Fault propagation path determination method and device, storage medium and electronic equipment

Abstract

The invention discloses a fault propagation path determination method and device, a storage medium and electronic equipment, and relates to the field of Internet of Things. The method comprises the following steps: acquiring a target data set; a service calling graph is constructed according to the target data set, then strong connectivity vectors in the service calling graph are determined, and the service calling graph is used for representing a topological structure and a dependency relationship of service interaction in the distributed system; the strong connectivity vector is used for representing a connectivity component of which the bidirectional calling frequency is greater than a preset frequency and / or the data interaction strength is greater than a preset strength in the service calling graph; performing neighborhood feature aggregation on the service call graph to obtain a time sequence feature matrix; and determining a fault propagation path of the distributed system according to the strong connectivity vector and the time sequence characteristic matrix. According to the method and the device, the technical problem of poor fault propagation path prediction accuracy caused by lack of quantification of dependency relationship and topological association between services in fault propagation prediction in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of Internet of Things, in particular to a fault propagation path determination method and device, a storage medium and an electronic device. BACKGROUND

[0002] At present, a distributed system has become a basic framework for building complex applications and services, which is composed of multiple components or services that cooperate with each other, communicate and exchange data through a network, and jointly complete complex functions and tasks. The dynamic interaction and dependency relationship between services form a multi-dimensional and high-density network structure, and the complexity of this structure has a decisive influence on the overall stability and fault propagation characteristics of the system.

[0003] Traditional fault diagnosis techniques and fault propagation prediction models often take an isolated perspective when dealing with faults in distributed systems, i.e., mainly focusing on fault detection and recovery of a single service or component, without fully considering the deep dependency relationship between services within the system and the complex characteristics of the topology structure. This limitation leads to the problem that when a fault occurs, the system may not be able to quickly and accurately identify the affected service chain, resulting in low accuracy of fault propagation path prediction.

[0004] In view of the above problems, no effective solutions have been proposed so far. SUMMARY

[0005] The embodiments of the present application provide a fault propagation path determination method, device, storage medium and electronic device to at least solve the technical problem in the prior art that fault propagation prediction lacks quantification of the dependency relationship between services and topology association, resulting in poor accuracy of fault propagation path prediction.

[0006] According to an aspect of an embodiment of the present application, a fault propagation path determination method is provided, comprising: obtaining a target data set, wherein the target data set at least includes service logic data and resource physical data of a distributed system; constructing a service call graph according to the target data set, and then determining a strongly connected vector in the service call graph, wherein the service call graph is used to represent the topology structure and dependency relationship of service interaction in the distributed system, and the strongly connected vector is used to represent a connected component in the service call graph with bidirectional call frequency greater than a preset frequency and / or data interaction intensity greater than a preset intensity; performing neighborhood feature aggregation on the service call graph to obtain a time sequence feature matrix; and determining a fault propagation path of the distributed system according to the strongly connected vector and the time sequence feature matrix.

[0007] Optionally, the service call graph is constructed according to the target data set, including: constructing a vertex set according to all service nodes in the target data set; constructing a directed edge set according to the call relationship between each two associated service nodes in the target data set, wherein the time sequence weight of each directed edge in the directed edge set is determined by the call frequency and call delay between the two service nodes associated with the edge; and generating the service call graph according to the vertex set and the directed edge set.

[0008] Optionally, the strongly connected vector in the service call graph is determined, including: performing a traversal operation on the service call graph, wherein the traversal operation is used to push the accessed service nodes into a stack while maintaining a search sequence number and a low link value; the search sequence number is a unique identifier assigned in order to each accessed service node, and is used to track the progress of the traversal search and the access order of the service nodes; the low link value represents a search sequence number lower limit value set for each service node, and the search sequence number lower limit values corresponding to different service nodes are the same or different; in the case where the low link value of the i th service node is equal to the search sequence number of the i th service node, backtracking is performed from the stack until the i th service node, and a node set on the backtracking path is defined as a strongly connected vector, wherein i is an integer greater than or equal to 1.

[0009] Optionally, the neighborhood feature of the service call graph is aggregated to obtain a time sequence feature matrix, including: dimensionally processing the service nodes in the service call graph into node feature vectors, wherein the node feature vectors include at least one of the processor utilization rate and the service error rate; for each service node, traversing the first-order neighbor nodes of the service node, and then determining a target function from a plurality of aggregation functions according to the business scenario demand corresponding to the service node; fusing the node feature vectors of the service node and the first-order neighbor nodes of the service node according to the target function to obtain the neighborhood dynamic feature of the service node, wherein the neighborhood dynamic feature at least includes the node call frequency, the delay fluctuation information of the node call, and the error rate trend information of the node call; and generating the time sequence feature matrix according to the neighborhood dynamic feature of each service node.

[0010] Optionally, the fault propagation path of the distributed system is determined according to the strongly connected vector and the time sequence feature matrix, including: constructing a state space model according to the strongly connected vector, wherein the state space model is used to simulate and predict the propagation path and time length of the fault between different service nodes; taking the node call frequency, the delay fluctuation information of the node call, and the error rate trend information of the node call in the time sequence feature matrix as state transition weights; and determining the fault propagation path according to the state transition weights and the state space model.

[0011] Optionally, the determining the fault propagation path according to the state transition weight and the state space model comprises: determining a transition state of each two strongly connected vectors at the target time based on the state space model and the state transition weight, wherein the transition state of each two strongly connected vectors at the target time comprises at least one propagation path when the fault transfers from one of the each two strongly connected vectors to another of the each two strongly connected vectors, and a time consumption corresponding to each propagation path; and determining the fault propagation path of each two strongly connected vectors at the target time according to the transition state of each two strongly connected vectors at the target time.

[0012] Optionally, in the process of determining the transition state of each two strongly connected vectors at the target time based on the state space model and the state transition weight, the method further comprises:

[0013] determining the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time based on the state space model, wherein the ith strongly connected vector and the jth strongly connected vector are any two different strongly connected vectors, the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time represents at least one propagation path when the fault transfers from the ith strongly connected vector to the jth strongly connected vector and a time consumption corresponding to the propagation path; obtaining a neighbor component set corresponding to the jth strongly connected vector; and determining the transition state of the ith strongly connected vector and the jth strongly connected vector at the target time according to the state transition weight of each component in the neighbor component set and the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time, wherein the first time is a time preceding the target time.

[0014] Optionally, after determining the fault propagation path of each two strongly connected vectors at the target time according to the transition state of each two strongly connected vectors at the target time, the method further comprises: determining a fault propagation path set according to the fault propagation path of each two strongly connected vectors at the target time; and determining a target fault propagation path from the fault propagation path set according to a target condition, wherein the target condition comprises a constraint condition of a shortest time length of the fault propagation and a constraint condition of a least number of service nodes associated with the fault.

