve sensitive identification and trend quantification of changes in data access behavior, a fluctuation difference enhancement model for access behavior is designed. This model integrates three key behavioral features—access growth rate, access decay rate, and access interval change—and performs nonlinear integral processing over time to obtain a fluctuation difference value for determining behavioral abrupt changes. This fluctuation difference value is calculated using a non-stationary integral enhancement function, and its mathematical expression is shown below:

[0048] ;

[0049] in,

[0050] Indicates at time The obtained access fluctuation differential enhancement amount;

[0051] The duration of the fluctuation observation window;

[0052] Indicates time The rate of increase in access recorded at that location;

[0053] Indicates time The access decay rate recorded at the location;

[0054] Indicates time The amount of change in the access interval recorded at the location;

[0055] , , These are the weighting coefficients corresponding to the three behavioral characteristics mentioned above, used to control the intensity of the influence of each characteristic on the final fluctuation difference value;

[0056] This is a time-decay enhancement kernel function, whose value varies with... Approaching the current time point Increasing the value is used to amplify the contribution of fluctuations closer to the current moment within the window;

[0057] This is the time variable during the integration process.

[0058] This integral function can continuously analyze the behavioral change trend within the observation window. Compared with the traditional difference method, this function can significantly improve the identification sensitivity in the early stage of data mutation, and also has the ability to handle non-periodic fluctuations. When the value exceeds the set threshold, the system triggers the data state transition judgment logic and enters the path scheduling preparation stage.

[0059] Furthermore, to enhance the detection capability for abnormal operational states, three types of access anomaly judgment indicators are constructed. First, the access error rate is calculated by the ratio of the number of erroneous requests occurring per unit time to the total number of requests, used to identify abnormal behaviors such as request failures, protocol conflicts, or authentication failures. Second, the access timeout rate is obtained by accumulating the number of requests exceeding a preset response time limit, used to identify abnormal states such as resource congestion, system response delays, or decreased processing capacity. Finally, the system compares the time distribution and source node distribution of all access requests within the current time period with the historical normal access baseline curve. If the deviation increases significantly, it is judged that there is an abnormal access distribution, including access spikes, node imbalances, or potential attack behaviors.

[0060] When any of the above abnormal conditions is met or multiple abnormal indicators exceed a set threshold, the system will immediately enter the reverse rollback processing phase of the scheduling path. This phase includes three core judgment rules. The first is rollback operation identification, where the system identifies rollbackable nodes in the current path based on task execution logs and dependency chain graphs, while excluding irreversible, strongly dependent, or globally transactional nodes. The second is rollback impact area location, where the system traces the associated task paths, data reference relationships, and resource allocation information of rollbackable nodes forward and backward to delineate the set of tasks affected by the rollback. The third is rollback scope limitation, where the system limits the execution of rollback operations within a safe and controllable boundary by setting a maximum rollback depth, a maximum number of affected nodes, and a resource usage recovery threshold, ensuring that system stability is not affected by a cascading effect.

[0061] function The derivation process includes:

[0062] Will access growth rate Access decay rate and the amount of change in access intervals It is considered as the fundamental variable describing changes in access behavior, and it is believed that the fluctuation of access at any given moment is not determined by a single variable, but by a composite effect formed by the combined action of the three.

[0063] To capture the trend of this composite effect over time, this invention introduces feature weights into each basic variable. , , To express the strength of different features in behavioral changes, the three features are combined in the manner of enhancement, cancellation, and perturbation. Since changes in access behavior are not triggered instantaneously but accumulate gradually over an observation window, this invention integrates the combined feature contributions over the time window. This is accumulated to form a continuous change in volatility.

[0064] On the other hand, the closer to the current moment The access behavior is more sensitive to actual fluctuations, while the historical behavior at the window edge has a weaker impact. Therefore, a time decay enhancement kernel function is multiplied outside the integral term. The kernel function is based on The contribution is adjusted based on its distance from the current moment, so that it is within the range of time. near Enhancement is gained when far away Time decay; taking the above considerations into account, this invention defines the access fluctuation differential enhancement amount as .

