A performance index prediction method, device and medium for a pod

Abstract

This invention provides a method, device, and medium for predicting Pod performance metrics, relating to the field of data processing technology. In the method, a set A of multidimensional performance metric datasets is obtained based on a preset historical time period. Based on A, an initial Pod count list C corresponding to a preset future time period T2, and a time-series large model corresponding to A, an initial performance metric prediction dataset D corresponding to A is obtained. D is input into a self-healing system to obtain a predicted Pod count list E corresponding to T2. When C and E are not completely consistent, D is updated based on C and E, and E is used as C. Then, D is input into the self-healing system again for the next round of judgment. When C and E are completely consistent, the target performance metric prediction dataset corresponding to each Pod currently serving the target service is obtained based on D. This method considers the impact of dynamic changes in the number of Pods on the load and performance metrics of individual Pods, which helps improve the accuracy of the prediction results.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, device and medium for predicting the performance indicators of a Pod. Background Technology

[0002] In large-scale distributed online service systems (such as airline business systems), backend microservices are typically deployed dynamically in containerized form (such as Kubernetes Pods). To ensure service quality and improve resource utilization, it is necessary to predict the performance metrics of the corresponding Pods for the service over a future period. Most existing performance metric prediction methods are based on historical time-series data of a single metric and use neural network models for univariate prediction.

[0003] However, the above method also has the following technical problems:

[0004] In actual operation, a service corresponds to at least one Pod, and the number of corresponding Pods will be dynamically adjusted according to the load. When the number of Pods changes, external requests or tasks will be redistributed to more (or fewer) instances, which will significantly change the actual load borne by a single Pod and thus directly affect its various performance indicators. However, when predicting performance indicators based on the above method, it is usually assumed that the number of Pods corresponding to the service remains unchanged, without taking into account the impact of dynamic changes in the number of Pods on the performance indicators of Pods. Therefore, the indicator data predicted by the above method cannot accurately reflect the real load distribution, and the accuracy of the prediction results is low. Summary of the Invention

[0005] To address the aforementioned technical problems, the technical solution adopted by this invention is as follows:

[0006] According to a first aspect of the present invention, a method for predicting performance metrics for a Pod is provided, the method comprising the following steps:

[0007] S1. Obtain the set A = {A1, A2, ..., A...} of the multidimensional performance index dataset. i A n};A i A is a multi-dimensional performance metric dataset for the i-th Pod corresponding to the target service within a predefined historical time period T1, where 1 ≤ i ≤ n, and n is the number of Pods corresponding to the target service within T1; i It includes a time series data list corresponding to each preset performance indicator. The time series data list includes the continuous sampled values ​​of the corresponding preset performance indicator within T1. The end time of T1 is the current time point, and the duration is the first preset duration t1.

[0008] S2. Based on A, the initial Pod count list C corresponding to the preset future time period T2, and the time series model corresponding to A, obtain the initial performance metric prediction dataset D corresponding to A; the starting time of T2 is the current time, and the duration is the second preset duration t2; C includes the initial Pod count corresponding to each preset time point in T2; D includes the initial metric prediction value list corresponding to each preset performance metric, and the initial metric prediction value list includes the initial metric prediction value corresponding to each preset time point in T2; t2 < t1; the initial Pod count is consistent with the current Pod count corresponding to the target service.

[0009] S3. Input D into the self-healing system to obtain the predicted Pod count list E corresponding to T2; E includes the predicted Pod count corresponding to each preset time point in T2.

[0010] S4. When C and E are completely consistent, obtain the target performance metric prediction dataset for each Pod corresponding to the target service based on D; when C and E are not completely consistent, update D based on C and E, take E as C, and proceed to step S3.

[0011] According to a second aspect of the present invention, a non-transitory computer-readable storage medium is provided, wherein a computer program is stored in the storage medium, and the computer program is loaded and executed by a processor to implement the aforementioned method.

[0012] According to a third aspect of the present invention, an electronic device is provided, comprising: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned method.

