Distributed micro-service architecture service component efficient management and control method based on intelligent learning model

Abstract

The invention discloses a distributed micro-service architecture service component efficient management and control method based on an intelligent learning model, and belongs to the field of intelligent operation and maintenance, and the method comprises the steps: S1, constructing a multi-granularity service component map, collecting service calling data, mapping the service calling data into a third-order tensor, eliminating non-service flow, and carrying out smooth processing; s2, performing parallel orthogonal subspace slice mapping on the tensor, and extracting an abnormal path subtensor set; s3, constructing a dynamic health assessment model based on the heterogeneous service dependency graph, and fusing indexes such as response time delay and overload frequency; s4, constructing a nested meta-learning model fusing support vector classification and graph convolution, and generating a strategy group; s5, expanding a service behavior response function in a disturbance environment, and improving the coverage rate of a strategy to a rare state; s6, based on an execution engine of a rescheduling factor, guiding the resources to be redirected to a high-aggregation-degree substructure; and S7, comparing the structure entropy change to judge whether to enter the next round of intervention or adjust the disturbance source. The beneficial effects are that abnormity is accurately detected, health is dynamically evaluated, and scheduling efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance, and more specifically, to an efficient management and control method for service components of a distributed microservice architecture based on an intelligent learning model. Background Art

[0002] With the widespread adoption of microservices architectures, distributed systems have significantly improved in terms of flexibility and scalability. To ensure stable system operation, traditional management and control methods have gradually evolved from static configuration management to intelligent strategies with real-time monitoring, automatic circuit breaking, and service degradation capabilities. Microservices management and control architectures based on service meshes (such as Istio) and container orchestration platforms (such as Kubernetes) have implemented core functions such as metric collection, service discovery, and load balancing, driving the development of microservices systems towards automation and adaptability.

[0003] However, existing technologies still have several shortcomings: the accuracy of anomaly perception is limited, relying on static threshold rules, making it difficult to accurately capture complex anomaly patterns; the ability to identify fault propagation is weak, and there is a lack of deep modeling of topological dependencies between services; the response mechanism is extensive, using global circuit breakers or static current limiting, which can easily lead to resource waste or incorrect service blocking; the service recovery mechanism is lagging, failing to achieve aggregation of abnormal nodes and local self-healing scheduling; and it lacks adaptability to heterogeneous environments, making it difficult to achieve consistent governance under a multi-cloud, cross-platform architecture. Summary of the Invention

[0004] The purpose of the present invention is to provide an efficient management and control method for service components of a distributed microservice architecture based on an intelligent learning model to solve the problems raised in the above-mentioned background technology: the accuracy of anomaly perception is limited, and it relies on static threshold rules, making it difficult to accurately capture complex anomaly patterns; the fault propagation identification capability is weak, and there is a lack of deep modeling of topological dependencies between services; the response mechanism is extensive, and the use of global circuit breakers or static current limiting can easily lead to resource waste or incorrect service closure; the service recovery mechanism lags behind, and fails to achieve aggregation and local self-healing scheduling of abnormal nodes; it lacks adaptability to heterogeneous environments, and it is difficult to achieve consistent governance under a multi-cloud, cross-platform architecture.