[0065] Combined with appendix Figure 3 The system incorporates several fundamental behavioral feature extraction and calculation mechanisms, primarily including methods for obtaining the access growth rate, extracting intermittent changes in access volume, and determining the number of erroneous requests. Regarding the acquisition of the access growth rate, the system compares the total access volume over two consecutive access cycles to obtain the growth ratio for the current cycle. The system records the access volume of the first cycle as the previous cycle value and the access volume of the second cycle as the current cycle value. By calculating the difference between the two and comparing it to the previous cycle value, the system quantifies the growth trend of access traffic. This growth ratio measures whether data requests exhibit accelerated upward behavior and serves as a crucial input to the fluctuation enhancement model.

[0066] Regarding the calculation of access interval variation, the system reconstructs the access records of a single data object within a unit of time on a timeline and calculates the time interval between any two adjacent accesses. By analyzing the magnitude of the change in this interval, the system determines whether the access has shifted from a regular access pattern to a jumpy access pattern. If there are significant fluctuations between adjacent time intervals, the access interval behavior is considered unstable and thus marked as a high-fluctuation task. This feature helps to capture nonlinear access behaviors, such as burst reads, abnormal source retries, or intermittent retries, further improving the system's ability to perceive abnormal access trends.

[0067] In the process of counting erroneous requests, the system makes a joint judgment based on three types of indicators: HTTP status code response information, communication protocol error types, and authentication failure results that occurred during the access process. Specifically, the system filters the HTTP status codes that appear in all task access responses, identifying status codes that represent abnormal situations such as request failure, access denial, and resource not found. At the same time, it parses protocol anomalies in the communication process, such as handshake failure, format mismatch, or timeout retransmission, and further combines these with failure records of the accessing user or calling program during the interface verification phase, such as incorrect authentication information or insufficient permissions, and includes them in the number of erroneous requests in the current time window.

[0068] The system constructs access behavior offset analysis mechanisms and scheduling link structure dependency analysis mechanisms to assist in anomaly detection and reverse operation security control. In judging abnormal access distribution, the system first constructs a normal access curve for the target data or interface under stable operating conditions. This curve is obtained by smoothing and fitting historical multi-period access request distribution data over time, representing the periodic characteristics and fluctuation patterns of access behavior under natural conditions. The system compares the distribution of access requests in the current period along the time axis with the aforementioned normal access curve. If the current distribution shows a significant concentrated offset, high-frequency spikes, or window delays, it is determined that its behavior over the time period has a non-linear offset, and the deviation magnitude between the access behavior and the normal curve is measured accordingly. The system uses this magnitude as a key indicator to measure the degree of abnormal access distribution offset, used to identify problems such as sudden access, bot spamming, service imbalance, or cache failure.

[0069] To determine the degree of offset of the source node of access requests, the system calculates the number of access requests received by each node per unit time based on node-level resource statistics, and calculates the ratio of this number of requests to the average access volume of all nodes to obtain the load deviation ratio of the current node. If the request load of a certain node is much higher than the average level, and the historical access pattern does not show obvious hotspot behavior, the system marks this deviation ratio as abnormal, and performs a horizontal comparison with the load of other nodes to determine whether this node will become a potential bottleneck, attack entry point, or source of uneven link distribution.

[0070] To address the potential for abnormal execution results among task nodes in the scheduling path, this invention constructs a reversible execution node identification mechanism. This mechanism quickly identifies nodes with rollback capabilities once the system's undo logic is triggered during execution, avoiding critical dependency sections and ensuring both undoing safety and link continuity. The system structurally models the dependencies between all task nodes in the scheduling link, including various types such as data dependencies, execution order dependencies, and resource binding dependencies. For each pair of nodes, the system marks its dependency direction, i.e., the forward or backward triggering order between tasks, to determine whether rollback will cause upstream or downstream logic conflicts. Furthermore, the system introduces a quantification standard for dependency strength, comprehensively considering the frequency of data calls between dependent tasks, the proportion of shared resources, and task continuity requirements to establish dependency level scores. When an anomaly is triggered, the system determines whether the current node meets the undo conditions based on the dependency type, direction, and strength. If the node is a member of a non-strong dependency chain and undoing it will not disrupt the upper-level logic structure, it is identified as a reversible node and selected as the primary target for rollback execution.

