CN-121980230-A - Micro-service-oriented risk perception multitasking graph time sequence prediction method

CN121980230ACN 121980230 ACN121980230 ACN 121980230ACN-121980230-A

Abstract

The invention discloses a micro-service-oriented risk perception multitask graph time sequence prediction method which is applied to a target cluster and used for supporting operation of at least one micro service, determining time sequence feature tensors corresponding to at least one micro service according to data to be processed of at least one target device associated with each micro service in a plurality of evaluation dimensions within preset time length, inputting the time sequence feature tensors into a pre-trained risk evaluation model to obtain prediction results of the micro service in the plurality of evaluation dimensions, performing risk evaluation on the prediction results of the micro service in the plurality of evaluation dimensions according to a service level target threshold value in each evaluation dimension to obtain risk evaluation results of the micro service, and determining target risk evaluation information according to the risk evaluation results of at least one micro service. The method improves the accuracy and global consistency of the prediction result, ensures that the corresponding risk assessment information meets the actual operation and maintenance requirements, and provides reliable data support for subsequent scheduling and resource allocation.

Inventors

ZHAO TAIYIN
LU GUOMING
LUO GUANGCHUN
ZHAN SIYU
ZHANG YUNXIAO
DU JIAYI

Assignees

电子科技大学

Dates

Publication Date: 20260505
Application Date: 20260303

Claims (10)