[0005] Technical solution: The efficient management and control method for service components in a distributed microservice architecture based on an intelligent learning model includes the following steps: S1. Build a multi-granularity distributed service component graph by collecting 5-minute granularity historical call data between all service nodes in the microservice cluster, extracting call volume, call latency, standard deviation, average resource utilization, including multi-dimensional indicators of CPU / memory, and mapping them into a third-order tensor , n is the number of components, and t is the number of time windows; non-business requests such as system liveness detection and health check are excluded from the call relationship, and exponential weighted smoothing is used for access layer traffic; S2. Perform tensor slice mapping based on parallel orthogonal subspace decomposition on the third-order tensor, construct a path dimension projection tensor from the third-order tensor in the direction of the fixed service path, and introduce minimum cosine similarity correction to extract a set of abnormal path subtensors with reconstruction inconsistencies in the static call sequence. The abnormal path subtensors are the candidate bottleneck substructure set. S3. Construct a dynamic service health assessment model, which is based on a heterogeneous service dependency graph. Based on, V is the service node set, E is the dependent edge set, Deployment metadata mapping function for the service, It is a state attribute mapping function, combined with the deployment affinity measurement function , is the average response delay between nodes, is the overload frequency ratio; S4. Construct a nested meta-learning model for cross-cycle training, using a fusion architecture of support vector classification and graph convolution. In each learning cycle, the service state transition chain is used as the input sequence, and the strategy function space is selected based on the forward maximum entropy. Perform preliminary constraints and train the basic strategy group in the nested structure; S5. After injecting the disturbance training environment, according to the original service state-behavior distribution , perform interpolation and expansion of multiple groups of behavioral perturbations, establish behavioral response function surfaces through counterfactual path playback, and improve the behavioral coverage of intervention strategies in rare state areas; S6. Introducing a dynamic reconfiguration execution engine based on the deployment node traffic rescheduling factor. The reconfiguration logic is based on distributed behavior scenario simulation, taking into account cross-AZ communication costs and real-time link RTT changes, and redirecting resource deployment to local high-density substructures. S7. Generate a new service call tensor graph based on the rescheduling results, perform service topology entropy change model calculation, and compare the changes in structural entropy before and after. The next round of intervention evaluation is performed when the intervention is negative, otherwise a rollback is triggered and the disturbance source is iteratively adjusted.

[0006] Preferably, the step S1 of constructing a multi-granularity distributed service component graph further includes the following steps: S1-1. Set up the data collection framework as an embedded observation agent, deployed on the network interface of each Kubernetes node Pod, to collect link layer call data, HTTP message headers, gRPC request bodies, and TCP latency; call time series are uniformly numbered in Lamport logical sequence according to distributed timestamps; S1-2. Structuring the collected data into a three-dimensional time-service-metric tensor. The time dimension is layered and constructed according to sliding windows, with a minimum granularity of 5 minutes. A node hardware resource baseline model is introduced to standardize the service peak resource utilization to form a comparable indicator vector. ; S1-3. When building a static microservice path map, filter the instantaneous peak traffic caused by scheduling behavior, including the CI / CD change deployment process, and use a three-point median filter smoothing mechanism to eliminate occasional data jumps; the final topology sparsity is sparse interaction graph of ; Preferably, in the bottleneck path identification stage, a cross-period structural inconsistency rate indicator is introduced to dynamically identify path nodes in the same service chain that exhibit significant behavioral differences in multiple time windows. The indicator is defined as: ; Among them, the is the calling frequency at time t, is the mean, It is the reconstruction factor of the AZ to which the service belongs; the indicator is used as the bottleneck path pruning weight to serve the path sub-tensor priority selection strategy.

[0007] Preferably, the S1-2 further comprises the following steps: S1-2-1. Introduce the deployment node delay disturbance fitting function in the standardization processing stage, and build a response delay-load fitting model for each node in the form of , is the actual response time curve of the node; S1-2-2. Perform first-order derivative trend analysis on the fitting curve and screen out nodes with first-order derivative fluctuations greater than 3 times the median offset. These nodes are treated as unstable node sets and subsequently subjected to separate attention weight noise reduction in service health assessment.

[0008] Preferably, the S3 constructing a dynamic service health assessment model includes the following steps: S3-1. Use a structurally heterogeneous graph model to organize service components into a heterogeneous node set based on deployment location, resource affinity, and call logic. Node attributes include the standard deviation of instantaneous CPU usage, memory utilization fluctuation frequency, and the sliding average of interface call success rate. S3-2. Graph modeling uses a re-parameterized structured attention mechanism, where each attention path is controlled by a dual factor, the node resource behavior angle difference and the historical delay covariance of neighbor paths ; S3-3. The output service embedding vector enters the state anomaly regressor, and the output score is a continuous distribution [0, 1]. Nodes with an anomaly greater than 0.8 will be included in the perturbation simulation critical path and fed back to the strategy model for feature adjustment.