[0071] The system constructs a normal access baseline curve based on historical behavior modeling and designs a quantification algorithm for the periodic deviation magnitude and a multi-node deviation judgment mechanism. During long-term operation in the data center, the system collects access log information of the target data, including behavioral indicators such as access time, access frequency, and access source. After data cleaning and time-series normalization, the system statistically aggregates multi-period access sequences through a sliding time window, and then uses a fitting algorithm to smooth the historical access data, forming a continuous normal access curve. The system uses this normal curve as the benchmark for judging access behavior deviation in subsequent periods, providing a structured reference for the magnitude of time deviation.

[0072] During periodic deviation analysis, the system compares the peak value of the actual access traffic on the time axis within the current period with the corresponding peak value of the normal access curve, calculates the difference between the two, and introduces an amplitude weighting mechanism to enhance the sensitivity of the deviation. The weighting coefficient is determined by the current growth rate of the access curve and the time density at which the deviation occurs, resulting in a higher sensitivity response to sudden, high-frequency access deviations. If the difference value exceeds the set time deviation threshold after weighting, the system considers the current access behavior to have significantly deviated from the normal pattern and triggers the abnormal behavior identification module to enter a further judgment process.

[0073] Meanwhile, to assess whether there is structural imbalance in access anomalies, the system introduces a node-level access load assessment mechanism to identify resource scheduling imbalances within the data center. The system statistically analyzes the total number of requests received by each access node within a unit of time and compares this to the average access volume of all nodes within the same period to obtain the node's access load deviation ratio. To accommodate the differentiated behavioral characteristics of different business scenarios, the system sets deviation tolerance ranges based on the business type to which the access data belongs. This tolerance reflects the permissible fluctuation space of node load distribution for different businesses, including online queries, batch calculations, and cache prefetching. If the system detects that the load deviation ratios of multiple nodes simultaneously exceed the tolerance threshold set for their respective business types, and this fluctuation is spatially concentrated rather than globally synchronous, the system classifies this access pattern as a node-level unbalanced access anomaly and records this anomaly event for subsequent path scheduling, load migration, and risk isolation processing.

[0074] To achieve safe rollback and dependency control during the exception handling phase of scheduled tasks, the system constructs a dependency relationship identification and dependency direction resolution mechanism. This mechanism can comprehensively determine the dependency type, logical order, and binding strength between task nodes, and generate a dependency graph sequence table with reversible judgment capabilities based on the analysis results. This serves as the structural foundation for subsequent path rollback operations and task culling control. Regarding the determination of dependency types, the system extracts and categorizes the interaction relationships between any two task nodes in the scheduling chain. Based on their triggering mechanisms and resource association characteristics, dependencies are divided into three types: data dependency, execution order dependency, and resource binding dependency. Data dependency refers to the upstream task's output serving as the input condition for the downstream task, with a clear data flow relationship between them. Missing upstream data will prevent the downstream from starting normally. Execution order dependency refers to two tasks that, although not directly exchanging data, must execute according to a predetermined logical order. The scheduling phase of the next task can only begin after the previous task is completed. This type of dependency is commonly seen in transactional chains and staged scheduling structures. Resource binding dependency refers to the competition or constraint between two or more tasks for the same resource instance during execution, such as sharing a CPU core, a specific I / O port, or a dedicated memory block. Due to the exclusive or mutually exclusive operation of resources, the tasks need to coordinate the order, thus forming a soft exclusive dependency relationship.

[0075] Secondly, during the determination of dependency direction, the system logically analyzes the preceding and following trigger relationships between task nodes in the chain and establishes a directional index structure based on task start time, event response order, and triggering rules. By constructing a mapping table between task start timelines and trigger paths, the system identifies the direct predecessor and successor nodes of each task, thus forming a dependency direction graph with unidirectional causal relationships. The explicit definition of dependency direction is not only used to construct the task execution topology but, more importantly, provides path constraints for subsequent rollback node determination, preventing downstream tasks in the dependency direction from being affected during rollback and ensuring the integrity and logical consistency of the chain.

