A method for ensuring reliability of data storage

Abstract

The application discloses a kind of reliability guarantee methods of data storage, it is related to distributed data storage technical field.The method is to determine first near failure node and candidate target node set, and collect migration environment state data, then the data input contains multiple decision-making intelligent agent reinforcement learning model, and generates migration path decision by group decision verification, finally executes data migration and continuously collects environment state and target node health degree: when expected completion time is over threshold value, reactivation model is carried out online reprogramming, and when target node health degree is lower than threshold value, trigger task splitting fault-tolerant mechanism, confirm migrated fragment, reassign alternative target node to un-migrated fragment and update migration transaction state synchronously, whereby through multiple intelligent agent collaborative decision-making, online dynamic adjustment and fine fault-tolerant, adaptive optimization of migration path and active protection of data reliability are realized, which is convenient for practical application and promotion.

Description

Technical Field

[0001] This invention belongs to the field of distributed data storage technology, and specifically relates to a method for ensuring the reliability of data storage. Background Technology

[0002] Distributed data storage systems improve data reliability and availability by distributing data across multiple storage nodes and utilizing data redundancy techniques (such as replication mechanisms and erasure coding). In distributed data storage systems, ensuring that data stored on a storage node is not lost when the node experiences performance degradation, hardware aging, or impending failure is one of the core issues in ensuring data reliability.

[0003] Currently, existing technologies typically address node failure risks through data migration: when a storage node is detected to be at risk of failure, the data on that node is migrated to other healthy nodes, thereby preventing data loss due to node failure. However, existing technologies still have the following technical problems during the data migration process: (1) There is a contradiction between the static nature of the migration path and the dynamic nature of environmental changes. That is, although the existing technology can generate an initial migration path, once the path is determined, it will start to be executed, lacking the ability to continuously respond to changes in the environmental state during the migration process. In practical applications, environmental factors such as network link quality, target node load and source node health status may change at any time. Static migration paths cannot adapt to these dynamic changes, resulting in reduced migration efficiency or even migration failure. (2) There is a contradiction between the passivity of the fault tolerance mechanism and the initiative of reliability assurance. That is, the existing technology uses the breakpoint resume mechanism to deal with abnormal situations during the migration process. This mechanism is a passive fault tolerance - that is, recovery is only carried out after the migration fails. This method cannot actively avoid risks when the health of the target node declines but has not completely failed, nor can it dynamically adjust the migration target during the migration process, resulting in insufficient timeliness of data reliability assurance. (3) There is a contradiction between the singular goal of migration decision and the multidimensional requirements of reliability assurance. That is, the existing technology mainly focuses on the single goal of migration efficiency, and lacks comprehensive consideration of multidimensional factors such as business importance, data lineage, storage cost and access latency. It is difficult to ensure data reliability while taking into account business continuity and storage resource optimization.

[0004] There is currently no effective solution to the aforementioned technical problems. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, computer device, computer-readable storage medium, and computer program product for ensuring the reliability of data storage, in order to solve the contradictions between the static nature of migration paths and the dynamic nature of environmental changes, the passivity of fault tolerance mechanisms and the initiative of reliability assurance, and / or the singularity of migration decision objectives and the multidimensional requirements of reliability assurance in existing distributed data storage technologies.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a method for ensuring the reliability of data storage is provided, including: Determine the critical node and the set of candidate target nodes for receiving and storing the data blocks to be migrated from the distributed storage cluster, wherein the data blocks to be migrated refer to the data blocks currently stored on the critical node; Real-time acquisition of migration environment status data, wherein the migration environment status data includes at least one of the following: load of critical nodes, load of candidate target nodes, network link quality, and global load distribution of the cluster; The migration environment state data is input into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and the outputs of each decision agent are fused through a group decision verification mechanism to generate a migration path decision result. The migration path decision result includes the selected target node, migration shard size, migration concurrency thread count and / or migration compression algorithm. Data migration is performed on the data block to be migrated according to the migration path decision results, and environmental status changes and the health score of the selected target node are continuously collected during the migration process; When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online, wherein the online replanning includes at least reselecting the target node. When the health score of the selected target node is detected to have decreased and fallen below the preset score threshold, the task splitting fault tolerance mechanism is triggered: for the data shards in the data block to be migrated that have completed migration, confirmation is submitted; for the data shards in the data block to be migrated that have not completed migration, the corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

[0007] Based on the above-mentioned invention, a novel data storage reliability assurance scheme is provided that integrates multi-agent reinforcement learning and group decision-making to dynamically plan migration paths and achieve online replanning and task splitting fault tolerance collaboration. First, a set of potentially failing nodes and candidate target nodes is determined, and migration environment status data is collected. Then, this data is input into a reinforcement learning model containing multiple decision-making agents. A migration path decision is generated through group decision-making verification. Finally, data migration is executed while continuously collecting environmental status and target node health data. When the expected completion time exceeds a threshold, the model is reactivated for online replanning. When the target node health falls below a threshold, a task splitting fault tolerance mechanism is triggered, confirming migrated shards and reallocating replacement target nodes for unmigrated shards while simultaneously updating the migration transaction status. Thus, through multi-agent collaborative decision-making, online dynamic adjustment, and refined fault tolerance, adaptive optimization of the migration path and proactive assurance of data reliability are achieved, facilitating practical application and promotion.

[0008] In one possible design, the critical nodes and the set of candidate target nodes for receiving and storing the data blocks to be migrated are determined from the distributed storage cluster, including: Obtain multidimensional health status data of each storage node in a distributed storage cluster, wherein the multidimensional health status data includes at least one of I / O latency time series, disk error rate, bandwidth jitter amplitude, data verification failure rate, disk SMART information and SSD wear level; For each storage node, the corresponding multidimensional health status data is input into the pre-trained health score model, and the health score model uses a long short-term memory network or a gated recurrent unit to extract the temporal features of the multidimensional health status data and outputs the corresponding health score. Storage nodes whose health scores are lower than a first score threshold but higher than a second score threshold are identified as nodes on the verge of failure, wherein the second score threshold is lower than the first score threshold; Storage nodes with health scores higher than the first score threshold are identified as candidate target nodes for receiving and storing data blocks to be migrated, forming a set of candidate target nodes. The data blocks to be migrated refer to the data blocks currently stored on the faulty node.

[0009] In one possible design, the group decision verification mechanism includes: Obtain candidate migration paths output by the at least two decision agents, and construct a verifiable decision topology network based on the obtained results; Perform domain-aware group advantage tuning on each candidate migration path in the verifiable decision topology network to obtain the tuning confidence of each candidate migration path; Based on the optimization credibility, the verifiable decision topology network is subjected to group decision credibility convergence verification to obtain the comprehensive credibility evaluation value of each candidate migration path. The candidate migration path with the highest overall credibility score will be used as the final migration path decision.

[0010] In one possible design, the method further includes: Obtain the business metadata of the data object to which the data block to be migrated belongs, wherein the business metadata includes the data importance level, expected access frequency, related business service level agreement requirements, and data lineage; Based on the historical access logs of the data object, a data access pattern profile of the data object is constructed. The data access pattern profile is used to reflect the transformation patterns and transformation time windows of cold data, warm data and hot data. Combining the business metadata and the data access pattern profile, a differentiated migration strategy is generated for the data object using a multi-objective optimization algorithm. The objective function of the multi-objective optimization algorithm includes storage cost, access latency, data reliability, and energy consumption. The differentiated migration strategy includes migration priority, migration time window, and atomic migration requirements. The differentiated migration strategy is input as a constraint into the multi-agent reinforcement learning migration decision model to guide the migration path decision of the data block to be migrated. When performing the data migration task of the data block to be migrated, the migration operation is scheduled to be performed during the predicted low business hours according to the migration time window in the differentiated migration strategy. For the data objects that have a data lineage, in accordance with the atomic migration requirements in the differentiated migration strategy, the migration operation is performed atomically on all data blocks contained in the data objects to ensure the lineage integrity of the data objects.