[0009] Preferably, the S3-2 further comprises the following steps: S3-2-1. Introducing the node scheduling history state transition probability in the attention score calculation , forming a dynamic attention distribution with constrained normalization to avoid hot nodes from interfering with the expansion of the neighborhood; S3-2-2. Apply cluster-driven compression after embedding update to cluster and compress nodes with a Cosine distance less than 0.1 in the embedding space to reduce feature redundancy of high-frequency adjacent nodes, and output the compressed representation for downstream prediction.

[0010] Preferably, the step S6 further comprises the following steps: S6-1. Construct a betweenness centrality-time evolution graph for each component and select components with betweenness value slope greater than a threshold of 0.3 and a standard deviation less than 0.15 as long-term high centrality path nodes and as migration targets; S6-2. Migration strategy considering deployment cost function ,in is the startup delay, D is the cross-domain data migration amount, and a multi-target arrangement model is established in the target node set according to the cost gradient; S6-3. The reorganization structure verifies that the call volume of the migration path is restored to more than 90% of the original value within 10 minutes by calling the consistency verifier. If it fails to meet the standard, it will automatically roll back and enter the backup reconstruction path.

[0011] Preferably, the S1-2-2 further comprises the following steps: S1-2-2-1. Build a high-order Fourier coefficient change rate sequence based on the historical 3-hour delay sequence, extract the main harmonics of the delay variation, and establish a symmetrical envelope coefficient model to determine whether the jump is caused by system-level periodic fluctuations and local faults; S1-2-2-2. Remove the periodic fluctuation component from the original sequence and apply a second-order sliding minimum variance filter to the residual term to obtain the actual business behavior deviation trend for microservice evolution modeling.

[0012] Preferably, the S3-3 further comprises the following steps: S3-3-1. Construct an abnormal service node aggregation pool, in which members of the abnormal service node aggregation pool are subjected to subgraph cutting in a multi-scale recursive neighborhood aggregation manner to form a spin group structure; S3-3-2. With the spin group as the center, a spherical adjacent boundary is constructed and the dynamic growth coefficient of the anomaly radius is calculated. When the anomaly propagation speed exceeds a fixed threshold of 0.8× the average neighborhood response rate, the local path freezing mechanism is triggered to block the anomaly diffusion chain.

[0013] Compared with the prior art, the advantages of the present invention are: (1) By modeling the call relationship between services through models such as graph convolutional networks (GCN) / GraphSAGE, it is possible to identify abnormal propagation paths and service bottlenecks in real time, which is superior to traditional static threshold rules or monitoring indicator triggering methods.

[0014] (2) Innovatively adopt multi-scale recursive neighborhood aggregation to construct a “spin group” subgraph to improve the accuracy of abnormal cluster identification and replace the traditional flat fault attribution mechanism.

[0015] (3) The propagation boundary is constructed using a spherical model, and the anomaly radius is dynamically adjusted in combination with the anomaly diffusion rate to achieve an elastic response range, which is superior to the existing fixed threshold alarm and processing mechanism.

[0016] (4) Introducing a local topology freezing strategy (edge ​​label freezing + service fallback flow control) can more accurately control abnormal areas and avoid cascading amplification compared to conventional global circuit breaking.

[0017] (5) Comprehensively consider the call chain information and service health distribution, form a closed-loop perception and feedback at the model level, and improve the level of intelligent management and control.

[0018] (6) It can be dynamically deployed and migrated in multi-platform environments such as Kubernetes and Service Mesh, and can perform service optimization scheduling based on resource status.

[0019] (7) Achieve a closed-loop control time of 1 to 2 seconds from abnormality recognition to freeze response, significantly improving fault recovery speed and system stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a schematic diagram of the overall process of the distributed microservice architecture service component efficient management and control method based on the intelligent learning model of the present invention. DETAILED DESCRIPTION