[0076] Based on the dependency type determination results and dependency direction analysis data, the system further generates a reversible execution node dependency graph sequence list. This sequence list uses tasks as nodes and dependencies as edges, and includes type labels and direction identifiers to indicate which nodes meet the rollback conditions under the current dependency structure. When access conflicts or exception handling requirements occur, the system quickly filters out reversible nodes that do not logically constitute strong dependency barriers by pruning paths in this list, forming a precise, safe, and limited-scope task rollback plan, effectively preventing chain-like exception propagation.

[0077] Example 1:

[0078] Combined with appendix Figure 4In this embodiment, a week before a promotional event, the access frequency of the OrderLog table in the core data center of a large e-commerce platform surged from approximately 5,000,000 times per day to approximately 36,000,000 times per day. Simultaneously, the triggering entity for these access requests shifted from a background scheduler to primarily front-end user behavior. To improve system response efficiency and prevent resource bottlenecks, the data center implemented the AI-based data management method described in this invention.

[0079] The system monitors the access behavior of the OrderLog data object in 6-hour intervals, first statistically analyzing the access time intervals. Under normal conditions, the average access time interval for this object is approximately 10 seconds, with a standard deviation of approximately 1 second. However, in the two days leading up to the promotional event, the standard deviation of this time interval increased to approximately 5 seconds, indicating a significant fluctuation in the access rhythm. Subsequently, the system measures request burstiness, calculating the ratio of the maximum access request volume to the average request volume within a 1-minute time window, which increased from approximately 1.2 times to approximately 3.5 times. Furthermore, the system detects the frequency of trigger subject switching, finding that the number of trigger subject types increased from the original 2 to 7, and the switching frequency increased to approximately 4 times the historical average. The system weights the above three indicators with weights of 0.4, 0.4, and 0.2, calculating the data state variation rate for the current period to be 0.78, significantly higher than the historical average variation rate of 0.25 for this data object.

[0080] The system uses a dynamic adjustment mechanism for behavior transformation thresholds to retrospectively analyze the stability of access patterns over the past seven consecutive days in the same period. It found that the historical maximum mutation rate did not exceed 0.42, therefore the behavior transformation threshold for this period was set to 0.50. Since the current mutation rate (0.78) has exceeded this threshold, the system automatically determines that the OrderLog data object has entered a behavior transformation state and generates a corresponding target processing path sequence. Priority is given to cache boosting paths and access rate limiting paths to handle high-frequency access and sudden loads.

[0081] The system then assesses resource pressure and sets a resource elasticity range of ±15% based on the historical average load. In the previous period, the system's average CPU utilization was 57%, and storage bandwidth usage was 35%. Current monitoring data shows that after implementing the cache enhancement path, CPU utilization rose to 65%, bandwidth usage increased to 50%, and the number of concurrent tasks increased from 1,500 to 2,400. Based on this, the system calculates path tension metrics, concluding that the concurrency percentage increased by 60%, the bandwidth increase rate was 42%, and the CPU usage changed by 8%.

[0082] Based on the system's preset indicator weights—concurrency increase of 0.5, bandwidth increase of 0.3, and CPU usage change of 0.2—the final path tension level is calculated as follows:

[0083] ;

[0084] Since the system's current resource critical point has been dynamically adjusted from the original 65% to 72% based on the frequency of triggers over the past 30 consecutive days, and the currently calculated tension value has not yet reached this critical point, the system allows the immediate execution of the cache boosting path and marks the data object's status as "Cache acceleration activated." Simultaneously, the system will continue to monitor the fluctuations of this object for the next cycle to support subsequent scheduling optimizations.

[0085] After identifying that the data state variation rate exceeds the behavior transformation threshold and initially generating cache enhancement paths and access rate limiting paths, the system further activates tension trade-off rules to ensure the system-level coordination of overall scheduling and task priority control. The system assesses the sensitivity of OrderLog data, which belongs to the core log of real-time order processing and is highly sensitive to business continuity and user experience. Based on this, the system determines that its path type is a high-sensitivity path and triggers a priority protection strategy. That is, when the path tension value does not exceed the resource critical point and there is contention for tasks, the execution order and scheduling ratio of this task should be reserved first.