[0015] According to a further aspect of the embodiments of the present application, a device for determining a failure propagation path is also provided, comprising: an obtaining unit configured to obtain a target data set, wherein the target data set comprises at least service logic data and resource physical data of a distributed system; a first processing unit configured to construct a service call graph according to the target data set, and then determine a strongly connected vector in the service call graph, wherein the service call graph is used to represent a topology structure and a dependency relationship of service interaction in the distributed system, and the strongly connected vector is used to represent a connected component in the service call graph in which a bidirectional call frequency is greater than a preset frequency and / or a data interaction intensity is greater than a preset intensity; a second processing unit configured to perform neighborhood feature aggregation on the service call graph to obtain a time sequence feature matrix; and a third processing unit configured to determine a failure propagation path of the distributed system according to the strongly connected vector and the time sequence feature matrix.

[0016] According to a further aspect of the embodiments of the present application, a computer readable storage medium is also provided, which stores a computer program, wherein when the computer program is run, the computer readable storage medium causes a device in which the computer readable storage medium is located to perform the method for determining a failure propagation path described above.

[0017] According to a further aspect of the embodiments of the present application, an electronic device is also provided, which comprises one or more processors and a memory, and the memory is configured to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors perform the method for determining a failure propagation path described above.

[0018] According to a further aspect of the embodiments of the present application, a computer program product is also provided, which comprises a computer program or instructions, and the computer program or instructions implement the method for determining a failure propagation path described above when executed by a processor.

[0019] In the present application, the method for determining a failure propagation path first obtains a target data set, wherein the target data set comprises at least service logic data and resource physical data of a distributed system; constructs a service call graph according to the target data set, and then determines a strongly connected vector in the service call graph, wherein the service call graph is used to represent a topology structure and a dependency relationship of service interaction in the distributed system, and the strongly connected vector is used to represent a connected component in the service call graph in which a bidirectional call frequency is greater than a preset frequency and / or a data interaction intensity is greater than a preset intensity; performs neighborhood feature aggregation on the service call graph to obtain a time sequence feature matrix; and determines a failure propagation path of the distributed system according to the strongly connected vector and the time sequence feature matrix.

[0020] In the embodiment of the present application, the method for determining the fault propagation path is adopted, the target data set containing service logic data and resource physical data is collected, the service call graph quantifying the service call frequency and data interaction intensity is constructed, and the strongly connected component with bidirectional call frequency and data interaction intensity exceeding the preset threshold is identified. Then, the neighborhood feature aggregation technology is used, the time sequence information of the service call graph is combined, the time sequence feature matrix is generated, and the understanding of the service dynamic interaction is further deepened. Finally, with the help of the strongly connected vector and the time sequence feature matrix, the fault propagation in the distributed system can be accurately predicted, so as to realize the timely isolation and efficient recovery of the fault, enhance the system stability and service continuity, and further solve the technical problems that the fault propagation prediction lacks the quantification of the inter-service dependency relationship and topological association in the prior art, and the fault propagation path prediction accuracy is poor. BRIEF DESCRIPTION OF DRAWINGS

[0021] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0022] Figure 1 is a flow chart of an optional method for determining a fault propagation path according to an embodiment of the present application;

[0023] Figure 2 is a general structure diagram of an optional method for determining a fault propagation path according to an embodiment of the present application;

[0024] Figure 3 is a schematic diagram of an optional device for determining a fault propagation path according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] In order to enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should fall within the scope of protection of the present application.

[0026] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and in the above drawings are used to distinguish similar objects, and do not necessarily have to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0027] According to an embodiment of the present application, an embodiment of a method for determining a fault propagation path is provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a set of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0028] According to an embodiment of the present application, a fault propagation path determination system (hereinafter referred to as system) can be used as an execution subject of the method for determining a fault propagation path according to an embodiment of the present application, wherein the fault propagation path determination system can be a software system or a combination of software and hardware embedded system. Of course, the execution subject of the method according to an embodiment of the present application can also be other forms of execution subject, such as device, equipment, etc. Those skilled in the art should know that the specific form of the execution subject of the present application is not particularly limited.

[0029] Figure 1 is a flowchart of an optional method for determining a fault propagation path according to an embodiment of the present application, as shown in Figure 1 The method comprises the following steps:

[0030] Step S101, obtaining a target data set, wherein the target data set at least includes service logic data and resource physical data of a distributed system.

[0031] Optionally, the target data set refers to collecting real-time running data and historical interaction records from each service in the distributed system, including service logic data and resource physical data. The service logic data covers the execution flow, interaction mode and other information of the service, while the resource physical data involves resource consumption, network delay and other physical layer data during service running.

[0032] Optionally, the system continuously collects key indicators such as call records, data transmission volume, response time, etc. between services by deploying monitoring agents or utilizing log analysis tools, forming a data set that comprehensively reflects the running state of the system. This process may involve data cleaning, formatting, and real-time analysis techniques, contributing to the accuracy and usability of the data.

[0033] In step S102, a service call graph is constructed according to the target data set, and a strongly connected vector in the service call graph is determined, wherein the service call graph is used to represent the topology and dependency of service interaction in the distributed system, and the strongly connected vector is used to represent the connected components in the service call graph with bidirectional call frequency greater than a preset frequency and / or data interaction intensity greater than a preset intensity.

[0034] Optionally, the nodes in the service call graph represent services, and the edges represent the call relationship between services, and the weight of the edge represents the call frequency and data interaction intensity.

[0035] Optionally, the strongly connected vector is used to mark the connected components in the service call graph with bidirectional call frequency and data interaction intensity higher than the preset threshold.

[0036] Optionally, based on the target data set, the system uses graph theory algorithms to construct the service call graph, and assigns call frequency and data interaction intensity weights to each edge. Then, algorithms such as depth-first search are used to determine the strongly connected components in the graph, and the information is encoded into a strongly connected vector for subsequent analysis.

[0037] In step S103, the neighborhood feature aggregation of the service call graph is performed to obtain a time series feature matrix.

[0038] Optionally, neighborhood feature aggregation refers to extracting the adjacent service characteristics of each service node from the service call graph, including call order, frequency variation, response time fluctuation, etc. The time series feature matrix is a matrix obtained by aggregating neighborhood features, which is used to represent the dynamic interaction characteristics between services over time.