[0021] For examples, see Figure 1 The efficient management and control method for distributed microservice architecture service components based on intelligent learning models includes the following steps: S1. Build a multi-granularity distributed service component graph by collecting 5-minute granularity historical call data between all service nodes in the microservice cluster, extracting call volume, call latency, standard deviation, average resource utilization, including multi-dimensional indicators of CPU / memory, and mapping them into a third-order tensor , n is the number of components, t is the number of time windows; non-business requests such as system liveness detection and health check are excluded from the call relationship, and exponential weighted smoothing is applied to the access layer traffic; S2. Perform tensor slice mapping on the third-order tensor based on parallel orthogonal subspace decomposition. Construct a path-dimensional projection tensor from the third-order tensor in the direction of a fixed service path. Minimum cosine similarity correction is introduced to extract a set of abnormal path subtensors with reconstruction inconsistencies in the static call sequence. These abnormal path subtensors are considered candidate bottleneck substructures. S3. Build a dynamic service health assessment model, which is based on heterogeneous service dependency graphs. As the basis, V is the service node set, E is the dependent edge set, Mapping functions to the deployment metadata of a service, It is a state attribute mapping function, combined with the deployment affinity measurement function , is the average response delay between nodes, is the overload frequency ratio; S4. Construct a nested meta-learning model for cross-cycle training, using a fusion architecture of support vector classification and graph convolution. In each learning cycle, the service state transition chain is used as the input sequence, and the strategy function space is selected based on the forward maximum entropy. Perform preliminary constraints and train the basic strategy group in the nested structure; S5. After injecting the disturbance training environment, according to the original service state-behavior distribution , perform interpolation and expansion of multiple groups of behavioral perturbations, establish behavioral response function surfaces through counterfactual path playback, and improve the behavioral coverage of intervention strategies in rare state areas; S6. Introducing a dynamic reconfiguration execution engine based on the deployment node traffic rescheduling factor. The reconfiguration logic is based on distributed behavior scenario simulation, taking into account cross-AZ communication costs and real-time link RTT changes, and redirecting resource deployment to local high-density substructures. S7. Generate a new service call tensor graph based on the rescheduling results, perform service topology entropy change model calculation, and compare the changes in structural entropy before and after. The next round of intervention evaluation is performed when the intervention is negative, otherwise a rollback is triggered and the disturbance source is iteratively adjusted.

[0022] Specifically, the hardware operating environment should include a multi-node Kubernetes cluster, with nodes equipped with at least a 4-core CPU and 8GB of memory, and support for container sidecar proxy communication tracking. Algorithm framework: Tensor mapping and anomaly detection use TensorLy to implement tensor decomposition, and nested meta-learning uses PyTorch+PyG to implement a graph convolution-support vector machine hybrid model; Software environment: Support for gRPC and HTTP2 communication tracing, and support for Istio or Linkerd as a service mesh injection proxy. Training data: The data required for model training comes from historical call traces collected from real online microservice clusters, covering at least a two-week data window, including three layers of business domains and approximately 50-300 service instances; Graph structure mechanism: All service components are considered as graph nodes, and call relationships are edges. Cross-time dependency tensors are introduced to dynamically update edge weights. Delay mechanism: All path delays are dynamically estimated using a sliding window average and anomaly detection combination through multi-path shortest-term response simulation.

[0023] S1 builds a multi-granularity distributed service component graph and also includes the following steps: S1-1. Set up the data collection framework as an embedded observation agent, deployed on the network interface of each Kubernetes node Pod, to collect link layer call data, HTTP message headers, gRPC request bodies, and TCP latency; call time series are uniformly numbered in Lamport logical sequence according to distributed timestamps; S1-2. Structuring the collected data into a three-dimensional time-service-metric tensor. The time dimension is layered and constructed according to sliding windows, with a minimum granularity of 5 minutes. A node hardware resource baseline model is introduced to standardize the service peak resource utilization to form a comparable indicator vector. ; S1-3. When building a static microservice path map, filter the instantaneous peak traffic caused by scheduling behavior, including the CI / CD change deployment process, and use a three-point median filter smoothing mechanism to eliminate occasional data jumps; the final topology sparsity is sparse interaction graph of ; Specifically, the observation agent architecture: Based on the eBPF and Sidecar dual-channel collector, it is deployed at each Pod network interface and can non-invasively collect TCP connection information and application layer metadata; Data structure construction: Use Pandas + NumPy to build a third-order tensor data structure. The time dimension is aggregated using a sliding window. The service dimension is mapped using a unique service identifier. The indicator dimension contains five resource attributes. Topology graph construction tool: Uses NetworkX to build static service graphs and uses CSR compressed sparse representation to ensure that the topology graph can be used for efficient GPU batch modeling; Call smoothing mechanism: Exponentially weighted average flow rate (EWMA) and median filtering are used to jointly process traffic mutations to eliminate short-term fluctuations caused by unconventional behaviors such as CI / CD.