[0086] Meanwhile, the system detected another data object, PromoTrack, which belongs to promotional activity records. The current path involves both cache refresh and log replication operations. The path tension value of PromoTrack is 0.476, slightly higher than OrderLog's 0.442, but its data sensitivity is marked as medium priority, and the current task is a non-critical analysis path. Based on this, the system determines that PromoTrack should be downgraded. Through tension trade-off logic, the non-critical replication path is replaced with a low-bandwidth summary write path, and the migration task is marked as deferred.

[0087] After determining the execution path, the system loads different types of data tasks into the lazy control pool and assigns them to corresponding sub-pools based on task type. Specifically, the cache promotion path and access rate limiting path of OrderLog are assigned to the cache and cleanup sub-pools, while the log summary path of PromoTrack is assigned to the migration sub-pool. The lazy control pool periodically monitors the data behavior fluctuation trends of each task. The system performs fluctuation evaluations on the cache-type task containing OrderLog for three consecutive periods, recording access growth rates of 48%, 39%, and 42%, and decrease rates of 6%, 2%, and 4%, with pause time changes of less than 5 seconds, indicating an overall stable trend. Therefore, the system marks it as "delayed executable" and sets the next execution window for evaluation after three periods.

[0088] The fluctuation value of the PromoTrack data object changed drastically over two periods, with a growth rate as high as 70% and a significant fluctuation in the decline rate. The system determined that its fluctuation trend had not converged and maintained the task in a frozen state. Entering the next scheduling period, the system monitored the fluctuation difference. The PromoTrack fluctuation value in the previous period was 0.73, and in this period it was 0.94. The system compared the absolute value of the difference (0.21) with the threshold setting value (0.15) and determined that the difference had exceeded the threshold window. This triggered the lazy control pool activation logic, marked the PromoTrack migration path as pre-execution, and simultaneously restarted the task context.

[0089] After executing the above path, the system sets a conflict detection window for the execution status of all tasks, with a width of 10 minutes. The system monitoring found that the access error rate of PromoTrack data objects increased from the original 0.3% to 1.1% after the copy path was executed, the timeout rate exceeded 2.5%, and asymmetric access load appeared, with the load deviation of a specific source node exceeding 1.6 times. Based on this, the system judged that there was an abnormal distribution of access and triggered the reverse undo logic.

[0090] Based on the task scheduling link structure diagram, the system identifies the undo operation of the task path where PromoTrack is located, confirming that it is a reversible node and does not involve strong dependency constraints. Subsequently, the system locates the area affected by the undo and finds that the cache preheating node it depends on is a parallel structure with no strong dependency path propagating downwards. The system determines that the rollback boundary is controllable, implements task undo, and records the fault label for feedback to strengthen the path adjudication model.

[0091] In the continuous operation of e-commerce platform promotional scenarios, the system needs to identify drastic fluctuations in access behavior in a more refined and quantitative manner. To this end, this invention uses a non-stationary integral enhancement function to model the access fluctuation difference and introduces three dimensions—access growth rate, access decay rate, and access interval change—for joint weighted calculation, thereby improving the ability to perceive potential abrupt changes in access patterns.

[0092] Taking the behavior of the OrderLog data object in the first two periods of the activity as an example, the system sets the fluctuation observation window length. Hours, corresponding to the timeline The calculation is performed by sampling in hourly increments. The following parameters are obtained from the sampling:

[0093] exist hourly access growth rate Access decay rate Access intermittent changes ;

[0094] exist When I was a child, , , ;

[0095] exist When I was a child, , , ;

[0096] exist When I was a child, , , ;

[0097] exist When I was a child, , , ;

[0098] exist When I was a child, , , ;

[0099] Set the weight coefficients for each item as follows: , , And select a time decay enhancement kernel function of the following form:

[0100] ;

[0101] Substituting the above parameters into the fluctuation difference enhancement function:

[0102] ;

[0103] Using numerical integration as an approximation (in hours), we obtain:

[0104] 1st hour item ( The weighting factor is ;

[0105] 2nd hour item ( The weighting factor is ;

[0106] The 3rd hour item is ;

[0107] The 4th hour item is ;

[0108] The 5th hour item is ;

[0109] Current hour item is ;

[0110] Calculate the integrand value for each term:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] ;

[0117] Multiply each term by its corresponding kernel function weight and sum them up to obtain:

[0118] ;

[0119] ;

[0120] The final fluctuation difference enhancement at the current moment is:

[0121] ;

[0122] The system compares this value with the fluctuation difference threshold of 0.70 and confirms that the current fluctuation has entered the overheated zone, meeting the conditions for scheduling optimization or risk warning.