[0039] Optionally, for each node in the service call graph, the system performs neighborhood feature aggregation to collect the temporal features of the surrounding services, such as historical trends of call frequency, fluctuations of response time, etc., and integrates the characteristics into a time series feature vector. Then, all the time series feature vectors of the services are organized into a time series feature matrix to reflect the dynamic characteristics of the internal service interaction of the system. In addition, the generation of the time series feature matrix not only captures the static dependencies between services, but also reveals the dynamic interaction patterns over time, providing time series information for fault propagation prediction and enhancing the accuracy and timeliness of the prediction.

[0040] In step S104, the fault propagation path of the distributed system is determined according to the strongly connected vector and the time series feature matrix.

[0041] Optionally, the fault propagation path refers to the path along which a fault propagates from a service that has failed to other services in the system through the dependency chain in the service call graph.

[0042] Optionally, in combination with the strongly connected vector and the timing feature matrix, the system uses matrix operations and graph algorithms to analyze and identify the fault source and its associated paths with other services in the strongly connected component. By analyzing the associated paths in depth, the system can predict the direction and scope of the fault propagation. By comprehensively analyzing the service cluster relationship in the strongly connected vector and the dynamic interaction data in the timing feature matrix, the system can provide more accurate fault propagation path prediction, thereby accelerating the fault isolation and recovery process and improving the stability and business continuity of the system.

[0043] In an optional embodiment, constructing a service call graph based on a target data set includes: constructing a vertex set based on all service nodes in the target data set; constructing a directed edge set based on the call relationship between each pair of associated service nodes in the target data set, wherein the timing weight of each directed edge in the directed edge set is determined by the call frequency and call latency between the two service nodes associated with the edge; and generating a service call graph based on the vertex set and the directed edge set.

[0044] Optionally, the vertex set refers to a set of all service nodes in the distributed system, and each service node is represented as a vertex in the call graph, which helps the service call graph to cover all service nodes.

[0045] Optionally, by analyzing the service logic data in the target data set, all service instances participating in inter-service calls are identified and added as independent vertices to the vertex set. This process may involve data filtering and deduplication, which helps to reduce redundancy in the graph representation.

[0046] Optionally, the directed edge set is a set of edges representing the call relationship between services in the service call graph, with the direction indicating the direction of the call and the weight determined by the call frequency and call latency. By constructing the directed edge set, the system can accurately depict the call relationship between services in the service call graph, as well as the strength and dynamic characteristics of the call relationship. The weight setting can reflect important information about the call frequency and latency.

[0047] Optionally, based on the call events recorded in the target data set, a directed edge is created between each pair of associated service nodes, with the direction of the edge pointing to the target service node of the call. At the same time, the weight of each directed edge is calculated, specifically by counting the call frequency between service nodes within a certain period of time and combining the average latency of each call to determine the weight value. A directed edge with a higher weight implies frequent and stable call relationship between services, which may become the main channel for fault propagation.

[0048] Optionally, the service call graph is a graph representation consisting of vertices and directed edges, where vertices correspond to service nodes, and directed edges and edge weights represent the calling relationship between services and its strength. Generating the service call graph can visually present the topology and dependency of service interaction in the distributed system, facilitating subsequent analysis and fault propagation prediction. By combining the vertex set and the directed edge set, it can be seen which service nodes are high-frequency callers, which nodes are critical call objects, and the strength and directionality of the calling relationship.

[0049] Optionally, the system uses an algorithmic framework in graph theory, such as an adjacency matrix or an adjacency list, to combine the vertex set and the directed edge set, constructing a mathematical model of the service call graph. Then, according to this model, a visualization tool is used to draw the service call graph, where the position and size of each vertex may be adjusted according to its position in the network and the size of the weight, in order to show the dependency and network structure between services.

[0050] In an optional embodiment, determining a strongly connected component in the service call graph comprises: performing a traversal operation on the service call graph, wherein the traversal operation is used to push the accessed service nodes into a stack while maintaining a search order number and a low link value; the search order number is a unique identifier assigned in order to each accessed service node, and is used to track the progress of the traversal search and the access order of the service nodes; the low link value represents a lower limit value of the search order number set for each service node, and the lower limit values of the search order number corresponding to different service nodes are the same or different; in the case where the low link value of the ith service node is equal to the search order number of the ith service node, backtracking from the stack until the ith service node, and defining the set of nodes on the backtracking path as a strongly connected component, wherein i is an integer greater than or equal to 1.

[0051] Optionally, the traversal operation is a process of accessing each service node in the service call graph in turn, and the stack is a first-in-last-out data structure used to save the service nodes on the access path.

[0052] Optionally, the strongly connected component is a set of all service nodes in the service call graph that can be reached bidirectionally through a series of directed edges, which usually means that there is a high calling frequency and strong dependency between nodes.

[0053] Optionally, starting from an optional unvisited service node in the service call graph, it is marked and pushed into the stack. Then, all service nodes directly connected to it are recursively accessed while maintaining the search order number and the low link value.

[0054] Optionally, when first accessing a service node, an incremental search sequence number is assigned to it, and the low-link value is initialized to the same value. The low-link value is the lower limit of the search sequence number of each service node, which can be updated according to the search sequence number of the descendant nodes of the service node during traversal. Then, all adjacent nodes are traversed. For each adjacent node, if it has not been accessed, continue to traverse in depth, and recursively update the low-link value of the adjacent node. If the adjacent node has been accessed but not popped from the stack, update the low-link value of the current node to the smaller value of the low-link value of the adjacent node.

[0055] Optionally, during traversal, if it is detected that the low-link value of a certain service node is equal to its search sequence number, it means that the current node and its previous node form a closed strongly connected component. At this time, backtracking from the stack until the current node, and define the set of nodes on the backtracking path as a strongly connected vector. This process is repeated until all service nodes in the service call graph have been accessed.

[0056] In an optional embodiment, the neighborhood feature aggregation of the service call graph is performed to obtain a time sequence feature matrix, including: dimensionally processing the service nodes in the service call graph into node feature vectors, wherein the node feature vectors include at least one feature information of processor utilization rate and service error rate; for each service node, traversing the first-order neighbor nodes of the service node, and then determining a target function from a plurality of aggregation functions according to the business scenario demand corresponding to the service node; fusing the node feature vectors of the service node and the first-order neighbor nodes of the service node according to the target function to obtain the neighborhood dynamic feature of the service node, wherein the neighborhood dynamic feature at least includes node call frequency, node call delay fluctuation information and node call error rate trend information; generating a time sequence feature matrix according to the neighborhood dynamic feature of each service node.

[0057] Optionally, the node feature vector is a process of converting the key attributes of the service node into a multi-dimensional numerical vector, which usually includes feature information such as processor utilization rate and service error rate.