[0024] During the bottleneck path identification phase, the cross-period structural inconsistency rate indicator is introduced to dynamically identify path nodes in the same service chain that exhibit significant behavioral differences in multiple time windows. The indicator is defined as: ; in, is the calling frequency at time t, is the mean, It is the reconstruction factor of the AZ to which the service belongs; the indicator serves as the bottleneck path pruning weight and serves the path sub-tensor priority selection strategy.

[0025] Specifically, the inconsistency rate calculation mechanism is as follows: a call frequency vector is constructed with each time window as a slice unit, and a sliding window difference convolution is used to extract sequence differences; Definition of the AZ reconstruction factor: This factor represents the deployment tendency of the current service in different availability zones. It is used to estimate the impact of network jitter on behavioral differences. Priority selection strategy implementation: Use the maximum entropy path selection strategy to retain the top 10% paths with the highest inconsistency rate among multiple abnormal candidate paths as intervention input; Performance target: The bottleneck identification module can complete 500+ path identification and weight ranking within a 1-minute window, suitable for large-scale cluster environments.

[0026] S1-2 also includes the following steps: S1-2-1. Introduce the deployment node delay disturbance fitting function in the standardization processing stage, and build a response delay-load fitting model for each node in the form of , is the actual response time curve of the node; S1-2-2. Perform first-order derivative trend analysis on the fitting curve and screen out nodes with first-order derivative fluctuations greater than 3 times the median offset. These nodes are treated as unstable node sets and subsequently subjected to separate attention weight noise reduction in service health assessment.

[0027] Specifically, the response delay fitting function: the least squares method is used to build a delay-load polynomial model on each node, with the default second-order fitting, and the calculated slope change reflects the stability of the node performance; Unstable node screening: Based on the 3σ rule, nodes with abnormal first-order derivative fluctuations are identified, denoted as set U, and additional penalty terms are introduced into the evaluation model to adjust the weights; Model toolchain: Fitting and derivative analysis are implemented using SciPy+NumPy, and support GPU-accelerated computing to improve throughput.

[0028] S3 builds a dynamic service health assessment model including the following steps: S3-1. Use a structurally heterogeneous graph model to organize service components into a heterogeneous node set based on deployment location, resource affinity, and call logic. Node attributes include the standard deviation of instantaneous CPU usage, memory utilization fluctuation frequency, and the sliding average of interface call success rate. S3-2. Graph modeling uses a re-parameterized structured attention mechanism, where each attention path is controlled by a dual factor, the node resource behavior angle difference and the historical delay covariance of neighbor paths ; S3-3. The output service embedding vector enters the state anomaly regressor, and the output score is a continuous distribution [0, 1]. Nodes with an anomaly greater than 0.8 will be included in the perturbation simulation critical path and fed back to the strategy model for feature adjustment.

[0029] Specifically, the graph structure model is defined as follows: deployment location, resource affinity, and call chain are defined as three types of heterogeneous edges, and service node attributes are converted into vector form to form a node feature matrix; Behavior Angle Index: Calculates the angle between service resource fluctuation vectors to construct a behavioral difference attention factor, reflecting the consistency and resonance characteristics of resource utilization; State anomaly regressor: This is implemented based on a multi-layer perceptron (MLP) regression model. It generates an anomaly score for each service node, uses ReLU activation, and normalizes it to the range of 0 to 1.

[0030] S3-2 also includes the following steps: S3-2-1. Introducing the node scheduling history state transition probability in the attention score calculation , forming a dynamic attention distribution with constrained normalization to avoid hot nodes from interfering with the expansion of the neighborhood; S3-2-2. Apply cluster-driven compression after embedding update to cluster and compress nodes with a Cosine distance less than 0.1 in the embedding space to reduce feature redundancy of high-frequency adjacent nodes, and output the compressed representation for downstream prediction.