[0011] In one possible design, the method further includes: Construct a multi-level fault domain topology diagram of the distributed storage cluster, wherein the multi-level fault domain topology diagram includes the relationship between physical machines, racks, switches and data centers and historical fault statistics of each level; When determining the set of candidate target nodes, the common failure probability of the node about to fail and each other storage node is calculated based on the fault history statistics, and other storage nodes with a common failure probability lower than a preset probability threshold are selected as candidate target nodes. When reselecting target nodes, choose the node combination that minimizes the sum of common failure probabilities within the already selected target node group.

[0012] In one possible design, the task splitting fault tolerance mechanism also includes: When the health score of the critical node is detected to have dropped below the third score threshold, the migration priority of the data block to be migrated is automatically increased, and the migration concurrency is dynamically adjusted. And / or, when network link quality fluctuations are detected to exceed a preset fluctuation threshold, the migration fragment size is adaptively adjusted: the fragment size is reduced on low-quality network links and the migration fragment size is increased on high-quality network links to improve the transmission success rate; And / or, record fault-tolerant events and execution results, and use the recorded results as negative samples to incrementally train the multi-agent reinforcement learning transfer decision model.

[0013] In one possible design, the migration transaction status of the data block to be migrated includes a transaction identifier, the node identifier of the faulty node, the selected target node identifier, the data block identifier, the migration status, the retry count, and the migration transaction creation timestamp. The migration status includes pending migration, in migration, confirmed, rolled back, and pending retry. Synchronously update the migration transaction status of the data blocks to be migrated, including: When a recoverable abnormal event is detected during the data migration process of the data block to be migrated, the migration status is switched to pending retry and the retry count is incremented. The recoverable abnormal event refers to network timeout, temporary unreachability of the target node, or resource contention. When the retry count exceeds the preset count threshold, an anomaly type analysis is triggered: if the anomaly type is that the target node is unreachable, then the target node switching action is performed and the retry count is reset; if the anomaly type is that the network timeout, then the data migration task is re-executed after reducing the transmission rate.

[0014] In one possible design, the training process of the multi-agent reinforcement learning transfer decision model includes: For each of the at least two decision agents, a corresponding policy network and a value network are constructed. The value network is used to evaluate the quality of the migration path decision results output by the policy network. The optimization objective function of the policy network and the value network consists of at least two optimization objectives, including minimizing the migration completion time, maximizing the target node load balancing, and / or minimizing the impact of the migration process on business I / O. A training architecture combining centralized training and distributed execution is adopted: During the training phase, the experience data of each decision agent is collected centrally and shared among them, so as to use the experience data to synchronously update the policy network parameters and value network parameters of each decision agent through a multi-objective policy gradient algorithm; During the execution phase, each decision agent independently makes migration path decisions based solely on its own policy network.

[0015] Secondly, a data storage reliability assurance system is provided, including a node type determination unit, an environment status acquisition unit, a migration path decision unit, a path decision execution unit, an online replanning unit, and a task splitting trigger unit. The node type determination unit is used to determine the faulty node and the set of candidate target nodes for receiving and storing the data block to be migrated from the distributed storage cluster, wherein the data block to be migrated refers to the data block currently stored on the faulty node; The environment status acquisition unit is communicatively connected to the node type determination unit and is used to acquire migration environment status data in real time. The migration environment status data includes at least one of the following: load of the node on the verge of failure, load of the candidate target node, network link quality, and global load distribution of the cluster. The migration path decision unit is communicatively connected to the environment state acquisition unit. It is used to input the migration environment state data into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and to fuse the outputs of each decision agent through a group decision verification mechanism to generate a migration path decision result. The migration path decision result includes the selected target node, migration shard size, migration concurrent thread count and / or migration compression algorithm. The path decision execution unit is communicatively connected to the migration path decision unit and is used to perform data migration on the data block to be migrated according to the migration path decision result, and continuously collect environmental status changes and the health score of the selected target node during the migration process. The online replanning unit is communicatively connected to the path decision execution unit. When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online. The online replanning includes at least reselecting the target node. The task splitting triggering unit is communicatively connected to the path decision execution unit. When the health score of the selected target node is detected to have decreased and fallen below a preset score threshold, the task splitting fault tolerance mechanism is triggered: for data shards in the data block to be migrated that have completed migration, confirmation is submitted; for data shards in the data block to be migrated that have not completed migration, a corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

[0016] Thirdly, the present invention provides a computer device, comprising a storage module, a processing module, and a transceiver module connected in sequence for communication, wherein the storage module is used to store a computer program, the transceiver module is used to send and receive messages, and the processing module is used to read the computer program and execute the reliability assurance method as described in the first aspect or any possible design in the first aspect.

[0017] Fourthly, the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, perform the reliability assurance method as described in the first aspect or any possible design within the first aspect.

[0018] Fifthly, the present invention provides a computer program product, including a computer program or instructions, wherein the computer program or instructions, when executed by a computer, implement the reliability assurance method as described in the first aspect or any possible design in the first aspect.

[0019] The beneficial effects of the above scheme are: (1) This invention creatively provides a new data storage reliability guarantee scheme that integrates multi-agent reinforcement learning and group decision-making to dynamically plan migration paths and realize online replanning and task splitting fault tolerance collaboration. That is, firstly, the set of nodes in danger and candidate target nodes is determined, and migration environment status data is collected. Then, the data is input into a reinforcement learning model containing multi-decision agents. The migration path decision is generated through group decision verification. Finally, the data migration is performed and the environment status and target node health are continuously collected. When the expected completion time exceeds the threshold, the model is reactivated for online replanning. When the target node health is lower than the threshold, the task splitting fault tolerance mechanism is triggered. The migrated shards are confirmed, and the unmigrated shards are reassigned to replace the target nodes and the migration transaction status is updated synchronously. Thus, through multi-agent collaborative decision-making, online dynamic adjustment and fine-grained fault tolerance, adaptive optimization of migration path and active guarantee of data reliability are realized. (2) By constructing a health rating model based on long short-term memory network or gated recurrent unit, the time-series features of the multidimensional health status data of storage nodes can be extracted, the health status of nodes can be quantified, and a dual threshold strategy can be used to accurately identify nodes on the verge of failure as migration sources and select healthy nodes as migration targets, providing a reliable starting point for migration tasks. (3) By introducing a reinforcement learning transfer decision model containing multiple decision agents, each agent is guided by different objectives such as transfer completion time, load balancing degree, and business impact. A verifiable decision topology network is constructed through a group decision verification mechanism. After domain-aware optimization and credibility convergence verification, the optimal transfer path is generated by fusion, thus achieving dual protection of multi-objective collaborative optimization and decision credibility. (4) By continuously collecting the environmental status and target node health during the migration process, when the expected completion time exceeds the threshold, the path is replanned online and the target node is dynamically switched. When the target node health deteriorates, the task splitting fault tolerance is triggered, the migrated shards are confirmed and the unmigrated shards are reallocated, and the migration transactions containing five states, namely, pending migration, in migration, confirmed, rolled back and pending retry, are updated synchronously to ensure data consistency and recoverability under abnormal conditions. (5) By increasing migration priority and concurrency when the source deteriorates, adaptively adjusting the fragment size when the network fluctuates, and using fault-tolerant events as negative samples to incrementally train the decision model, a closed loop of multi-dimensional active fault tolerance and self-evolution is formed. (6) By adopting a centralized training and distributed execution architecture, each agent shares experience during the training phase, updates the policy network and value network parameters synchronously through multi-objective policy gradients, and makes independent decisions during the execution phase, thus ensuring the model's efficient training and real-time response capabilities. (7) By acquiring business metadata, a data access pattern profile is constructed, and a differentiated migration strategy containing priority, time window and atomicity requirements is generated. This strategy is then integrated into the multi-agent reinforcement learning decision model as a constraint, realizing business-aware migration scheduling and lineage integrity assurance. This ensures that migration decisions can guarantee data reliability while taking into account business continuity and storage resource optimization. (8) By constructing a multi-level fault domain topology map and introducing historical fault statistics, the dynamic common fault probability between nodes is calculated. In the selection of candidate target nodes, low probability nodes are selected first, and the sum of probabilities within the combination is optimized in the selection of target node groups, thus realizing the proactive guarantee of data reliability from the physical topology level. (9) This solution uses multiple technical means to work together to enable data migration to dynamically adapt to environmental changes, proactively avoid risks and handle anomalies in a refined manner, which significantly improves the data reliability of the distributed storage system and facilitates practical application and promotion. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating the data storage reliability assurance method provided in this application embodiment.