[0123] In addition, to improve the sensitivity of anomaly detection, the system further introduces access error rate, timeout rate, and abnormal access distribution indicators. In the current period, the total number of requests is 1,200,000, the number of erroneous requests is 10,800, and the access error rate is calculated as follows:

[0124] ;

[0125] If the system-defined anomaly threshold is 0.6%, then this value has exceeded the limit. Simultaneously, 32,000 requests exceeding the 5-second response time limit were detected, resulting in a timeout rate of:

[0126] ;

[0127] The system determined that there was a structural shift in access distribution when a large number of access requests were concentrated on a single source node during peak hours, deviating from the load balancing strategy, and ultimately triggered the revocation module.

[0128] The system identifies the current task as not belonging to a strong dependency chain based on the scheduling link dependency graph. It only depends on the data archiving node and has no propagation effect to lower-level nodes. Finally, it locates the reversible node as the log digest task, performs the undo operation, and resets the context.

[0129] Example 2:

[0130] Combined with appendix Figure 5 Based on Example 1, in two consecutive access cycles, the system recorded 530,000 and 612,000 accesses in OrderLog, respectively, with a cycle length of one hour. The system calculates the access growth rate using the increment ratio formula:

[0131] ;

[0132] The growth rate is 15.47%, indicating that the current access pressure is actively increasing.

[0133] The system then extracted the timestamps of the five most recent accesses: 13:00:05, 13:00:19, 13:00:22, 13:00:58, and 13:01:07, and calculated the intervals between each adjacent access to be 14 seconds, 3 seconds, 36 seconds, and 9 seconds, respectively. The amplitude of the interval variation was calculated using variance as follows:

[0134] ;

[0135] in ,have to:

[0136] ;

[0137] This indicates that the current access interval variation is 155.25 seconds², which is significantly abnormal, prompting the system to enter the fluctuation-sensitive processing logic.

[0138] In the access anomaly identification process, the system analyzed the total number of erroneous requests through HTTP response status analysis. This included responses with status codes 4xx and 5xx, totaling 8,900. Additionally, 670 protocol format errors and 1,050 verification failure events were detected. The system's overall assessment of the erroneous request count was:

[0139] ;

[0140] Compared to a total of 650,000 requests, the access error rate was:

[0141] ;

[0142] It has far exceeded the system's warning threshold of 1%.

[0143] In abnormal access distribution detection, the system uses multi-period smoothing fitting based on historical normal access curves to obtain an average peak access frequency of 620,000 times, with a peak frequency of 755,000 times recorded in the current period, a difference of 135,000, and an offset magnitude of:

[0144] ;

[0145] The offset is significant. Comparing the access distribution from the source nodes, node A processed 312,000 accesses this period, with an average node processing volume of 180,000. The load skew ratio is:

[0146] ;

[0147] The number of visits exceeded the set tolerance limit by 1.5 times, further confirming that an anomalous access source aggregation effect has formed.

[0148] To ensure system stability and support dynamic recovery, the system initiates task dependency chain graph parsing logic to determine the type, direction, and strength of dependencies between nodes in the execution path. Taking the OrderLog write path as an example, the current task sequence is data preprocessing → log classification and writing → index update → cross-region backup. The system analyzes the dependencies as follows:

[0149] Data preprocessing and log classification are data dependencies, with the direction being a unidirectional weak dependency.

[0150] There is an execution order dependency between index updates and categorized writes, and the direction is a strong dependency;

[0151] Cross-region backup and index update are resource-bound dependencies;

[0152] The system identifies the log category write node as a reversible execution node, and its dependency direction does not form a strong binding with subsequent nodes. Therefore, when a write anomaly is detected, the system can skip the index update node and directly perform a rollback operation on the category write, ensuring that no cascading failures are triggered during the anomaly recovery process.