[0058] Optionally, the system converts the processor utilization rate and service error rate and other index information of each service node in the service call graph into a feature vector. The dimension and specific feature information of the feature vector can be flexibly adjusted according to the business scenario demand, which helps the selected indicators to have an impact on the prediction of the fault propagation path.

[0059] Optionally, the aggregation function is used to process the feature information of the service node and its neighbor nodes, and common functions include average value, weighted average, maximum value, etc.

[0060] Optionally, for each service node in the service call graph, first traverse its first-order neighbor nodes (i.e., other service nodes directly connected to the service node in the service call graph). Then, according to the requirements of the business scenario (such as real-time transaction system, file storage service, or data processing platform) to which the service node belongs, select a target function from a plurality of preset aggregation functions. For example, for a real-time transaction system, it may pay more attention to the delay fluctuation of node calls, so an aggregation function with high correlation with delay fluctuation is selected; while for a file storage service, it may pay more attention to the stability of the service, so an aggregation function that can reflect the error rate trend of the service is selected.

[0061] Optionally, the neighborhood dynamic feature is the result obtained by fusing the feature vectors of the service node and its first-order neighbor nodes through the target aggregation function, which contains key indicators such as node call frequency, delay fluctuation information, and error rate trend.

[0062] Optionally, for the selected service node, the target aggregation function is applied to fuse the node feature vector of the service node itself and the feature vectors of its first-order neighbor nodes to obtain the neighborhood dynamic feature of the service node, thereby reflecting the dynamic behavior of the local network, including the frequency change of service calls, the trend of delay fluctuation, and the fluctuation of error rate.

[0063] Optionally, the time series feature matrix is a matrix containing neighborhood dynamic features with service nodes as rows and timestamps as columns, used to represent the trend of dynamic features of each service node in the service call graph over time.

[0064] Optionally, the neighborhood dynamic feature information of all service nodes at different time points is collected and arranged in matrix form. Each row represents a service node, and each column corresponds to a timestamp. Each element in the matrix represents the neighborhood dynamic feature value of a particular service node at a particular time point. Through this matrix, the evolution trend of the running state of the service node over time can be observed.

[0065] In an optional embodiment, determining a fault propagation path of a distributed system according to the strongly connected vector and the time series feature matrix comprises: constructing a state space model according to the strongly connected vector, wherein the state space model is used to simulate and predict the propagation path and duration of faults between different service nodes; taking the node call frequency, the delay fluctuation information of node calls, and the error rate trend information of node calls in the time series feature matrix as state transition weights; and determining the fault propagation path according to the state transition weights and the state space model.

[0066] Optionally, the state space model is used to describe and predict the transition between different states of the system. In this application, the state of the model corresponds to the health status of the service node, while the state transition represents how the failure propagates from one service node to another. By simulating the dependency and interaction characteristics between service nodes, the state space model can more accurately predict the direction and possible impact range of failure propagation.

[0067] Optionally, the system determines the initial state and possible state transition paths of the model according to the service node set in the strongly connected vector. The state of each service node can be represented by a series of variables, such as whether it is running normally, the level of failure, etc. Then, the initial state of the model is set as all service nodes being in normal operation state. By defining state transition rules in the state space model, the interaction and dependency between service nodes in different states are simulated, including how the failure triggers state transition and the conditions and paths of failure propagation.

[0068] Optionally, the state transition weight refers to the probability or cost of state transition from one service node to another in the state space model, which is determined by the node call frequency, delay fluctuation information and error rate trend information in the time series feature matrix. The state transition weight quantifies the possibility and impact of service node state transition.

[0069] Optionally, the node call frequency in the time series feature matrix is considered as an indicator of frequent interaction between service nodes, and high frequency of calls means higher probability of state transition; the delay fluctuation information reflects the stability of the call, and large fluctuation may increase the risk of failure propagation; the error rate trend information directly shows the quality and stability of the service node operation. Based on the above information, each state transition path in the state space model is assigned a weight, and the size of the weight can reflect the possibility and duration of failure propagation from the current service node to the adjacent service node.

[0070] Optionally, in the state space model, graph algorithms (such as Dijkstra algorithm) are used in combination with state transition weights to find the shortest path or lowest cost path from the simulated failure source node to other service nodes, i.e. the most likely propagation path of the failure. At the same time, the time required for the failure to propagate on the path is calculated by the state transition weight, giving the time range of the failure impact.

[0071] In an optional embodiment, the method for determining the fault propagation path based on the state transition weight and the state space model comprises: determining the transition state of each pair of strongly connected vectors at the target time based on the state space model and the state transition weight, wherein the transition state of each pair of strongly connected vectors at the target time comprises at least one propagation path when the fault transfers from one of the pair of strongly connected vectors to the other one of the pair of strongly connected vectors, and the time consumption of each propagation path; and determining the fault propagation path of each pair of strongly connected vectors at the target time based on the transition state of each pair of strongly connected vectors at the target time.

[0072] Optionally, the state space model is used to calculate, for each pair of strongly connected vectors in the service call graph, the possible path and the corresponding time consumption of the fault transfer from one strongly connected vector to another strongly connected vector at the target time based on the state transition weight. This calculation can use the shortest path algorithm in graph theory, taking the weight as the path cost, to find the fault propagation path with the shortest time consumption.

[0073] Optionally, for any two strongly connected vectors in the service call graph, the time consumption of each propagation path obtained by calculation is selected, and the path with the shortest time consumption or the highest probability is selected as the fault propagation path at the target time. Wherein, the selection may need to consider additional factors, such as the real-time state of the service node and the system load, thereby helping to improve the accuracy of the prediction.

[0074] In an optional embodiment, in the process of determining the transition state of each pair of strongly connected vectors at the target time based on the state space model and the state transition weight, the method further comprises: determining the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time based on the state space model, wherein the ith strongly connected vector and the jth strongly connected vector are any two different strongly connected vectors, and the transition state at the first time represents at least one propagation path when the fault transfers from the ith strongly connected vector to the jth strongly connected vector and the time consumption of the propagation path; obtaining a neighbor component set corresponding to the jth strongly connected vector; and determining the transition state of the ith strongly connected vector and the jth strongly connected vector at the target time based on the state transition weight of each component in the neighbor component set and the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time, wherein the first time is the last time before the target time.

[0075] Optionally, the state space model is used to calculate the possible propagation path and the time consumption from the ith strongly connected vector to the jth strongly connected vector in combination with the system state and historical data at the current time (i.e. the first time). Wherein, the other strongly connected vectors having a direct or indirect connection relationship with the jth strongly connected vector are determined, which helps to analyze the propagation characteristics of the fault between these strongly connected vectors.

[0076] Optionally, based on the service call graph, all service nodes directly connected to the jth strongly connected vector and other strongly connected vectors indirectly connected through first-order neighbor relationship are identified to form a "neighbor component set".