[0022] Figure 2 This is a schematic diagram of the data storage reliability assurance system provided in the embodiments of this application.

[0023] Figure 3 A schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in conjunction with the accompanying drawings and descriptions of the embodiments or the prior art. Obviously, the following description of the structure of the accompanying drawings is only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these embodiments without creative effort. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0025] It should be understood that although the terms "first" and "second", etc., may be used herein to describe various objects, these objects should not be limited by these terms. These terms are only used to distinguish one object from another. For example, the first object may be referred to as the second object, and similarly, the second object may be referred to as the first object, without departing from the scope of the exemplary embodiments of the invention.

[0026] It should be understood that the term "and / or" that may appear in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, or A and B exist simultaneously. Another example is A, B and / or C, which can mean that any one of A, B, and C or any combination thereof exists. The term " / and" that may appear in this document describes another relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone or A and B exist simultaneously. In addition, the character " / " that may appear in this document generally indicates that the related objects before and after it are in an "or" relationship.

[0027] Example like Figure 1 As shown, the data storage reliability assurance method provided in the first aspect of this embodiment can be executed, but is not limited to, by a computer device with certain computing resources. For example, it can be executed by a master node or control node in a distributed storage cluster, or by a independently deployed management server. The aforementioned device needs to include at least one processor and at least one memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the method described in this embodiment. Figure 1 As shown, the reliability assurance method includes, but is not limited to, the following steps S1 to S6.

[0028] S1. Determine the faulty node and the set of candidate target nodes for receiving and storing the data blocks to be migrated from the distributed storage cluster, wherein the data blocks to be migrated refer to the data blocks currently stored on the faulty node.

[0029] In step S1, the distributed storage cluster refers to a data storage system composed of multiple storage nodes interconnected by a network, where each storage node is used to store data blocks. A failing node refers to a storage node currently in a sub-healthy state and / or at risk of failure. The candidate target node set refers to a set of at least one healthy storage node selected from the distributed storage cluster and used to receive and store the data blocks to be migrated (i.e., storage objects requiring guaranteed reliability). To quantify node health status for accurately identifying failing nodes as migration sources and selecting healthy nodes as migration targets, preferably, determining failing nodes and the candidate target node set for receiving and storing the data blocks to be migrated from the distributed storage cluster includes, but is not limited to, the following steps S11 to S15.

[0030] S11. Obtain multidimensional health status data of each storage node in the distributed storage cluster, wherein the multidimensional health status data includes, but is not limited to, at least one of I / O latency time series, disk error rate, bandwidth jitter amplitude, data verification failure rate, disk SMART information and SSD wear level.

[0031] In step S11, the operating status data (i.e., the multidimensional health status data) of each storage node can be periodically collected through the monitoring system of the distributed storage cluster or the local agent of each storage node. In the multidimensional health status data, the I / O latency time series reflects disk response speed, the disk error rate indicates the reliability of the storage medium, the bandwidth jitter amplitude characterizes network stability, the data verification failure rate reveals data integrity risks, and the disk SMART (Self-Monitoring, Analysis and Reporting Technology) information (including relocation sector counts and power-on time, etc.) and the SSD (Solid State Drive) wear level respectively reflect the hardware lifecycle status. The purpose of collecting the aforementioned multidimensional data is to provide comprehensive and objective input parameters for subsequent health assessments, avoiding misjudgments based on a single indicator. Furthermore, in practical applications, at least one of the aforementioned indicators can be selected for collection based on the hardware type and deployment environment of the storage node.

[0032] S12. For each storage node, the corresponding multidimensional health status data is input into the pre-trained health score model, and the health score model uses a long short-term memory network or a gated recurrent unit to extract the temporal features of the multidimensional health status data and outputs the corresponding health score.

[0033] In step S12, the health rating model is built and pre-trained using a deep learning architecture, specifically employing Long Short-Term Memory (LSTM) networks or gated recurrent units (GRUs). These recurrent neural network structures excel at processing time-series data and can capture the trend characteristics of multidimensional health status data evolving over time. The health rating model is pre-trained before deployment: historical operational data (i.e., historical multidimensional health status data) and corresponding fault labels are collected, and the model parameters are trained using supervised learning. This allows the model to output a health rating between 0 and 1 based on the input multidimensional health status data (preferably time-series data). (A higher rating indicates a healthier node, while a lower rating indicates a higher risk of node failure.) Furthermore, the health rating is preferably based on continuous time-series data, so that it can not only reflect the current state but also predict short-term health trends.

[0034] S13. Identify storage nodes whose health scores are lower than a first scoring threshold and higher than a second scoring threshold as nodes on the verge of failure, wherein the second scoring threshold is lower than the first scoring threshold.

[0035] In step S13, based on the health score output in step S12, two scoring thresholds are set to classify storage nodes into three levels: the first scoring threshold is used to distinguish between healthy nodes and unhealthy nodes, and the second scoring threshold is used to distinguish between nodes on the verge of failure and nodes that have already failed. Specifically, storage nodes with a health score higher than the first scoring threshold are classified as healthy nodes; storage nodes with a health score lower than the first scoring threshold but higher than the second scoring threshold are identified as nodes on the verge of failure. Although these nodes have not completely failed, their health status has significantly declined, posing a high risk of failure, and their data should be proactively migrated to healthy nodes; nodes with a health score lower than the second scoring threshold are considered as failed or about to fail nodes and are no longer considered as data storage targets.

[0036] S14. The storage nodes with health scores higher than the first score threshold are identified as candidate target nodes for receiving and storing data blocks to be migrated, forming a set of candidate target nodes, wherein the data blocks to be migrated refer to the data blocks currently stored on the faulty node.

[0037] In step S14, based on the classification results of step S13, healthy nodes with health scores higher than the first score threshold are selected as candidate target nodes that can receive and store the data blocks to be migrated. These nodes have good health and reliability, ensuring the security of the migrated data. Through the aforementioned screening mechanism, a reliable pool of source nodes (potentially failing nodes) and target nodes (a set of healthy nodes) can be provided for subsequent data migration decisions, ensuring the reliability and security of data migration from the source.

[0038] Based on the above steps S11 to S14, by performing time-series modeling analysis on the multidimensional health status data of storage nodes, the health status of each node can be continuously and dynamically quantified, accurately identifying nodes at risk of failure as migration sources, while selecting healthy nodes as migration targets, providing a reliable data foundation for subsequent migration path decisions.

[0039] S2. Real-time collection of migration environment status data, wherein the migration environment status data includes, but is not limited to, at least one of the following: load of critical nodes, load of candidate target nodes, network link quality, and global load distribution of the cluster.

[0040] In step S2, to support the subsequent generation of the optimal migration path by the multi-agent reinforcement learning transfer decision model, it is necessary to collect the dynamic state data of the current migration environment (i.e., the migration environment state data) in real time. In this migration environment state data, the load of the failing node reflects the processing pressure of the source storage node (i.e., the failing node), including indicators such as CPU (Central Processing Unit) utilization, memory usage, and disk I / O queue depth, to assess the available bandwidth and potential latency for data reading; the load of the candidate target node reflects the real-time processing capability of each candidate target storage node, to assess the expected time and reliability of data writing; the network link quality includes parameters such as bandwidth, latency, and packet loss rate, to assess the stability of data transmission; and the cluster global load distribution reflects the resource usage of the entire distributed storage cluster, avoiding new reliability risks caused by local overload. The aforementioned data can be obtained in real time through heartbeat reporting from each storage node, network monitoring probes, or centralized monitoring services, and used as input to the decision model.