[0153] To achieve accurate detection of access anomalies and intelligent identification of dependency scheduling, the system introduces a normal access curve modeling and a node-level anomaly judgment mechanism. The system extracts access data from the OrderLog over the most recent 10 consecutive promotional periods, dividing each period by hour, collecting a total of 240 access data points. Using a weighted moving average method for smoothing, the system constructs a normal access curve for the OrderLog. The results show that peak access values ​​are relatively stable between 620,000 and 635,000 times, with an average peak of 627,000 times. Subsequently, the system compares the actual access data of the OrderLog within the current period and finds that the access volume surges sharply to 752,000 times between 14:00 and 15:00. The difference between this abnormal access peak and the peak of the normal access curve is 125,000 times. After magnitude weighting:

[0154] ;

[0155] Among them, the weighting coefficient The system's preset promotional anomaly sensitivity factor indicates that the current access deviation is severe, exceeding the system's maximum tolerance range of 80,000 times, thus triggering the abnormal access deviation identification mechanism.

[0156] To further pinpoint the source of the offset, the system analyzed the access load of each front-end access node during the abnormal period. The results are as follows: Node A processed 312,000 requests, Node B processed 180,000 requests, Node C processed 192,000 requests, and the system average node access volume was:

[0157] ;

[0158] The load deviation ratio of node A is:

[0159] ;

[0160] The deviation ratio for node B is 0.789, and for node C it is 0.842. According to the business configuration file, the system sets a maximum deviation tolerance of ±25% for "order log writing" tasks, i.e., a tolerance range of [0.75, 1.25]. Node A clearly exceeds this limit, with a deviation exceeding 13%, and other nodes are also approaching the critical value. Because node A's deviation value has exceeded the threshold, the system, based on the logic of multiple nodes simultaneously reaching critical offsets, classifies the current access as a node-level unbalanced access anomaly.

[0161] After triggering the exception check, the system initiates the scheduling chain dependency resolution module to identify the structure of the current OrderLog processing task chain. This chain contains five nodes: data writing, content verification, index building, log packaging and archiving, and cross-region synchronization. The system identifies the dependency types based on the task relationship graph as follows:

[0162] Data writing and content validation are data-dependent;

[0163] Validation and index building are dependent on the execution order.

[0164] Log archiving and cross-region synchronization are resource-bound dependencies;

[0165] Further, the dependency direction is confirmed through the task trigger sequence. Data writing is the starting node, and archiving and synchronization are the ending nodes. The system generates a dependency graph sequence table as: [N1→N2→N3→N4→N5], where N1 is a reversible node, N2 and N3 are unidirectional strong dependency chains and are irreversible, and N4 is a resource-bound weak dependency node with conditional reversibility. Based on this dependency graph, the system identifies N1 and N4 as candidate rollback points. Considering that the load deviation is mainly concentrated at the initial log writing node, the system ultimately chooses to reverse the rollback of N1 and freeze the subsequent task scheduling of N4.

[0166] Based on the normal access behavior benchmark constructed by smooth fitting, structural anomalies caused by high-amplitude deviation behavior are identified. The uneven access is judged in detail at the node level. Combined with the dependency graph sequence logic, reversible task nodes are located, which effectively improves the identification accuracy and scheduling flexibility of the data center in the event of anomalies.