[0077] Optionally, the system uses information at the first time point, including the fault propagation path and time consumption at that time point, and the state transition weight of each component in the neighbor component set. Then, combined with the prediction ability of the state space model, a suitable time series prediction algorithm (such as ARIMA, state transition matrix multiplication, etc.) is used to predict the shortest path and shortest time consumption of the fault transfer from the ith strongly connected vector to the jth strongly connected vector at the target time point, which helps to improve the timeliness and accuracy of fault propagation analysis.

[0078] In an optional embodiment, after determining the fault propagation path of each two strongly connected vectors at the target time point according to the transition state of each two strongly connected vectors at the target time point, the method further comprises: determining a fault propagation path set according to the fault propagation path of each two strongly connected vectors at the target time point; and determining a target fault propagation path from the fault propagation path set according to a target condition, wherein the target condition includes a constraint condition of the shortest time length of fault propagation and a constraint condition of the least number of service nodes related to fault.

[0079] Optionally, the fault propagation path set refers to the set of all possible fault propagation paths at a certain target time point, each path originating from a strongly connected vector and eventually affecting another strongly connected vector.

[0080] Optionally, all calculated fault propagation paths and their corresponding time consumption information are sorted into a matrix or list form of fault propagation path set according to the combination of source strongly connected vector and target strongly connected vector. Each path and its related time consumption should be clearly marked with its origin and destination to facilitate subsequent analysis and screening.

[0081] Optionally, the target condition is a specific condition or standard that needs to be met when selecting a fault propagation path, including but not limited to a constraint condition of the shortest time length of fault propagation and a constraint condition of the least number of service nodes related to fault. By setting the target condition, the fault propagation path that best meets the system operation and maintenance requirements is selected, which helps to quickly respond and minimize the impact of faults.

[0082] Optionally, the system uses the shortest path algorithm in graph theory or other applicable optimization algorithms to find the path with the shortest time consumption from the fault propagation path set, i.e. the target fault propagation path. This helps to quickly locate the affected area when a fault occurs and promptly initiate recovery measures.

[0083] Optionally, finally, the number of service nodes associated with each path is calculated, and the path with the fewest associated service nodes is selected as the target fault propagation path. This helps reduce the scope of fault propagation, protects unaffected service nodes, and thus maintains the overall service continuity of the system.

[0084] Figure 2 This diagram illustrates the overall structure of an optional method for determining a fault propagation path according to an embodiment of this application. Figure 2 As shown, the system collects data from multiple dimensions—both service logic and resource physical layers—through a multi-source data acquisition and preprocessing module. This data is preprocessed to form a structured node-edge-temporal feature dataset, such as constructing a cyclical call chain from account to payment to settlement, and a unidirectional call path from payment to risk control. Subsequently, a directed temporal graph is constructed using a strong connected component identification module. For example, it identifies and marks strong connected component 1, which forms the core transaction loop of account, payment, and settlement, and strong connected component 2, which forms an independent subsystem of risk control, achieving in-depth analysis of the system topology. Next, a neighborhood feature aggregation module initializes node embedding in the data and captures potential temporal dependencies between services by aggregating the temporal features of neighboring nodes. For example, the phenomenon of "sudden increase in payment, then settlement delay" will be marked as a high-risk path, providing important clues for subsequent fault prediction. Next, the dynamic programming-driven fault propagation prediction module employs a three-stage algorithm collaboration: state space reconstruction, temporal weight injection, and Pareto optimal path calculation, to predict the propagation path and duration of the fault. It outputs the predicted path (e.g., from b to c and then to a), the predicted duration (e.g., 0.05 seconds), and a fault heatmap, visually demonstrating the potential impact range and intensity of the fault. Following this, the composite scenario generation and verification module comprehensively considers core cluster cascading and cross-cluster linkage scenarios, verifying whether the prediction accuracy meets expectations (e.g., verification error less than 10%). The output and quantitative evaluation module generates fault heatmaps, propagation indicators, and optimization suggestions. For example, for the scenario of "a fault in strongly connected component 1 cascading to strongly connected component 2," it proposes specific fault impact assessments and preventative measures. Finally, the entire process achieves iterative iteration through continuous monitoring and closed-loop optimization. Based on the preliminary evaluation results, such as suggestions to increase cross-cluster redundancy, the system automatically triggers system topology optimization and resource scheduling adjustments, then returns to monitoring status, continuously monitoring the optimized system performance. This closed-loop mechanism, from data acquisition to fault prediction, and then to composite scenario verification and system optimization, helps to improve the comprehensiveness and timeliness of distributed system fault management, and enhances system stability and operational efficiency.

[0085] The descriptions of each module in the overall structure are as follows:

[0086] Multi-source data collection and preprocessing: interface service logs (call relationship, QPS, delay), server configuration (CPU / memory), deployment relationship (service-machine room mapping), and generate "node-edge-time sequence feature" dataset after deduplication and completion, laying the foundation for subsequent analysis.

[0087] Strongly connected component identification: build a graph with services as nodes and call relationships as directed edges, then identify strongly connected components SCC (such as core transaction cluster SCC ), label core closed loop and edge service, replace manual topology analysis.

[0088] Neighborhood feature aggregation: aggregate node neighborhood features (call frequency, delay fluctuation) through graph neural networks, learn "call influence" and "fault sensitivity", and output node embedding matrix with time sequence dependence to capture high-risk propagation paths.

[0089] Dynamic programming driven fault propagation prediction: reconstruct the state space with SCC as "super node", inject time sequence weight, and use dynamic programming state transition equation (dp[SCC i ][SCC j ][t]=min()+w()))to predict cluster-level propagation path and duration, output Pareto optimal path set and fault heat map (e.g. red / orange / yellow risk identification), where dp represents the solution array of dynamic programming, [SCC i ] and [SCC j ] represent the i-th and j-th strongly connected components, respectively, and the min() function calculates the shortest time to propagate the fault from SCC i to SCC j at time t, and the w() function represents the weight.

[0090] Composite scenario generation and verification: based on the predicted path, generate "core cluster cascade + cross-cluster linkage" scenarios (such as SC fault triggers SC ), dynamically inject faults (service downtime, network isolation) to verify prediction accuracy (error <10%).

[0091] Output and quantitative evaluation: generate fault heat map (visualize propagation range), quantitative indicators (propagation duration, number of affected nodes), and output optimization suggestions (such as adding cross-cluster redundancy between SC and SC ).

[0092] Continuous monitoring and closed-loop optimization: optimize system topology / resources based on evaluation recommendations, return to monitoring state, form a "prediction-verification-optimization" closed loop, and improve financial system fault response efficiency.