[0041] S3. Input the migration environment state data into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and fuse the outputs of each decision agent through a group decision verification mechanism to generate migration path decision results. The migration path decision results include, but are not limited to, the selected target node, migration shard size, migration concurrent thread count and / or migration compression algorithm.

[0042] In step S3, within the multi-agent reinforcement learning transfer decision model, each decision agent, after being configured to pursue different optimization objectives (e.g., one agent aims to minimize transfer completion time, another aims to maximize target node load balancing, a third aims to minimize the impact of the transfer process on business I / O, etc.), independently outputs its own candidate transfer path decision based on the current environment state (i.e., the transfer environment state data). This decision (which includes, but is not limited to, at least one of the following: target node selection, transfer shard size, number of concurrent transfer threads, and transfer compression algorithm) is made. To achieve multi-objective collaborative optimization and improve model training efficiency and decision quality, the training process of the multi-agent reinforcement learning transfer decision model preferably includes, but is not limited to, the following steps S311-S312.

[0043] S311. For each decision agent in the at least two decision agents, construct a corresponding policy network and a value network, wherein the value network is used to evaluate the quality of the migration path decision results output by the policy network, and the optimization objective function of the policy network and the value network consists of at least two optimization objectives, including but not limited to minimizing the migration completion time, maximizing the load balancing of the target node, and / or minimizing the impact of the migration process on business I / O.

[0044] In step S311, a policy network and a value network are specifically constructed for each decision-making agent, forming an Actor-Critic architecture. The policy network (Actor), acting as the "actor" of the decision-making agent, receives the current migration environment state data as input and outputs migration path decision results (such as the selected target node, migration shard size, and number of concurrent migration threads). Its network parameters determine the quality of the decision policy. The value network (Critic), acting as the "evaluator" of the decision-making agent, receives the current migration environment state data and the decision results output by the policy network as input and outputs a quality evaluation value for the decision results, which guides the parameter update direction of the policy network. The optimization objective function of the policy network and the value network preferably consists of multiple optimization objectives, such as: minimizing migration completion time (pursuing efficient migration), maximizing target node load balancing (avoiding new hotspots after migration), and minimizing the impact of the migration process on business I / O (ensuring business continuity). By introducing multiple objectives into the optimization function, the trained model can make a comprehensive optimal decision that balances efficiency, balance, and business friendliness in complex scenarios.

[0045] S312. A training architecture combining centralized training and distributed execution is adopted: During the training phase, the experience data of each decision agent is collected centrally and shared among them, so as to use the experience data to synchronously update the policy network parameters and value network parameters of each decision agent through a multi-objective policy gradient algorithm; During the execution phase, each decision agent independently makes migration path decisions based solely on its own policy network.

[0046] In step S312, a centralized training and distributed execution architecture is adopted to solve the training challenge of environmental non-stationarity in multi-agent environments. Specifically, during the training phase, a centralized training server is set up. The experience data (including states, actions, rewards, and next states) generated by the interactions of each decision-making agent in the simulated or real environment are uniformly collected into an experience replay pool and shared among them. Using this shared experience data, the gradients of the policy network and value network in each agent can be calculated using a multi-objective policy gradient algorithm (such as MADDPG, Multi-Agent Deep Deterministic Policy Gradient, an existing algorithm for multi-agent reinforcement learning environments), and the network parameters of all agents are updated synchronously. This centralized training method enables each agent to perceive the decision intentions of other agents, thereby learning collaborative decision-making capabilities. During the execution phase, each decision-making agent no longer relies on the centralized server, but only loads its own pre-trained policy network parameters and independently makes migration path decisions based on the current environmental state without communicating with each other, ensuring the real-time nature of the decision and the scalability of the system.

[0047] Based on the above steps S311 to S312, the decoupling of decision-making and evaluation is achieved by constructing a policy network and a value network for each decision-making agent; centralized training enables each agent to learn collaborative decision-making capabilities, and distributed execution ensures decision-making efficiency; the multi-objective optimization function enables the model to take into account migration efficiency, load balancing and business impact, ultimately achieving high-quality and high-efficiency migration path decision-making.

[0048] In step S3, the group decision verification mechanism is used to perform fusion verification of all candidate decisions. That is, this mechanism performs consistency checks and conflict resolution on the outputs of each agent, comprehensively evaluates the combined advantages of each candidate path under multi-objective constraints, and finally generates a comprehensive optimal migration path decision result to guide subsequent data migration execution. To fuse multi-agent decisions and achieve the goal of ensuring the reliability and credibility of the final decision through verification and optimization, preferably, the group decision verification mechanism includes, but is not limited to, the following steps S321 to S324.

[0049] S321. Obtain candidate migration paths output by the at least two decision agents, and construct a verifiable decision topology network based on the obtained results.

[0050] In step S321, candidate migration paths independently output by each decision-making agent are first collected. Each candidate migration path is a complete migration scheme, including decision variables such as the selected target node, migration fragment size, number of concurrent migration threads, and migration compression algorithm. Based on these candidate paths, a verifiable decision topology network is constructed: this topology network uses each candidate path as a node and the relationships between paths (such as sharing the same target node and / or using similar migration parameters) as edges, forming a graph structure. The purpose of this topology network is to map the independent decisions of each agent into a graph structure, providing a computable data model for subsequent group advantage tuning and credibility convergence. Thus, by constructing the aforementioned topology network, the correlation and differences between decisions can be intuitively expressed, facilitating comprehensive analysis at the group level.

[0051] S322. Perform domain-aware group advantage tuning on each candidate migration path in the verifiable decision topology network to obtain the tuning confidence of each candidate migration path.

[0052] In step S322, specifically, prior knowledge from the distributed storage domain (such as high target node load leading to performance degradation after migration, and cross-fault domain migration improving data reliability) can be incorporated to calculate the relative advantage of each candidate path within the group. The entire optimization process considers multiple dimensions: the degree of consistency between the candidate path and other paths in the group (higher consensus with other paths indicates higher credibility), whether the path's decision variables conform to domain best practices, and the path's performance in historically similar environments. Through domain knowledge weighting and group consensus analysis, an optimization credibility score can be generated for each candidate path, reflecting its relative reliability in group decision-making.

[0053] S323. Based on the optimization credibility, perform group decision credibility convergence verification on the verifiable decision topology network to obtain the comprehensive credibility evaluation value of each candidate migration path.

[0054] In step S323, the specific process of the convergence verification of the group decision credibility is implemented through an iterative propagation mechanism: in the verifiable decision topology network, the optimization credibility of each candidate path is propagated along the edges to adjacent nodes, and adjacent nodes update their own credibility based on their own optimization credibility and the propagated information; this process is repeated until the credibility distribution of the entire network converges and stabilizes. This convergence verification mechanism draws on the idea of ​​the belief propagation algorithm, which can integrate global information to eliminate the bias of individual decisions. Finally, each candidate path obtains a comprehensive credibility evaluation value, which comprehensively considers the path's own optimization credibility and its influence in the group topology network, representing the comprehensive reliability of the path in group decision-making.

[0055] S324. The candidate migration path with the highest overall credibility rating will be used as the final migration path decision.

[0056] In step S324, a candidate migration path represents the most reliable and trustworthy migration scheme under the multi-agent group decision-making framework, after domain knowledge optimization and group consensus verification. This scheme can then be output as a decision result to the execution layer to guide subsequent data migration operations. This group decision verification mechanism avoids potential biases or blind spots in single-agent decisions, ensuring the robustness and reliability of the final decision.

[0057] Based on the above steps S321 to S324, by constructing a verifiable decision topology network, optimizing the domain-aware group advantage, and verifying the convergence of group decision credibility, the candidate path with the highest comprehensive credibility is finally selected, realizing the fusion and verification of multi-agent decision-making, and ensuring the reliability and credibility of the migration path decision results.

[0058] S4. Perform data migration on the data block to be migrated according to the migration path decision result, and continuously collect environmental status changes and the health score of the selected target node during the migration process.