[0167] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A data center data management method based on artificial intelligence, characterized by The method comprises: Recording the access mode, trigger subject and time distribution of the target data object to be managed in the data center within a preset period, wherein the access mode comprises reading, writing, deleting or migrating, and the trigger subject comprises a terminal user, an upper application, an automatic scheduler or a third-party interface request; and judging the data state variation rate based on the behavior track, interaction feature and load change, wherein the data state variation rate is a quantitative parameter for measuring the degree of change of the current data usage state, the calculation of which comprises weighted fusion of the access time interval variance, request burst degree and trigger subject switching frequency, and the access time interval variance, request burst degree and trigger subject switching frequency are respectively given adjustable weight parameters and the weighted comprehensive value is calculated; the data target processing path sequence generated when the variation rate reaches the behavior transition threshold; the behavior transition threshold is periodically adjusted according to the access mode stability degree of the historical data within the same period; the data target processing path sequence is a subsequent management operation suggestion for the target data object, and at least one of the cache promotion path, storage migration path, access flow limiting path or cleaning and recycling path is included; According to the path sequence, the tension degree of each path to the resource is evaluated in combination with the elastic range, resource critical point and real-time occupation of the current storage and computing resource of the data center, wherein the elastic range of the resource is expanded or contracted in real time according to the system load fluctuation rate; the resource critical point is dynamically re-scaled in combination with the frequency of the threshold history being triggered; the tension degree is obtained by calculating the concurrent occupation growth, storage bandwidth rising rate and CPU occupation change caused by path execution, and the path execution, delay or replacement is judged by the tension weighing rule to form an actual processing plan; the processing plan is temporarily stored in an inert control pool, the delay execution is set for the stable task by monitoring the data behavior fluctuation trend, the freezing is set for the continuous high fluctuation task, and the actual execution of the processing plan is triggered only when the fluctuation difference exceeds the threshold window; A conflict detection window is set for the executed action, and a reverse rollback logic is triggered when concurrent conflict, access anomaly or resource overload is detected, the most recent non-mandatory task is rolled back, and the state evaluation value of the path task associated with the data is updated. 2.The artificial intelligence-based data center data management method of claim 1, wherein The tension weighing rule comprises the judgment logic of prioritizing high sensitivity data paths, reducing high load paths and replacing low priority paths; the inert control pool comprises a plurality of sub-pools for processing cache, migration, replication or cleaning tasks according to different types of tasks; the data behavior fluctuation trend is obtained by comparing the access growth rate, access decline rate and pause time change in adjacent periods. 3.The artificial intelligence-based data center data management method of claim 1, wherein The condition that the fluctuation difference exceeds the threshold window is judged based on the difference absolute value of the fluctuation values of the current window and the previous window; the access anomaly is realized by identifying the access error rate, timeout rate and abnormal access distribution; the reverse rollback logic comprises three rules of rollback operation identification, rollback impact area positioning and rollback range limitation. 4.The artificial intelligence-based data center data management method of claim 3, wherein The calculation of the fluctuation difference comprises jointly weighting the access growth rate, the access decay rate and the access intermittent change amount; the access error rate is obtained by a ratio of the number of error requests to the total number of requests in a unit time; the timeout rate is obtained by cumulatively calculating the access requests exceeding a preset upper time limit; the abnormal access distribution is determined based on the offset degree of the access requests on the time axis and the source nodes; and the revocation operation recognition comprises recognizing nodes with reversible execution in a scheduling link and skipping nodes with strong dependency relationship. 5.The artificial intelligence-based data center data management method of claim 4, wherein The access growth rate is obtained based on a growth ratio calculated from a difference in access amount in two consecutive access periods; the access intermittent change amount is obtained by calculating a change amplitude from a time interval between two adjacent accesses; and the number of error requests is determined based on HTTP status codes, protocol error types and verification failure results. 6.The artificial intelligence-based data center data management method of claim 4, wherein The offset degree of the abnormal access distribution is determined by comparing a periodic deviation of the access requests from a normal access curve on the time axis; the judgment of the offset degree of the source nodes comprises comparing load deviation ratios of the access requests in different nodes; and the recognition of the nodes with reversible execution in the scheduling link comprises determining types, directions and strengths of dependency relationships between the nodes. 7.The artificial intelligence-based data center data management method of claim 6, wherein The normal access curve is obtained by smoothing and fitting historical multi-period access data, and is used to provide a time offset benchmark; the amplitude judgment of the periodic deviation comprises amplitude weighting a difference between an abnormal access peak value and a normal curve peak value; the load deviation ratio of the nodes is obtained by comparing a ratio of node access amount to average node access amount, and different deviation tolerances are set according to business types, and when load deviation ratios of multiple nodes simultaneously exceed the tolerance range, the access mode is identified as a node-level uneven access abnormality. 8.The artificial intelligence-based data center data management method of claim 6, wherein The determination of the types of the dependency relationships comprises recognizing three types of relationships, i.e. data dependency, execution sequence dependency and resource binding dependency; the determination of the dependency directions is determined according to a front and back trigger sequence between tasks, and is used to generate a dependency graph sequence table of the reversible nodes.

Images