[0093] Optionally, the core logic of the system is to realize the quantitative prediction of fault propagation through the "topology decomposition, time series feature analysis, dynamic programming prediction, and heat map visualization" closed loop. The core process is:

[0094] 1. Data collection and preprocessing: Collect service logic data (call logs, time series features) and resource physical data (server configuration, deployment relationship), and generate "node-edge-time series feature" structured data set after deduplication and completion.

[0095] Example: Collect the "account (A)-payment (B, QPS 500)-clearing (C, delay 20ms)-account (A)" loop call chain and "payment-risk control (D, error rate 0.1%)" one-way call of the bank core system, and associate A, B, and C deployed in computer room A and D deployed in computer room B.

[0096] 2. Strongly connected component identification:

[0097] Construct a directed time series graph, with service nodes (A, B, C, D) as vertex set {V}, call relationship as directed edge set {E} (such as A-B represents service A calling service B), and assign call frequency and delay as time series weights to each edge.

[0098] Identify strongly connected components: traverse the graph structure by depth-first search, while maintaining two key variables: search order number and low link value. When the depth-first search recursively visits a node, push it onto the stack and mark it as visited; if the low link value of a node is equal to its search order number, backtrack from the top of the stack to the node to form a strongly connected component.

[0099] For example: identify SC ={A, B, C} as the core transaction closed loop, which has a bidirectional call path (A-B-C-A) inside, while SC ={D} as an independent risk control service node only receiving one-way calls.

[0100] Mark the core cluster: the system then outputs a logical topology report, marking SC as "core transaction strongly connected cluster" (highest test priority), and annotating the call frequency and delay threshold between each strongly connected component (such as SC -SC call error rate exceeding 5% triggers an alarm). This process helps to achieve linear time complexity (O(|V|+|E|)) for strongly connected component identification through stack backtracking mechanism, which is better than manual topology analysis method.

[0101] 3. Neighborhood feature aggregation:

[0102] Node embedding initialization: Map service nodes into a low-dimensional vector space, each node vector contains CPU utilization, error rate and other key attribute indicators.

[0103] Neighborhood feature aggregation: The system uses a multi-level aggregation strategy: for each target node (such as node B), first traverse its first-order neighbors (A, C, D), then select Mean / Max / LSTM aggregation function according to business scenario requirements for feature fusion. Mean aggregation function is suitable for smoothing statistical features such as call frequency, Max function is used to capture abnormal feature peaks such as delay fluctuations, and LSTM is used to model time step dependent patterns.

[0104] Through aggregation operations, key dynamic features can be extracted, for example, in call chain analysis, the system automatically marks the path as a high-risk propagation path when it discovers abnormal patterns such as "B-C delay surge (20ms-100ms)" and "C-A error rate rise (0.1%-5%)".

[0105] Output time series feature matrix: Contains two core dimensions: "call influence" dimension quantifies the strength and stability of the call relationship between nodes, and "fault sensitivity" dimension reflects the vulnerability of nodes to upstream and downstream faults. These features will be used as key input weights for dynamic programming algorithms.

[0106] 4. Dynamic programming driven fault propagation prediction:

[0107] The present application breaks through the limitations of traditional single algorithm and creates a "strongly connected component + time series features + dynamic programming" three-algorithm collaborative framework. It includes the following steps:

[0108] (1) State space reconstruction: Take the identified strongly connected component as the basic unit of dynamic programming (such as SCC As a "super node"), upgrade the fault propagation from "single-point-single-point" to "cluster-cluster", solving the "local perspective" defect of traditional solutions;

[0109] (2) Time series weight injection: The output node embedding (including delay fluctuations, error rate trends) is used as the "dynamic weight" of dynamic programming state transition, replacing the fixed propagation time, making the prediction fit the "time dependence" of real fault propagation;

[0110] (3) Path-time length joint optimization: The dynamic state transition equation combines "shortest propagation time" and "maximum impact range" double objectives, outputs a set of Pareto optimal paths, providing multiple strategy options for emergency resource scheduling.

[0111] 4.1 State definition and transition equation:

[0112] State definition: dp[SCC i ][SCCj ] represents the shortest time for the failure to arrive from SCCi to SCCj at time t, path[SCC i ] records the propagation path of the failure from SCCi to SCCj (e.g., SCC1-SCC2-SCC3).

[0113] The transition equation is formula (1):

[0114] ) (1)

[0115] where N(SCC j ) is the neighbor component of SCC j , w(SCC k , SCC j ) is the inter-component propagation weight (including delay, error rate features) learned by GraphSAGE, t represents the current time step, and t-1 represents the previous time step.

[0116] 4.2 Example:

[0117] Scenario: The trigger node is the payment service (B) in SCC , and the predicted failure propagates to SCC (clearing C, account A) and SCC (risk control D).

[0118] Step 1, identify SCC ={A,B,C}, SCC ={D}, output B-C weight w=0.02 seconds (delay 20 ms), C-A weight w=0.03 seconds, and B-D weight w=0.01 seconds.

[0119] Step 2, dynamic programming state transition calculation:

[0120] dp[SC ][SC ][B-C]=0.02 seconds (path B-C), path=SC internal closed loop;

[0121] dp[SC ][SC ][B-C-A]=0.02+0.03=0.05 seconds (path B-C-A);

[0122] dp[SC ][SC ][B-D]=0.01 seconds (cross-cluster path B-D).

[0123] Step 3, output result: the primary impact node is D (e.g., 0.01 seconds), the secondary impact node within the core cluster is A (0.05 seconds), and a heat map is generated with the SC label is red (high risk), SC is orange (medium risk).

[0124] 4.3 Joint optimization of path and duration: dual-objective optimization is achieved through dynamic programming state transition equations, with "shortest propagation duration" and "maximum impact range" as parallel optimization targets.

[0125] In specific modeling, the state dp[SCCi][SCCj][t] is defined to represent the shortest duration (in seconds) of fault propagation from strongly connected component SCCi to SCCj, while path[SCCi][SCCj] is maintained to record the propagation path (e.g., SCC1-SCC2-SCC3). Node embedding is integrated as a dynamic weight in the transition equation, replacing the traditional fixed propagation duration parameter, so that the prediction result can reflect the timing dependency in the real system.

[0126] When balancing dual objectives, the Pareto optimal solution set method is used to generate a non-dominated solution set through dynamic programming algorithm. The "shortest propagation duration" target is achieved by minimizing the cumulative time in the state transition equation, while the "maximum impact range" target is reflected by maximizing the number of SCCs covered by the path. The final output of the Pareto frontier contains multiple optimization strategies, such as:

[0127] Strategy A: propagation path SCC1-SCC2-SCC3 (duration 12 seconds, impact 3 SCCs).