[0059] In step S4, specifically, the data blocks to be migrated on the failing node are fragmented and transferred to the selected target node according to the migration fragment size, migration concurrency thread count, and migration compression algorithm specified in the decision result. During the migration, environmental status changes are continuously collected, including real-time fluctuations in the load of the failing node, changes in network link quality, and dynamic adjustments to the cluster's global load distribution. Simultaneously, the health score of the selected target node is continuously collected (the specific collection process is described in steps S21-S22 above and will not be repeated here), monitoring whether its health status changes during the migration process. This real-time collected data is used for triggering judgments in subsequent steps S5 and S6, providing a basis for decision-making regarding online replanning and task splitting fault tolerance.

[0060] S5. When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online, wherein the online replanning includes, but is not limited to, reselecting the target node.

[0061] In step S5, the expected completion time can be dynamically calculated based on factors such as the current amount of migrated data, the remaining amount of data, the current transmission rate, the target node load trend, and changes in network link quality. When it is detected that the expected completion time exceeds a preset time threshold due to environmental degradation (such as a sudden increase in target node load, a sharp drop in network bandwidth, and / or an imbalance in the global load distribution of the cluster), an online replanning mechanism is triggered. Specifically, the multi-agent reinforcement learning migration decision model is reactivated, and the latest environmental state data (including the load of critically ill nodes, the load of candidate target nodes, network link quality, and the global load distribution of the cluster) is input into the model. Each decision agent regenerates candidate migration paths, and the new migration path decision results are fused and output through a group decision verification mechanism: that is, abandoning the target node with slow response or declining health, and switching to a better candidate target node to continue the migration, so as to ensure that the migration task can be completed within the expected time and avoid migration failure or data loss due to environmental degradation.

[0062] S6. When the health score of the selected target node is detected to have decreased and fallen below the preset score threshold, the task splitting fault tolerance mechanism is triggered: for the data shards in the data block to be migrated that have completed migration, confirmation is submitted; for the data shards in the data block to be migrated that have not completed migration, the corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

[0063] In step S6, specifically, when the health score of the selected target node is detected to drop below the preset score threshold (e.g., the first score threshold), it indicates that the node's health has deteriorated during the migration process. Continuing to write data to it may lead to migration failure or data loss. At this point, a task splitting fault tolerance mechanism is triggered. This mechanism divides the current migration task into two parts: for data shards that have completed migration, confirmation is submitted directly, and the migration is marked as successful; for data shards that have not yet completed migration, one or more alternative target nodes are reallocated based on the latest environmental status data (including the load of the failing node, the load of other candidate target nodes, and network link quality, etc.). (This may require reactivating the multi-agent reinforcement learning migration decision model to achieve reallocation), and the data shards that have not yet completed migration are switched to the alternative target nodes to continue migration. Simultaneously, the migration transaction status of the data block to be migrated is updated synchronously, for example, recording the latest information of confirmed shards and shards to be migrated, ensuring the integrity and traceability of the migration transaction. Through this task splitting fault tolerance mechanism, risks can be proactively avoided when the health of the target node deteriorates, ensuring the reliability and consistency of data migration.

[0064] In step S6, in order to enhance the fault tolerance capability and achieve the purpose of multi-dimensional active fault tolerance and self-evolution of the decision model, preferably, the task splitting fault tolerance mechanism also includes, but is not limited to, the following enhancement methods (A) to (C).

[0065] (A) When the health score of the critical node is detected to have decreased and fallen below the third score threshold, the migration priority of the data block to be migrated is automatically increased, and the migration concurrency is dynamically adjusted. The aforementioned third score threshold may be lower than the aforementioned second score threshold used to identify critical nodes, marking emergency nodes whose health has severely deteriorated and are about to fail. When the health score of the critical node is detected to have fallen below the third score threshold, it indicates that the node has an extremely high risk of failure, and data evacuation must be accelerated. At this time, the migration priority of the data block to be migrated on the critical node is automatically increased, giving it higher processing weight in system resource scheduling and prioritizing bandwidth and I / O resources. Simultaneously, the migration concurrency is dynamically adjusted, i.e., based on the remaining health resources and network conditions of the critical node, the number of concurrent threads is appropriately increased to accelerate data reading and transmission, ensuring that the migration of critical data is completed before the node completely fails. Thus, the aforementioned mechanism enables proactive response to source-end deterioration, minimizing the risk of data loss.

[0066] (B) When network link quality fluctuations exceed a preset fluctuation threshold, the migration fragment size is adaptively adjusted: the fragment size is reduced on low-quality network links and increased on high-quality network links to improve transmission success rate. Specifically, network link quality parameters between the source and target nodes, including bandwidth, latency, and packet loss rate, can be collected in real time through the network monitoring module. When link quality fluctuations exceed the preset fluctuation threshold, adaptive fragment size adjustment is triggered: on low-quality network links, the fragment size is reduced, making each data fragment smaller, thereby reducing the retransmission cost of a single transmission failure and improving the transmission success rate; while on high-quality network links, the fragment size is increased, reducing the number of fragments, thereby reducing protocol overhead and improving transmission efficiency. Through this dynamic adaptation, the available network resources are maximized while ensuring transmission success rate, avoiding migration failures or efficiency degradation caused by link quality fluctuations.

[0067] (C) Record fault-tolerant events and execution results, and use the recorded results as negative samples to incrementally train the multi-agent reinforcement learning transfer decision model. Specifically, during the execution of the transfer task, record all events that trigger the fault-tolerant mechanism and their execution results, including: triggering conditions (such as the health score of a critical node dropping below the third score threshold or network link quality fluctuations exceeding the preset fluctuation threshold), fault-tolerant actions taken (such as increasing priority, adjusting concurrency, or adjusting shard size), and execution results (successfully avoiding risks or failing to migrate). Use these recorded results as negative samples (i.e., cases of anomalies or risks occurring during the migration process) to incrementally train the multi-agent reinforcement learning transfer decision model. By periodically or in real-time adding negative samples to the training set and updating the model parameters, the model can proactively choose a better migration path or trigger the fault-tolerant mechanism in advance when encountering similar environmental states in the future, thereby continuously improving the robustness and reliability of the transfer decision.

[0068] Based on the above enhancement methods (A) to (C), by prioritizing and adjusting concurrency when the source deteriorates, adapting the fragment size when the network link quality fluctuates, and incremental training of negative samples for fault-tolerant events, multi-dimensional proactive fault tolerance can be achieved from the three dimensions of the source, network, and model, forming a self-evolving closed loop, which significantly improves the robustness and reliability of the data migration process in complex environments.

[0069] In step S6, in order to ensure the reliability and consistency of the migration transaction and achieve the purpose of fine-grained anomaly handling and recovery, preferably, the migration transaction status of the data block to be migrated includes, but is not limited to, transaction identifier, node identifier of the faulty node, selected target node identifier, data block identifier, migration status, retry count, and migration transaction creation timestamp, etc. The migration status includes, but is not limited to, pending migration, in migration, confirmed, rolled back, and pending retry; the migration transaction status of the data block to be migrated is updated synchronously, including but not limited to the following steps S61 to S62.

[0070] S61. When a recoverable abnormal event is detected during the data migration process of the data block to be migrated, the migration state is switched to pending retry, and the retry count is incremented, wherein the recoverable abnormal event refers to network timeout, temporary unreachability of the target node, or resource contention.

[0071] In step S61, when a recoverable anomaly is detected, such as network timeout (i.e., data transmission exceeds the preset response time), target node temporarily unreachable (i.e., the node is temporarily offline or temporarily unresponsive due to high load), or resource contention (i.e., bandwidth or I / O resources are preempted, causing transmission obstruction), the system determines that the anomaly is temporary and non-fatal, and there is no need to immediately roll back the entire migration transaction. At this time, the migration transaction status of the data block to be migrated is switched from the current state to the "pending retry" state, indicating that the migration of the data block is temporarily blocked but can be recovered later, and the retry count is incremented to record the number of times recovery has been attempted. Thus, by introducing the "pending retry" state, recoverable anomalies can be distinguished from unrecoverable anomalies, avoiding erroneous rollbacks caused by temporary failures and improving migration efficiency.