1. The micro-service-oriented risk perception multitasking graph time sequence prediction method is characterized by being applied to a target cluster, wherein the target cluster comprises a plurality of target devices and is used for supporting at least one micro-service operation, and the method comprises the following steps: determining time sequence feature tensors corresponding to at least one micro service according to data to be processed of at least one target device associated with each micro service in a plurality of evaluation dimensions within a preset duration, wherein the plurality of evaluation dimensions at least comprise a central processing unit use dimension, a memory use dimension, a service request delay dimension and a service request failure dimension; Inputting the time sequence characteristic tensor into a pre-trained risk assessment model to obtain a prediction result of the micro-service under a plurality of assessment dimensions, wherein the risk assessment model at least comprises a graph time sequence encoder and a plurality of prediction task output modules, the graph time sequence encoder at least comprises a target graph convolution network and a time sequence encoding module, the target graph convolution network is obtained after a preset graph convolution network is adjusted based on a mixed adjacency matrix, and the mixed adjacency matrix is related to call information between the micro-service; Performing risk assessment on the prediction results of the micro-service in a plurality of assessment dimensions according to the service level target threshold value in each assessment dimension to obtain risk assessment results of the micro-service; And determining target risk assessment information corresponding to the target cluster according to the risk assessment result of at least one micro service.
2. The method of claim 1, wherein the determining a timing characteristic tensor corresponding to at least one micro-service based on data to be processed for a plurality of evaluation dimensions for a preset duration of at least one target device associated with each micro-service comprises: for a plurality of time slices within the preset duration, acquiring data to be processed in a plurality of evaluation dimensions within the time slices of at least one target device associated with each micro-service; and carrying out standardization processing on the data to be processed of at least one micro service according to the time slice number and the service number of the micro service so as to determine a three-dimensional tensor, and determining the three-dimensional tensor as a time sequence characteristic tensor.
3. The method of claim 1, wherein the hybrid adjacency matrix is determined by: determining call information among the micro services according to at least one service log information of the target cluster; constructing a micro-service call topological graph according to the call information, and determining a normalized adjacency matrix corresponding to the micro-service call topological graph; Determining at least one quality of service feature for each micro-service based on historical data for each micro-service in a plurality of evaluation dimensions for at least one time slice within a historical time period; Performing inter-service similarity evaluation on at least one service quality characteristic of at least one micro service to obtain an inter-service similarity matrix and an adaptive adjacency matrix corresponding to the inter-service similarity matrix; and carrying out weighted fusion processing on the normalized adjacent matrix and the self-adaptive adjacent matrix to obtain a mixed adjacent matrix.
4. The method of claim 1, wherein inputting the temporal feature tensor into a pre-trained risk assessment model to obtain the predicted results of the microservice in multiple assessment dimensions comprises: performing multi-layer graph volume aggregation processing on the time sequence feature tensor based on a target graph convolution network of the graph time sequence encoder to obtain a time sequence service feature sequence corresponding to the micro service; processing the time sequence service feature sequence based on a time sequence coding module of the graph time sequence coder to obtain a time sequence embedded vector corresponding to the micro service; and respectively predicting the time sequence embedded vector of the micro service based on a plurality of prediction task output modules, and determining the prediction result of the micro service under each evaluation dimension.
5. The method according to claim 4, wherein the target graph convolution network includes a plurality of graph convolution layers, the target graph convolution network based on the graph timing encoder performs multi-layer graph-volume aggregation processing on the timing feature tensor to obtain the timing service feature sequence corresponding to the micro-service, including: Determining a tensor to be processed of each microservice under each time slice according to the time sequence characteristic tensor; for the tensor to be processed of each micro-service under at least one time slice, performing linear transformation mapping processing on the tensor to be processed under each time slice to obtain initial service node characteristics; sequentially processing the initial service node characteristics according to the mixed adjacent matrix, the learnable weight matrix and the nonlinear activation function in the graph convolution layer to obtain output service node characteristics; taking the output service node characteristic as an initial service node characteristic of a next graph convolution layer, and repeating the processing of the initial service node characteristic by the graph convolution layer until the output service node characteristic of a final graph convolution layer is obtained; Taking the output service node characteristics of the convolution layer of the last graph as service characteristics to be fused; and splicing the service features to be fused of each micro service under at least one time slice according to the time slices to obtain a time sequence service feature sequence corresponding to each micro service.
6. The method of claim 4, wherein the timing encoding module at least comprises a gating loop unit and/or a one-dimensional convolutional network, the timing encoding module based on the graph timing encoder processes the timing service feature sequence to obtain the timing embedded vector corresponding to the microservice, and the method comprises the following steps: And capturing time dependency relationship of the time sequence service characteristic sequence based on a gating circulation unit and/or a one-dimensional convolution network of the time sequence coding module so as to obtain a time sequence embedded vector of each micro service.
7. The method of claim 4, wherein the plurality of predicted task output modules includes at least a first output module corresponding to the central processor usage dimension, a second output module corresponding to the memory usage dimension, a third output module corresponding to the service request delay dimension, a fourth output module corresponding to the service request failure dimension, The predicting task output module respectively predicts the time sequence embedded vector of the micro-service based on a plurality of the predicting task output modules, and determines the predicted result of the micro-service under each evaluation dimension, including: For the first output module, the second output module and the third output module, taking the historical data of the last time slice in the historical duration under the corresponding evaluation dimension as reference data; processing the time sequence embedded vector based on a residual prediction function under the corresponding evaluation dimension to obtain an output result under the corresponding dimension; and adding the output result and the reference data to obtain a prediction result under the corresponding evaluation dimension.
8. The method of claim 7, wherein the method further comprises: for the fourth output module, processing the time sequence embedded vector based on a residual prediction function corresponding to the service request failure dimension to obtain a to-be-processed result; and processing the to-be-processed result based on a nonlinear activation function to obtain a predicted result under the service request failure dimension.
9. The method according to claim 1, wherein the method further comprises: Acquiring a plurality of training samples, wherein the training samples comprise historical data of each micro service in a plurality of evaluation dimensions in a historical time length and theoretical results of each micro service in a plurality of evaluation dimensions, and the historical time length is consistent with the time length of the preset time length; for a plurality of training samples, determining a sample feature tensor corresponding to the training sample; Inputting the sample characteristic tensor into a risk assessment model to be trained to obtain an actual result of each micro-service under a plurality of assessment dimensions; Carrying out loss processing on the actual result and the theoretical result according to a target loss function to obtain a first loss value, wherein the target loss function is determined based on a multi-task regression loss sub-function, a classification loss sub-function and a first weight coefficient corresponding to the sub-function; Adjusting the first loss value according to the risk weight coefficient corresponding to the training sample to obtain a target loss value; and adjusting model parameters of the risk assessment model to be trained based on the target loss value to obtain a trained risk assessment model.
10. The method of claim 9, wherein the risk weighting factor is determined by: Determining comprehensive risk assessment attributes corresponding to each training sample according to a service level target threshold of the micro service in each assessment dimension; and mapping the comprehensive risk assessment attribute to obtain a risk weight loss coefficient corresponding to the training sample.