[0128] Strategy B: propagation path SCC1-SCC4-SCC5-SCC6 (duration 18 seconds, impact 4 SCCs).

[0129] This joint optimization mechanism provides multi-dimensional decision-making basis for emergency resource scheduling, allowing for trade-offs between propagation speed and impact range according to actual scene requirements.

[0130] 5. Compound scenario generation and verification: based on the predicted path, generate "core cluster cascade + cross-cluster linkage" scenarios (e.g., "SC Fault-triggering SC cascade"), dynamically inject faults (service downtime, network isolation) to verify prediction accuracy (error <10% passes).

[0131] Figure 3is a schematic diagram of an optional fault propagation path determination apparatus according to an embodiment of the present application. According to another aspect of the present application, there is also provided a fault propagation path determination apparatus, comprising: an obtaining unit 301, a first processing unit 302, a second processing unit 303, and a third processing unit 304.

[0132] The obtaining unit 301 is configured to obtain a target data set, wherein the target data set comprises at least service logic data and resource physical data of a distributed system; the first processing unit 302 is configured to construct a service call graph according to the target data set, and then determine a strongly connected vector in the service call graph, wherein the service call graph is used to represent a topological structure and a dependency relationship of service interaction in the distributed system, and the strongly connected vector is used to represent a connected component in the service call graph in which bidirectional call frequency is greater than a preset frequency and / or data interaction intensity is greater than a preset intensity; the second processing unit 303 is configured to perform neighborhood feature aggregation on the service call graph to obtain a time sequence feature matrix; and the third processing unit 304 is configured to determine a fault propagation path of the distributed system according to the strongly connected vector and the time sequence feature matrix. The above-mentioned serial numbers of the embodiments of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0133] Optionally, the first processing unit 302 comprises: a set construction subunit, configured to construct a vertex set according to all service nodes in the target data set; a directed edge set is constructed according to a call relationship between each two associated service nodes in the target data set, wherein a time sequence weight of each directed edge in the directed edge set is determined by a call frequency and a call delay between the two service nodes associated with the edge; and a generation subunit, configured to generate the service call graph according to the vertex set and the directed edge set.

[0134] Optionally, the first processing unit 302 further comprises: a traversal subunit, configured to perform a traversal operation on the service call graph, wherein the traversal operation is used to press a service node being visited into a stack while maintaining a search serial number and a low link value; the search serial number is a unique identifier assigned to each visited service node in order, and is used to track the progress of the traversal search and the visiting order of the service nodes; the low link value represents a lower limit value of the search serial number set for each service node, and the lower limit values of the search serial numbers corresponding to different service nodes are the same or different; and a backtracking subunit, configured to backtrack from the stack until an i-th service node in a case where it is detected that a low link value of the i-th service node is equal to a search serial number of the i-th service node, and define a node set on a backtracking path as a strongly connected vector, wherein i is an integer greater than or equal to 1.

[0135] Optionally, the second processing unit 303 comprises: a dimension reduction processing subunit, configured to perform dimension reduction processing on the service nodes in the service call graph to obtain node feature vectors, wherein the node feature vectors comprise at least one of processor utilization and service error rate; a neighborhood dynamic feature determination subunit, configured to, for each service node, traverse first-order neighbor nodes of the service node, and then determine a target function from a plurality of aggregation functions according to a business scenario requirement corresponding to the service node; and perform fusion of the node feature vectors of the service node and the first-order neighbor nodes of the service node according to the target function to obtain neighborhood dynamic features of the service node, wherein the neighborhood dynamic features comprise at least node call frequency, node call delay fluctuation information, and node call error rate trend information; and a time sequence feature matrix determination subunit, configured to generate a time sequence feature matrix according to the neighborhood dynamic features of each service node.

[0136] Optionally, the third processing unit 304 comprises: a model construction subunit, configured to construct a state space model according to the strongly connected vectors, wherein the state space model is used to simulate and predict a propagation path and a time length of a fault between different service nodes; a weight determination subunit, configured to take the node call frequency, the node call delay fluctuation information, and the node call error rate trend information in the time sequence feature matrix as state transition weights; and a path determination subunit, configured to determine a fault propagation path according to the state transition weights and the state space model.

[0137] Optionally, the path determination subunit comprises: a transition state determination module, configured to determine a transition state of each two strongly connected vectors at a target time based on the state space model and the state transition weights, wherein the transition state of each two strongly connected vectors at the target time comprises at least one propagation path when a fault transfers from one of the two strongly connected vectors to the other, and a time consumption corresponding to each propagation path; and a propagation path determination module, configured to determine a fault propagation path of each two strongly connected vectors at the target time according to the transition state of the two strongly connected vectors at the target time.

[0138] Optionally, the transition state determining module comprises: a first determining submodule configured to determine a transition state of the ith strongly connected vector and the jth strongly connected vector at a first time based on the state space model, wherein the ith strongly connected vector and the jth strongly connected vector are any two different strongly connected vectors, and the transition state at the first time represents at least one propagation path and a corresponding time consumption of the propagation path when the fault transitions from the ith strongly connected vector to the jth strongly connected vector; an obtaining submodule configured to obtain a neighbor component set corresponding to the jth strongly connected vector; and a second determining submodule configured to determine a transition state of the ith strongly connected vector and the jth strongly connected vector at a target time according to a state transition weight of each component in the neighbor component set and the transition state of the ith strongly connected vector and the jth strongly connected vector at the first time, wherein the first time is a time preceding the target time.

[0139] Optionally, the propagation path determining module comprises: a third determining submodule configured to determine a fault propagation path set according to the fault propagation path of each two strongly connected vectors at the target time; and a fourth determining submodule configured to determine a target fault propagation path from the fault propagation path set according to a target condition, wherein the target condition comprises a constraint condition of a shortest time length of fault propagation and a constraint condition of a least number of service nodes associated with the fault.

[0140] According to another aspect of the embodiments of the present application, a computer readable storage medium is also provided, and the computer readable storage medium stores a computer program. When the computer program runs, the computer readable storage medium makes a device where the computer readable storage medium is located execute the above-mentioned method for determining a fault propagation path.

[0141] According to another aspect of the embodiments of the present application, an electronic device is also provided, and the electronic device comprises one or more processors and a memory. The memory is configured to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors execute the above-mentioned method for determining a fault propagation path.

[0142] According to another aspect of the embodiments of the present application, a computer program product is also provided, and the computer program product comprises a computer program or instructions. When the computer program or instructions are executed by a processor, the computer program or instructions implement the above-mentioned method for determining a fault propagation path.