[0072] S62. When the retry count exceeds the preset count threshold, an anomaly type analysis is triggered: if the anomaly type is that the target node is unreachable, then the target node switching action is performed and the retry count is reset; if the anomaly type is that the network timeout, then the data migration task is re-executed after reducing the transmission rate.

[0073] In step S62, specifically, when the retry count exceeds a preset threshold, it indicates that the recoverable anomaly has occurred repeatedly, and a single retry strategy is no longer sufficient to resolve the issue. At this point, an anomaly type analysis is triggered, and a differentiated recovery strategy is adopted based on the specific type of the anomaly: If the anomaly type is "target node unreachable," it indicates that the currently selected target node may have a persistent fault (such as node crash or network partition), and continuing to retry is meaningless. Therefore, a target node switching action is performed, reselecting a healthy node from the candidate target node set as the new target node (this may require reactivating the multi-agent reinforcement learning transfer decision model to achieve reselection), and resetting the retry count to zero, retrying the migration along a new path. If the anomaly type is "network timeout," it indicates that the current network link may have congestion or insufficient bandwidth. In this case, the target node is not switched; instead, the transmission rate is reduced before re-executing the migration task to reduce the impact on network resources and improve transmission stability. Thus, this anomaly type-driven differentiated recovery strategy significantly improves the recovery success rate of the migration task in complex fault scenarios.

[0074] Based on the above steps S61 to S62, by introducing the "waiting for retry" state to classify recoverable anomalies in a refined manner, and by combining the retry count threshold and anomaly type analysis, proactive switching when the target node is unreachable and adaptive rate adjustment when the network times out are realized, effectively ensuring the reliability and consistency of migration transactions.

[0075] Therefore, based on the reliability assurance method described in steps S1 to S6 above, a new data storage reliability assurance scheme is provided that integrates multi-agent reinforcement learning and group decision-making to dynamically plan migration paths and achieve online replanning and task splitting fault tolerance collaboration. First, a set of potentially failing nodes and candidate target nodes is determined, and migration environment state data is collected. Then, the data is input into a reinforcement learning model containing multi-decision agents. A migration path decision is generated through group decision-making verification. Finally, data migration is executed while continuously collecting environment state and target node health data. When the expected completion time exceeds a threshold, the model is reactivated for online replanning. When the target node health falls below a threshold, a task splitting fault tolerance mechanism is triggered, confirming migrated shards and reassigning replacement target nodes to unmigrated shards while synchronously updating the migration transaction status. Thus, through multi-agent collaborative decision-making, online dynamic adjustment, and refined fault tolerance, adaptive optimization of the migration path and proactive assurance of data reliability are achieved, facilitating practical application and promotion.

[0076] Based on the technical solution of the first aspect, this embodiment also provides a possible design for how to introduce business awareness to achieve differentiated and intelligent migration decisions, that is, the method also includes, but is not limited to, the following steps S71 to S76.

[0077] S71. Obtain the business metadata of the data object to which the data block to be migrated belongs, wherein the business metadata includes, but is not limited to, data importance level, expected access frequency, related business service level agreement requirements, and data lineage.

[0078] In step S71, the business metadata of the data object to which the data block to be migrated belongs can be obtained through metadata services or data catalogs. Specifically, the data importance level identifies the degree of impact of the data on business continuity (e.g., core transaction data, log data, and archived data); the expected access frequency reflects the expected data usage to guide storage media selection; the associated business service level agreement specifies performance indicators such as data availability and latency; and the data lineage records the dependencies between data objects, such as data A being generated from data B and data C. The aforementioned business metadata can be obtained from data governance platforms, business systems, or configuration files, providing business-dimensional input for the generation of subsequent differentiated migration strategies.

[0079] S72. Based on the historical access logs of the data object, construct a data access pattern profile of the data object, wherein the data access pattern profile is used to reflect the transformation patterns and transformation time windows of cold data, warm data and hot data.

[0080] In step S72, historical access records of data objects can be extracted from the access logs of the distributed storage cluster, including but not limited to access time, access frequency, accessing user, and access type. By performing statistical analysis (such as time series clustering and frequency statistics) on the historical access logs, an access pattern profile of the data objects can be constructed. This profile reflects the changing patterns of data access popularity over a period of time, for example: which data belongs to high-frequency access "hot data," which belongs to low-frequency access "cold data," which belongs to "warm data" between "hot data" and "cold data," and the time window for data to transition from hot to cold or from cold to hot (e.g., hot data within 7 days of being written, turning into cold data after 30 days). The aforementioned profile can provide a data-driven decision-making basis for subsequently determining the migration timing and storage level.

[0081] S73. Combining the business metadata and the data access pattern profile, a differentiated migration strategy is generated for the data object using a multi-objective optimization algorithm. The objective function of the multi-objective optimization algorithm includes, but is not limited to, storage cost, access latency, data reliability, and energy consumption. The differentiated migration strategy includes, but is not limited to, migration priority, migration time window, and atomic migration requirements.

[0082] In step S73, the multi-objective optimization algorithm (such as Pareto front analysis or NSGA-II) is used to weigh multiple objectives, including storage cost, access latency, data reliability, and energy consumption, to find the optimal combination of migration strategies. Specifically, for data with high importance and high access frequency, the multi-objective optimization algorithm assigns a higher migration priority and selects high-performance storage media as the migration target; for data with low importance and low access frequency, the migration priority is reduced, and low-cost storage media is selected. The differentiated migration strategy includes at least: migration priority (determining the execution order of migration tasks), migration time window (limiting the time period during which migration operations are allowed to be executed), and atomic migration requirements (specifying whether data objects with a lineage must be migrated as a whole). This strategy provides business constraints for subsequent migration decisions.

[0083] S74. Input the differentiated migration strategy as a constraint into the multi-agent reinforcement learning migration decision model to guide the migration path decision of the data block to be migrated.

[0084] In step S74, specifically, the migration priority constraint affects the weight allocation of different data blocks by each decision-making agent during decision-making, enabling high-priority data blocks to receive better resource scheduling; the migration time window constraint limits the execution period of the migration operation, causing the decision model to output a "wait" decision outside the window and an "execute" decision inside the window; the atomized migration requirement constrains data blocks with a lineage relationship to be assigned to the same group of target nodes, ensuring the integrity of the data after migration. Thus, by incorporating business constraints into the decision model, migration path decisions can consider not only environmental conditions but also business needs.

[0085] S75. When executing the data migration task of the data block to be migrated, the migration operation is scheduled to be executed during the predicted low business hours according to the migration time window in the differentiated migration strategy.

[0086] In step S75, specifically, future peak and off-peak business periods are predicted based on historical access patterns, and the migration operation is scheduled to be executed during the predicted off-peak periods (such as nighttime or weekends). If the migration time window is "weekday nighttime," the migration task will start during the daily off-peak period and pause or complete before the peak business period arrives; if the migration time window is "unlimited," decisions can be made in real time based on environmental conditions. By decoupling the migration operation from business periods, the impact of the migration process on normal business I / O is reduced, ensuring business continuity.

[0087] S76. For the data objects that have a data lineage relationship, according to the atomic migration requirements in the differentiated migration strategy, the migration operation is performed atomically on all data blocks contained in the data objects to ensure the lineage integrity of the data objects.

[0088] In step S76, since data lineage records the dependencies between data, such as data A being generated from data B and data C, if data B and data C are migrated at different times and on different nodes, data A may not be correctly reconstructed during the migration transition. Therefore, atomic migration requires that for a group of data objects with a lineage relationship, all data blocks contained within it must be migrated as a whole; either all blocks must be successfully migrated to the target node, or all blocks must remain unchanged. In actual execution, the system binds this group of data blocks into atomic operations through migration transactions, ensuring the integrity of data lineage during the migration process and avoiding data inconsistency issues caused by partial migration.

[0089] Based on the above possible design one, a data access pattern profile is constructed by acquiring business metadata, generating a differentiated migration strategy that includes priority, time window and atomicity requirements, and integrating this strategy as a constraint into a multi-agent reinforcement learning decision model, thereby realizing business-aware migration scheduling and lineage integrity assurance, so that migration decision can ensure data reliability while taking into account business continuity and storage resource optimization.