Description

Micro-service-oriented risk perception multitasking graph time sequence prediction method Technical Field The invention relates to the technical field of data processing, in particular to a micro-service-oriented risk perception multitasking graph time sequence prediction method. Background With the application of cloud native technology and micro-service architecture, complex business systems are typically composed of tens or even hundreds of micro-services that form a directed call graph through remote calls, running entirely on top of shared cluster resources. In order to ensure service stability and user experience, resource requirements and quality indexes in a future period of time are generally predicted based on a set service level Objective (SERVICE LEVEL Objective, SLO), so as to perform resource scheduling or capacity expansion processing according to the prediction result. At present, the traditional prediction method is often used for training a model aiming at a single evaluation index or respectively training a model based on a plurality of evaluation indexes, and the problems that the evaluation indexes are isolated and data information among the evaluation indexes cannot be shared exist, so that a prediction result lacks global consistency, and the actual operation and maintenance scene is difficult to adapt. In addition, the traditional prediction method takes global average indexes such as mean square error or average absolute error and the like as an optimization target when a model is trained, and the overall risk assessment error can be reduced, but under a peak load scene or a tail delay scene, the prediction result based on the traditional prediction method is difficult to meet the actual operation and maintenance requirements. Disclosure of Invention The invention provides a micro-service-oriented risk perception multitask graph time sequence prediction method, which improves the accuracy and global consistency of a prediction result, ensures that corresponding risk assessment information meets the actual operation and maintenance requirements and provides reliable data support for subsequent scheduling and resource allocation. According to an aspect of the present invention, there is provided a micro-service oriented risk awareness multitasking graph timing prediction method applied to a target cluster, where the target cluster includes a plurality of target devices for supporting at least one micro-service operation, the method including: Determining a time sequence feature tensor corresponding to at least one micro service according to data to be processed of at least one target device associated with each micro service in a plurality of evaluation dimensions within a preset duration, wherein the plurality of evaluation dimensions at least comprise a central processing unit use dimension, a memory use dimension, a service request delay dimension and a service request failure dimension; Inputting a time sequence characteristic tensor into a pre-trained risk assessment model to obtain a prediction result of the micro-service under a plurality of assessment dimensions, wherein the risk assessment model at least comprises a graph time sequence encoder and a plurality of prediction task output modules, the graph time sequence encoder at least comprises a target graph convolution network and a time sequence encoding module, the target graph convolution network is obtained after a preset graph convolution network is adjusted based on a mixed adjacency matrix, and call information between the mixed adjacency matrix and the micro-service is related; Performing risk assessment on the prediction results of the micro-service in a plurality of assessment dimensions according to the service level target threshold value in each assessment dimension to obtain a risk assessment result of the micro-service; And determining target risk assessment information corresponding to the target cluster according to the risk assessment result of the at least one micro service. According to another aspect of the present invention, there is provided a micro-service oriented risk aware multitasking graph timing prediction apparatus for use in a target cluster, the target cluster including a plurality of target devices for supporting at least one micro-service operation, the apparatus comprising: the tensor determining module is used for determining a time sequence characteristic tensor corresponding to at least one micro-service according to data to be processed of at least one target device associated with each micro-service in a plurality of evaluation dimensions within a preset duration, wherein the plurality of evaluation dimensions at least comprise a central processing unit use dimension, a memory use dimension, a service request delay dimension and a service request failure dimension; The prediction result determining module is used for inputting the time sequence characteristic tensor into a pre-trained