[0143] In the above-mentioned embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0144] In several embodiments provided in the present application, it should be understood that the disclosed technology can be implemented by other ways. Among them, the above-described device embodiments are only schematic, for example, the division of the units can be a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, units or modules, and can be electrical or other forms.

[0145] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e. can be located in one place or can be distributed to multiple units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0146] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0147] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and various program codes that can be stored in the medium.

[0148] The above is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be considered as the protection scope of the present application.

Claims

1. A method for determining a fault propagation path, characterized in that, include: Obtain the target dataset, wherein the target dataset includes at least the service logic data and resource physical data of the distributed system; A service call graph is constructed based on the target dataset, and then strong connectivity vectors in the service call graph are determined. The service call graph is used to characterize the topology and dependencies of service interactions in the distributed system, and the strong connectivity vectors are used to characterize the connected components in the service call graph whose bidirectional call frequency is greater than a preset frequency and / or whose data interaction intensity is greater than a preset intensity. Neighborhood feature aggregation is performed on the service call graph to obtain a time-series feature matrix; The fault propagation path of the distributed system is determined based on the strongly connected vector and the temporal feature matrix.

2. The method for determining the fault propagation path according to claim 1, characterized in that, Constructing a service call graph based on the target dataset includes: A vertex set is constructed based on all service nodes in the target dataset; a directed edge set is constructed based on the call relationship between every two associated service nodes in the target dataset, wherein the temporal weight of each directed edge in the directed edge set is determined by the call frequency and call delay between the two service nodes associated with the edge; The service call graph is generated based on the set of vertices and the set of directed edges.

3. The method for determining the fault propagation path according to claim 2, characterized in that, Determining the strongly connected vectors in the service call graph includes: The service call graph is traversed, wherein the traversal operation is used to push accessed service nodes onto a stack while maintaining search sequence numbers and low link values; the search sequence number serves as a unique identifier assigned sequentially to each accessed service node, and is used to track the progress of the traversal search and the access order of service nodes; the low link value represents the lower limit value of the search sequence number set for each service node, and different service nodes may have the same or different lower limit values ​​of the search sequence number. If it is detected that the low link value of the i-th service node is equal to the search sequence number of the i-th service node, backtrack from the stack to the i-th service node, and define the set of nodes on the backtracking path as a strongly connected vector, where i is an integer greater than or equal to 1.

4. The method for determining the fault propagation path according to claim 1, characterized in that, The service call graph is subjected to neighborhood feature aggregation to obtain a time-series feature matrix, including: The service nodes in the service call graph are reduced in dimensionality to node feature vectors, wherein the node feature vectors include at least one feature information from processor utilization and service error rate; For each service node, the first-order neighbor nodes of the service node are traversed, and then the target function is determined from multiple aggregation functions according to the business scenario requirements corresponding to the service node. Based on the target function, the node feature vectors of the service node and the first-order neighbor nodes of the service node are fused to obtain the neighborhood dynamic features of the service node. The neighborhood dynamic features include at least the node call frequency, the latency fluctuation information of the node call, and the error rate trend information of the node call. The time-series feature matrix is ​​generated based on the neighborhood dynamic features of each service node.

5. The method for determining the fault propagation path according to claim 1, characterized in that, Determining the fault propagation path of the distributed system based on the strongly connected vector and the temporal feature matrix includes: A state-space model is constructed based on the strongly connected vectors, wherein the state-space model is used to simulate and predict the propagation path and duration of faults between different service nodes; The node call frequency, node call delay fluctuation information, and node call error rate trend information in the time-series feature matrix are used as state transition weights. The fault propagation path is determined based on the state transition weights and the state space model.

6. The method for determining the fault propagation path according to claim 5, characterized in that, Determining the fault propagation path based on the state transition weights and the state space model includes: Based on the state space model and the state transition weights, the transition state of each pair of strongly connected vectors at the target time is determined. The transition state of each pair of strongly connected vectors at the target time includes: at least one propagation path when the fault is transferred from one of the two strongly connected vectors to the other, and the time consumption corresponding to each propagation path. Based on the transition state of each pair of strongly connected vectors at the target time, the fault propagation path of each pair of strongly connected vectors at the target time is determined.

7. The method for determining the fault propagation path according to claim 6, characterized in that, In determining the transition state of every two strongly connected vectors at the target time based on the state-space model and the state transition weights, the method further includes: Based on the state space model, the transition states of the i-th strongly connected vector and the j-th strongly connected vector at the first moment are determined, wherein the i-th strongly connected vector and the j-th strongly connected vector are any two different strongly connected vectors, and the transition state at the first moment represents at least one propagation path when the fault is transferred from the i-th strongly connected vector to the j-th strongly connected vector and the time consumption corresponding to the propagation path. Obtain the set of neighbor components corresponding to the j-th strongly connected vector; Based on the state transition weight of each component in the neighbor component set and the transition states of the i-th strongly connected vector and the j-th strongly connected vector at the first time, the transition states of the i-th strongly connected vector and the j-th strongly connected vector at the target time are determined, wherein the first time is the time before the target time.

8. The method for determining the fault propagation path according to claim 6, characterized in that, After determining the fault propagation path of each pair of strongly connected vectors at the target time based on their transition states at the target time, the method further includes: Based on the fault propagation path of each pair of strongly connected vectors at the target time, determine the set of fault propagation paths; The target fault propagation path is determined from the set of fault propagation paths based on the target conditions, wherein the target conditions include: the constraint of the shortest fault propagation time and the constraint of the fewest service nodes associated with the fault.

9. A device for determining a fault propagation path, characterized in that, include: An acquisition unit is used to acquire a target dataset, wherein the target dataset includes at least the service logical data and resource physical data of the distributed system; The first processing unit is configured to construct a service call graph based on the target dataset, and then determine the strongly connected vectors in the service call graph. The service call graph is used to characterize the topology and dependencies of service interactions in the distributed system, and the strongly connected vectors are used to characterize the connected components in the service call graph whose bidirectional call frequency is greater than a preset frequency and / or whose data interaction intensity is greater than a preset intensity. The second processing unit is used to perform neighborhood feature aggregation on the service call graph to obtain a time-series feature matrix. The third processing unit is used to determine the fault propagation path of the distributed system based on the strong connectivity vector and the temporal feature matrix.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed, the device in which the computer-readable storage medium is located performs the fault propagation path determination method according to any one of claims 1 to 8.

11. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to perform the method for determining a fault propagation path as described in any one of claims 1 to 8.

12. A computer program product, characterized in that, It includes a computer program or instructions that, when executed by a processor, implement the method for determining the fault propagation path according to any one of claims 1 to 8.

Images