[0090] Based on the technical solution of the first aspect, this embodiment also provides a possible design two for how to achieve fault domain-aware replica isolation to improve physical reliability, that is, the method also includes, but is not limited to, the following steps S81 to S83.

[0091] S81. Construct a multi-level fault domain topology diagram of the distributed storage cluster, wherein the multi-level fault domain topology diagram includes the relationship between physical machines, racks, switches and data centers and historical fault statistics of each level.

[0092] In step S81, specifically, the multi-level fault domain topology map comprises four layers from bottom to top: the physical machine layer (a single storage node), the rack layer (multiple physical machines sharing the same rack power supply and network), the switch layer (multiple racks sharing the same access or aggregation switch), and the data center layer (multiple switch clusters sharing the same data center infrastructure). The relationships between each layer are determined through deployment configuration information, network topology information, and data center layout information. Simultaneously, historical fault statistics for each layer can be collected from the operation and maintenance monitoring system, fault work orders, and logs, including physical machine fault frequency, rack power outage counts, switch downtime duration, and data center disaster events. This statistical information is used to subsequently calculate the common fault probability among different storage nodes, providing a data foundation for fault domain-aware replica placement.

[0093] S82. When determining the set of candidate target nodes, calculate the common failure probability of the node about to fail and each other storage node according to the fault history statistics, and preferentially select other storage nodes with a common failure probability lower than a preset probability threshold as candidate target nodes.

[0094] In step S82, specifically, the common failure probability reflects the likelihood of two nodes failing simultaneously. Its calculation is based on whether the two nodes are located on the same physical machine, in the same rack, on the same switch, or in the same data center, and is weighted by combining historical failure statistics at each level (e.g., the common failure probability of nodes in the same rack is higher than that of nodes across racks, and the common failure probability of nodes in the same data center is higher than that of nodes across data centers). When determining the set of candidate target nodes, other storage nodes with a common failure probability lower than a preset probability threshold are preferentially selected. Therefore, the aforementioned screening mechanism ensures sufficient physical isolation between candidate target nodes and source nodes, preventing simultaneous data loss on both the source and target nodes due to a single failure domain, thus improving data reliability at the physical level.

[0095] S83. When reselecting target nodes, select the node combination that minimizes the sum of common failure probabilities within the already selected target node group.

[0096] In step S83, specifically, for the data block to be migrated, if multiple replicas are required, a group of target nodes needs to be selected. At this point, the sum of the common failure probabilities of all candidate node combinations within the group is calculated, i.e., the cumulative sum of the common failure probabilities between any two nodes within the combination. The node combination that minimizes this sum is selected as the final target node group. This optimization strategy ensures that the selected target node group has the lowest overall associated failure risk in physical deployment, avoiding placing multiple replicas on nodes with strong failure correlation (such as on the same rack or switch). Through this combination optimization, even if a single failure domain fails, only a very small number of replicas are affected, thereby maximizing the reliability of data at the physical layer.

[0097] Based on the above possible design two, by constructing a multi-level fault domain topology map and introducing fault history statistics, the dynamic common fault probability among nodes is calculated. In the selection of candidate target nodes, low probability nodes are selected first, and in the selection of target node groups, the sum of probabilities within the combination is optimized, thus realizing proactive protection of data reliability from the physical topology level.

[0098] like Figure 2 As shown, the second aspect of this embodiment provides a virtual system that implements the reliability assurance method described in the first aspect, possible design one or possible design two, including a node type determination unit, an environment state acquisition unit, a migration path decision unit, a path decision execution unit, an online replanning unit and a task splitting trigger unit; The node type determination unit is used to determine the faulty node and the set of candidate target nodes for receiving and storing the data block to be migrated from the distributed storage cluster, wherein the data block to be migrated refers to the data block currently stored on the faulty node; The environment status acquisition unit is communicatively connected to the node type determination unit and is used to acquire migration environment status data in real time. The migration environment status data includes at least one of the following: load of the node on the verge of failure, load of the candidate target node, network link quality, and global load distribution of the cluster. The migration path decision unit is communicatively connected to the environment state acquisition unit. It is used to input the migration environment state data into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and to fuse the outputs of each decision agent through a group decision verification mechanism to generate a migration path decision result. The migration path decision result includes the selected target node, migration shard size, migration concurrent thread count and / or migration compression algorithm. The path decision execution unit is communicatively connected to the migration path decision unit and is used to perform data migration on the data block to be migrated according to the migration path decision result, and continuously collect environmental status changes and the health score of the selected target node during the migration process. The online replanning unit is communicatively connected to the path decision execution unit. When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online. The online replanning includes at least reselecting the target node. The task splitting triggering unit is communicatively connected to the path decision execution unit. When the health score of the selected target node is detected to have decreased and fallen below a preset score threshold, the task splitting fault tolerance mechanism is triggered: for data shards in the data block to be migrated that have completed migration, confirmation is submitted; for data shards in the data block to be migrated that have not completed migration, a corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

[0099] The working process, working details and technical effects of the aforementioned device provided in the second aspect of this embodiment can be found in the reliability assurance method described in the first aspect, possible design one or possible design two, and will not be repeated here.

[0100] like Figure 3As shown, the third aspect of this embodiment provides a computer device for executing the reliability assurance method as described in the first aspect, possible design one, or possible design two. It includes a storage module, a processing module, and a transceiver module that are sequentially and communicatively connected. The storage module stores a computer program, the transceiver module sends and receives messages, and the processing module reads the computer program and executes the reliability assurance method as described in the first aspect, possible design one, or possible design two. Specifically, the storage module may include, but is not limited to, random-access memory (RAM), read-only memory (ROM), flash memory, first-in-first-out (FIFO) memory, and / or first-in-last-out (FILO) memory, etc.; the processing module may, but is not limited to, use a microprocessor of the STM32F105 series. Furthermore, the computer device may also include, but is not limited to, a power supply module, a display screen, and other necessary components.

[0101] The working process, working details and technical effects of the aforementioned computer device provided in the third aspect of this embodiment can be found in the reliability assurance method described in the first aspect, possible design one or possible design two, and will not be repeated here.

[0102] This fourth aspect of the embodiment provides a computer-readable storage medium storing instructions comprising the reliability assurance method as described in the first aspect, Possible Design 1, or Possible Design 2. Specifically, the computer-readable storage medium stores instructions that, when executed on a computer, perform the reliability assurance method as described in the first aspect, Possible Design 1, or Possible Design 2. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.

[0103] The working process, working details and technical effects of the aforementioned computer-readable storage medium provided in the fourth aspect of this embodiment can be found in the reliability assurance method described in the first aspect, possible design one or possible design two, and will not be repeated here.

[0104] This fifth aspect of the embodiment provides a computer program product, including a computer program or instructions, which, when executed by a computer, implement the reliability assurance method as described in the first aspect, possible design one, or possible design two. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.