[0031] Specifically, the dynamic attention distribution mechanism: establishes a Markov state transition matrix for the node's historical scheduling state and generates a probability prior to control the normalization boundary of the attention score; Clustering compression operation: Using the KMeans+DBSCAN joint compression model, clustering processing is performed on the GPU to significantly reduce the redundancy of high-density neighborhood features; Downstream prediction applications: Compressed vector input is used for graph convolution to predict changes in service health status, shortening prediction time by more than 30%.

[0032] S6 further comprises the following steps: S6-1. Construct a betweenness centrality-time evolution graph for each component and select components with betweenness value slope greater than a threshold of 0.3 and a standard deviation less than 0.15 as long-term high centrality path nodes and as migration targets; S6-2. Migration strategy considering deployment cost function ,in is the startup delay, D is the amount of cross-domain data migration, and a multi-target arrangement model is established in the target node set according to the cost gradient; S6-3. The reorganization structure verifies that the call volume of the migration path is restored to more than 90% of the original value within 10 minutes by calling the consistency verifier. If it fails to meet the standard, it will automatically roll back and enter the backup reconstruction path.

[0033] Specifically, the betweenness centrality graph is constructed: with the service node as the center, the betweenness graph is constructed based on the call volume statistics and the frequency of the impact path, and a time-weighted sliding window is used for updating; Deployment cost function definition: including cold start time, image pull time, resource binding time, and data migration size (in MB), forming a multi-objective optimization constraint; Consistency validator mechanism: The main metric is the degree of call recovery. A traffic tracking mechanism is used to perform a sliding mean comparison within 10 minutes after migration.

[0034] S1-2-2 also includes the following steps: S1-2-2-1. Build a high-order Fourier coefficient change rate sequence based on the historical 3-hour delay sequence, extract the main harmonics of the delay variation, and establish a symmetrical envelope coefficient model to determine whether the jump is caused by system-level periodic fluctuations and local faults; S1-2-2-2. Remove the periodic fluctuation component from the original sequence and apply a second-order sliding minimum variance filter to the residual term to obtain the actual business behavior deviation trend for microservice evolution modeling.

[0035] Specifically, Fourier coefficient change modeling: the historical 3-hour delay series is converted into frequency domain expression, and the first five main harmonic coefficients are extracted to construct the change rate feature vector; Periodic fluctuation identification method: By calculating the consistency of the upper and lower boundaries of the envelope function, the one with high symmetry is determined to be a systematic fluctuation, and the others are considered to be local anomalies; Fault diagnosis model: Integrated EMD (empirical mode decomposition) further decouples periodic jumps from noise fluctuations, improving fault detection accuracy by more than 20%.

[0036] S3-3 also includes the following steps: S3-3-1. Construct an abnormal service node aggregation pool. The members of the abnormal service node aggregation pool are subgraph cut using a multi-scale recursive neighborhood aggregation method to form a spin group structure. S3-3-2. With the spin group as the center, a spherical adjacent boundary is constructed and the dynamic growth coefficient of the anomaly radius is calculated. When the anomaly propagation speed exceeds a fixed threshold of 0.8× the average neighborhood response rate, the local path freezing mechanism is triggered to block the anomaly diffusion chain.

[0037] Specifically, the hardware environment requirements are as follows: Graph processing nodes with GPU acceleration (such as NVIDIA T4 or A100) must be deployed to support processing tens of thousands of service call relationships per second and dynamic graph composition; Abnormal node aggregation pool construction mechanism: By introducing a recursive k-order neighborhood iteration strategy, the upstream and downstream dependency sets of each abnormal node are expanded based on GraphSAGE, and multi-scale candidate subgraphs including the central node are dynamically generated; Subgraph cutting and spin group structure formation: The Louvain algorithm is used to divide the graph into communities, and the node clustering and edge sparsity are jointly weighted to achieve the construction of highly cohesive spin groups. The spin groups are topologically stable fault cluster structures. Spherical adjacent boundary definition: With the geometric center of the spin group as the sphere center and the neighborhood Hop distance as the radius, an "abnormal propagation shell" is dynamically formed for modeling the abnormal radiation range; Calculation method of dynamic growth coefficient of anomaly radius: Assume that the initial anomaly radius is , using the number of newly added abnormal nodes and the edge density change rate per unit time as weight factors, dynamically calculate: ; in Represents the increment of abnormal nodes in the i-th time step; Diffusion speed threshold calculation: Based on the average neighborhood response rate (ARCR), the abnormal propagation speed is set ,when It is judged as uncontrolled transmission; The path freezing mechanism is implemented by setting a "freeze label" on all edges connecting the spin group, inserting flow limiting and strong backoff logic in the service mesh proxy, and guiding abnormal outflow paths to degenerate to an isolated state through traffic mirroring and version rollback. System response time requirement: From the triggering of the abnormal radius dynamic exceeding the threshold to the freezing taking effect, the entire response chain is required to be completed in a closed loop within 1 to 2 seconds to avoid cascading amplification.