[0105] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for ensuring the reliability of data storage, characterized in that, include: Determine the critical node and the set of candidate target nodes for receiving and storing the data blocks to be migrated from the distributed storage cluster, wherein the data blocks to be migrated refer to the data blocks currently stored on the critical node; Real-time acquisition of migration environment status data, wherein the migration environment status data includes at least one of the following: load of critical nodes, load of candidate target nodes, network link quality, and global load distribution of the cluster; The migration environment state data is input into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and the outputs of each decision agent are fused through a group decision verification mechanism to generate a migration path decision result. The migration path decision result includes the selected target node, migration shard size, migration concurrency thread count and / or migration compression algorithm. Data migration is performed on the data block to be migrated according to the migration path decision results, and environmental status changes and the health score of the selected target node are continuously collected during the migration process; When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online, wherein the online replanning includes at least reselecting the target node. When the health score of the selected target node is detected to have decreased and fallen below the preset score threshold, the task splitting fault tolerance mechanism is triggered: for the data shards in the data block to be migrated that have completed migration, confirmation is submitted; for the data shards in the data block to be migrated that have not completed migration, the corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

2. The reliability assurance method according to claim 1, characterized in that, Identify the critical nodes and the set of candidate target nodes for receiving and storing the migrated data blocks from the distributed storage cluster, including: Obtain multidimensional health status data of each storage node in a distributed storage cluster, wherein the multidimensional health status data includes at least one of I / O latency time series, disk error rate, bandwidth jitter amplitude, data verification failure rate, disk SMART information and SSD wear level; For each storage node, the corresponding multidimensional health status data is input into the pre-trained health score model, and the health score model uses a long short-term memory network or a gated recurrent unit to extract the temporal features of the multidimensional health status data and outputs the corresponding health score. Storage nodes whose health scores are lower than a first score threshold but higher than a second score threshold are identified as nodes on the verge of failure, wherein the second score threshold is lower than the first score threshold; Storage nodes with health scores higher than the first score threshold are identified as candidate target nodes for receiving and storing data blocks to be migrated, forming a set of candidate target nodes. The data blocks to be migrated refer to the data blocks currently stored on the faulty node.

3. The reliability assurance method according to claim 1, characterized in that, The group decision verification mechanism includes: Obtain candidate migration paths output by the at least two decision agents, and construct a verifiable decision topology network based on the obtained results; Perform domain-aware group advantage tuning on each candidate migration path in the verifiable decision topology network to obtain the tuning confidence of each candidate migration path; Based on the optimization credibility, the verifiable decision topology network is subjected to group decision credibility convergence verification to obtain the comprehensive credibility evaluation value of each candidate migration path. The candidate migration path with the highest overall credibility score will be used as the final migration path decision.

4. The reliability assurance method according to claim 1, characterized in that, The method further includes: Obtain the business metadata of the data object to which the data block to be migrated belongs, wherein the business metadata includes the data importance level, expected access frequency, related business service level agreement requirements, and data lineage; Based on the historical access logs of the data object, a data access pattern profile of the data object is constructed. The data access pattern profile is used to reflect the transformation patterns and transformation time windows of cold data, warm data and hot data. Combining the business metadata and the data access pattern profile, a differentiated migration strategy is generated for the data object using a multi-objective optimization algorithm. The objective function of the multi-objective optimization algorithm includes storage cost, access latency, data reliability, and energy consumption. The differentiated migration strategy includes migration priority, migration time window, and atomic migration requirements. The differentiated migration strategy is input as a constraint into the multi-agent reinforcement learning migration decision model to guide the migration path decision of the data block to be migrated. When performing the data migration task of the data block to be migrated, the migration operation is scheduled to be performed during the predicted low business hours according to the migration time window in the differentiated migration strategy. For the data objects that have a data lineage, in accordance with the atomic migration requirements in the differentiated migration strategy, the migration operation is performed atomically on all data blocks contained in the data objects to ensure the lineage integrity of the data objects.

5. The reliability assurance method according to claim 1, characterized in that, The method further includes: Construct a multi-level fault domain topology diagram of the distributed storage cluster, wherein the multi-level fault domain topology diagram includes the relationship between physical machines, racks, switches and data centers and historical fault statistics of each level; When determining the set of candidate target nodes, the common failure probability of the node about to fail and each other storage node is calculated based on the fault history statistics, and other storage nodes with a common failure probability lower than a preset probability threshold are selected as candidate target nodes. When reselecting target nodes, choose the node combination that minimizes the sum of common failure probabilities within the already selected target node group.

6. The reliability assurance method according to claim 1, characterized in that, The task splitting fault tolerance mechanism also includes: When the health score of the critical node is detected to have dropped below the third score threshold, the migration priority of the data block to be migrated is automatically increased, and the migration concurrency is dynamically adjusted. And / or, when network link quality fluctuations are detected to exceed a preset fluctuation threshold, the migration fragment size is adaptively adjusted: the fragment size is reduced on low-quality network links and the migration fragment size is increased on high-quality network links to improve the transmission success rate; And / or, record fault-tolerant events and execution results, and use the recorded results as negative samples to incrementally train the multi-agent reinforcement learning transfer decision model.

7. The reliability assurance method according to claim 1, characterized in that, The migration transaction status of the data block to be migrated includes the transaction identifier, the node identifier of the faulty node, the selected target node identifier, the data block identifier, the migration status, the retry count, and the migration transaction creation timestamp. The migration status includes pending migration, in migration, confirmed, rolled back, and pending retry. Synchronously update the migration transaction status of the data blocks to be migrated, including: When a recoverable abnormal event is detected during the data migration process of the data block to be migrated, the migration status is switched to pending retry and the retry count is incremented. The recoverable abnormal event refers to network timeout, temporary unreachability of the target node, or resource contention. When the retry count exceeds the preset count threshold, an anomaly type analysis is triggered: if the anomaly type is that the target node is unreachable, then the target node switching action is performed and the retry count is reset; if the anomaly type is that the network timeout, then the data migration task is re-executed after reducing the transmission rate.

8. The reliability assurance method according to claim 1, characterized in that, The training process of the multi-agent reinforcement learning transfer decision model includes: For each of the at least two decision agents, a corresponding policy network and a value network are constructed. The value network is used to evaluate the quality of the migration path decision results output by the policy network. The optimization objective function of the policy network and the value network consists of at least two optimization objectives, including minimizing the migration completion time, maximizing the target node load balancing, and / or minimizing the impact of the migration process on business I / O. A training architecture combining centralized training and distributed execution is adopted: During the training phase, the experience data of each decision agent is collected centrally and shared among them, so as to use the experience data to synchronously update the policy network parameters and value network parameters of each decision agent through a multi-objective policy gradient algorithm; During the execution phase, each decision agent independently makes migration path decisions based solely on its own policy network.

9. A data storage reliability assurance system, characterized in that, It includes a node type determination unit, an environment status acquisition unit, a migration path decision-making unit, a path decision execution unit, an online replanning unit, and a task splitting trigger unit; The node type determination unit is used to determine the faulty node and the set of candidate target nodes for receiving and storing the data block to be migrated from the distributed storage cluster, wherein the data block to be migrated refers to the data block currently stored on the faulty node; The environment status acquisition unit is communicatively connected to the node type determination unit and is used to acquire migration environment status data in real time. The migration environment status data includes at least one of the following: load of the node on the verge of failure, load of the candidate target node, network link quality, and global load distribution of the cluster. The migration path decision unit is communicatively connected to the environment state acquisition unit. It is used to input the migration environment state data into a multi-agent reinforcement learning migration decision model containing at least two decision agents with different optimization objectives, and to fuse the outputs of each decision agent through a group decision verification mechanism to generate a migration path decision result. The migration path decision result includes the selected target node, migration shard size, migration concurrent thread count and / or migration compression algorithm. The path decision execution unit is communicatively connected to the migration path decision unit and is used to perform data migration on the data block to be migrated according to the migration path decision result, and continuously collect environmental status changes and the health score of the selected target node during the migration process. The online replanning unit is communicatively connected to the path decision execution unit. When it is detected that the expected completion time of the current migration path of the data block to be migrated exceeds a preset time threshold due to changes in environmental conditions, the multi-agent reinforcement learning migration decision model is reactivated to replan the current migration path online. The online replanning includes at least reselecting the target node. The task splitting triggering unit is communicatively connected to the path decision execution unit. When the health score of the selected target node is detected to have decreased and fallen below a preset score threshold, the task splitting fault tolerance mechanism is triggered: for data shards in the data block to be migrated that have completed migration, confirmation is submitted; for data shards in the data block to be migrated that have not completed migration, a corresponding replacement target node is reallocated according to the current environment status, and the migration transaction status of the data block to be migrated is updated synchronously.

10. A computer device, characterized in that, The device includes a storage module, a processing module, and a transceiver module that are sequentially connected in communication. The storage module is used to store a computer program, the transceiver module is used to send and receive messages, and the processing module is used to read the computer program and execute the reliability assurance method as described in any one of claims 1 to 8.

Images