[0038] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed in the present invention is defined by the attached claims and their equivalents.

Claims

1. An efficient management and control method for distributed microservice architecture service components based on an intelligent learning model, characterized by: The method for efficiently managing and controlling service components in a distributed microservice architecture based on an intelligent learning model includes the following steps: S1. Build a multi-granularity distributed service component graph by collecting 5-minute granularity historical call data between all service nodes in the microservice cluster, extracting call volume, call latency, standard deviation, average resource utilization, including multi-dimensional indicators of CPU / memory, and mapping them into a third-order tensor , n is the number of components, and t is the number of time windows; non-business requests such as system liveness detection and health check are excluded from the call relationship, and exponential weighted smoothing is used for access layer traffic; S2. Perform tensor slice mapping based on parallel orthogonal subspace decomposition on the third-order tensor, construct a path dimension projection tensor from the third-order tensor in the direction of the fixed service path, and introduce minimum cosine similarity correction to extract a set of abnormal path subtensors with reconstruction inconsistencies in the static call sequence. The abnormal path subtensors are the candidate bottleneck substructure set. S3. Construct a dynamic service health assessment model, which is based on a heterogeneous service dependency graph. Based on, V is the service node set, E is the dependent edge set, Deployment metadata mapping function for the service, It is a state attribute mapping function, combined with the deployment affinity measurement function , is the average response delay between nodes, is the overload frequency ratio; S4. Construct a nested meta-learning model for cross-cycle training, using a fusion architecture of support vector classification and graph convolution. In each learning cycle, the service state transition chain is used as the input sequence, and the strategy function space is selected based on the forward maximum entropy. Perform preliminary constraints and train the basic strategy group in the nested structure; S5. After injecting the disturbance training environment, according to the original service state-behavior distribution , perform interpolation and expansion of multiple groups of behavioral perturbations, establish behavioral response function surfaces through counterfactual path playback, and improve the behavioral coverage of intervention strategies in rare state areas; S6. Introducing a dynamic reconfiguration execution engine based on the deployment node traffic rescheduling factor. The reconfiguration logic is based on distributed behavior scenario simulation, taking into account cross-AZ communication costs and real-time link RTT changes, and redirecting resource deployment to local high-density substructures. S7. Generate a new service call tensor graph based on the rescheduling results, perform service topology entropy change model calculation, and compare the changes in structural entropy before and after. The next round of intervention evaluation is performed when the intervention is negative, otherwise a rollback is triggered and the disturbance source is iteratively adjusted.

2. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 1 is characterized in that: The S1 construction of a multi-granularity distributed service component graph further includes the following steps: S1-1. Set up the data collection framework as an embedded observation agent, deployed on the network interface of each Kubernetes node Pod, to collect link layer call data, HTTP message headers, gRPC request bodies, and TCP latency; call time series are uniformly numbered in Lamport logical sequence according to distributed timestamps; S1-2. Structuring the collected data into a three-dimensional time-service-metric tensor. The time dimension is layered and constructed according to sliding windows, with a minimum granularity of 5 minutes. A node hardware resource baseline model is introduced to standardize the service peak resource utilization to form a comparable indicator vector. ; S1-3. When building a static microservice path map, filter the instantaneous peak traffic caused by scheduling behavior, including the CI / CD change deployment process, and use a three-point median filter smoothing mechanism to eliminate occasional data jumps; the final topology sparsity is Sparse interaction graph of .

3. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 1 is characterized in that: During the bottleneck path identification phase, the cross-period structural inconsistency rate indicator is introduced to dynamically identify path nodes in the same service chain that exhibit significant behavioral differences in multiple time windows. The indicator is defined as: ; Among them, the is the calling frequency at time t, is the mean, It is the reconstruction factor of the AZ to which the service belongs; the indicator is used as the bottleneck path pruning weight to serve the path sub-tensor priority selection strategy.

4. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 2 is characterized in that: Said S1-2 further comprises the following steps: S1-2-1. Introduce the deployment node delay disturbance fitting function in the standardization processing stage, and build a response delay-load fitting model for each node in the form of , is the actual response time curve of the node; S1-2-2. Perform first-order derivative trend analysis on the fitting curve and screen out nodes with first-order derivative fluctuations greater than 3 times the median offset. These nodes are treated as unstable node sets and subsequently subjected to separate attention weight noise reduction in service health assessment.

5. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 1 is characterized in that: The S3 dynamic service health assessment model includes the following steps: S3-1. Use a structurally heterogeneous graph model to organize service components into a heterogeneous node set based on deployment location, resource affinity, and call logic. Node attributes include the standard deviation of instantaneous CPU usage, memory utilization fluctuation frequency, and the sliding average of interface call success rate. S3-2. Graph modeling uses a re-parameterized structured attention mechanism, where each attention path is controlled by a dual factor, the node resource behavior angle difference and the historical delay covariance of neighbor paths ; S3-3. The output service embedding vector enters the state anomaly regressor, and the output score is a continuous distribution [0, 1]. Nodes with an anomaly greater than 0.8 will be included in the perturbation simulation critical path and fed back to the strategy model for feature adjustment.

6. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 5 is characterized in that: The S3-2 further includes the following steps: S3-2-1. Introducing the node scheduling history state transition probability in the attention score calculation , forming a dynamic attention distribution with constrained normalization to avoid hot nodes from interfering with the expansion of the neighborhood; S3-2-2. Apply cluster-driven compression after embedding update to cluster and compress nodes with a Cosine distance less than 0.1 in the embedding space to reduce feature redundancy of high-frequency adjacent nodes, and output the compressed representation for downstream prediction.

7. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 1 is characterized in that: The S6 further comprises the following steps: S6-1. Construct a betweenness centrality-time evolution graph for each component and select components with betweenness value slope greater than a threshold of 0.3 and a standard deviation less than 0.15 as long-term high centrality path nodes and as migration targets; S6-2. Migration strategy considering deployment cost function ,in is the startup delay, D is the cross-domain data migration amount, and a multi-target arrangement model is established in the target node set according to the cost gradient; S6-3. The reorganization structure verifies that the call volume of the migration path is restored to more than 90% of the original value within 10 minutes by calling the consistency verifier. If it fails to meet the standard, it will automatically roll back and enter the backup reconstruction path.

8. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 4 is characterized in that: Said S1-2-2 further comprises the following steps: S1-2-2-1. Build a high-order Fourier coefficient change rate sequence based on the historical 3-hour delay sequence, extract the main harmonics of the delay variation, and establish a symmetrical envelope coefficient model to determine whether the jump is caused by system-level periodic fluctuations and local faults; S1-2-2-2. Remove the periodic fluctuation component from the original sequence and apply a second-order sliding minimum variance filter to the residual term to obtain the actual business behavior deviation trend for microservice evolution modeling.

9. The method for efficient management and control of distributed microservice architecture service components based on intelligent learning models according to claim 5 is characterized in that: The S3-3 further includes the following steps: S3-3-1. Construct an abnormal service node aggregation pool, in which members of the abnormal service node aggregation pool are subjected to subgraph cutting in a multi-scale recursive neighborhood aggregation manner to form a spin group structure; S3-3-2. With the spin group as the center, a spherical adjacent boundary is constructed and the dynamic growth coefficient of the anomaly radius is calculated. When the anomaly propagation speed exceeds a fixed threshold of 0.8× the average neighborhood response rate, the local path freezing mechanism is triggered to block the anomaly diffusion chain.

Images