[0013] The present invention has at least the following beneficial effects:

[0014] This invention provides a method, device, and medium for predicting Pod performance metrics. In the method, a set A of multidimensional performance metric datasets is obtained based on a preset historical time period. Based on A, an initial Pod count list C corresponding to a preset future time period T2, and a time-series large model corresponding to A, an initial performance metric prediction dataset D corresponding to A is obtained. D is input into a self-healing system to obtain a predicted Pod count list E corresponding to T2. When C and E are not completely consistent, D is updated based on C and E, and E is used as C. Then, D is input into the self-healing system again for the next round of judgment. When C and E are completely consistent, the target performance metric prediction dataset corresponding to each Pod currently serving the target service is obtained based on D. As can be seen, in the process of multidimensional performance index prediction, this invention uses a list of initial Pod counts corresponding to a preset future time period as input, combines it with a time-series large model to generate an initial performance index prediction dataset, and inputs the initial performance index prediction dataset into a self-healing system, which then generates a predicted Pod count list. Further, when the predicted Pod count list and the initial Pod count list are not completely consistent, the initial performance index prediction dataset and the initial Pod count list are updated, and a predicted Pod count sequence is generated again based on the self-healing system for the next round of judgment. When the predicted Pod count list and the initial Pod count list are completely consistent, the target performance index prediction dataset is obtained based on the current initial performance index prediction dataset. This approach considers the impact of dynamic changes in the number of Pods on the load and performance index of individual Pods, making the prediction results more accurately reflect the real load distribution and improving the accuracy of the prediction results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating a method for predicting performance metrics for Pods, provided as an embodiment of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar tasks and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that comprises a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0019] Embodiments of the present invention provide a method for predicting performance metrics for Pods, the method comprising the following steps, such as... Figure 1 As shown:

[0020] S1. Obtain the set A = {A1, A2, ..., A...} of the multidimensional performance index dataset. i A n};A i A is a multi-dimensional performance metric dataset for the i-th Pod corresponding to the target service within a predefined historical time period T1, where 1 ≤ i ≤ n, and n is the number of Pods corresponding to the target service within T1; i It includes a time-series data list corresponding to each preset performance indicator. The time-series data list includes the continuous sampled values ​​of the corresponding preset performance indicator within T1. The end time of T1 is the current time point, and the duration is the first preset duration t1. Those skilled in the art know that the first preset duration is a duration preset by those skilled in the art according to actual needs, such as 24 hours, which will not be elaborated here.

[0021] Specifically, the sampled values ​​in the time series data list correspond one-to-one with the historical sampling time points in T1, and are arranged in order from early to late according to the corresponding historical sampling time points; the historical sampling time points are determined based on the preset sampling period; the preset sampling period is a sampling period pre-set by those skilled in the art according to actual needs, such as 5 seconds, 10 seconds, 15 seconds, 30 seconds, which will not be elaborated here.

[0022] Specifically, the target service is a service that is predetermined by those skilled in the art based on actual needs, such as flight status management service, passenger notification service, boarding gate / baggage carousel allocation service, and surface support scheduling service, which will not be elaborated here.

[0023] Specifically, the Pod corresponding to the target service is a running and ready Pod instance that is dynamically matched by the target service through its tag selector.

[0024] Specifically, a Pod must deploy at least one container.

[0025] Specifically, the preset performance indicators are performance indicators that are predetermined by those skilled in the art based on actual needs, including but not limited to: CPU utilization, memory usage, garbage collection times, response time, and QPS.

[0026] Specifically, for each Pod corresponding to the target service within T1, if the Pod has not been created or has been terminated at a certain historical sampling time point, the sampling value corresponding to that historical sampling time point in the time series data list corresponding to each preset performance metric of the Pod is set to 0.

[0027] Specifically, the number of Pods corresponding to the target service within T1 can be understood as the total number of different Pod instances that have been associated with the target service within T1. For example, if the target service is associated with Pod1, Pod2, and Pod3 in T1, and then Pod3 terminates and Pod4 is created, then Pod1, Pod2, Pod3, and Pod4 are all Pods corresponding to the target service within T1, and the number of Pods corresponding to the target service within T1 is 4.

[0028] S2. Based on A, the initial Pod count list C corresponding to the preset future time period T2, and the time series model corresponding to A, obtain the initial performance metric prediction dataset D corresponding to A; the starting time of T2 is the current time, and the duration is the second preset duration t2; C includes the initial Pod count corresponding to each preset time point in T2; D includes the initial metric prediction value list corresponding to each preset performance metric, and the initial metric prediction value list includes the initial metric prediction value corresponding to each preset time point in T2; t2 < t1; the initial Pod count is consistent with the current Pod count corresponding to the target service; those skilled in the art know that the second preset duration is a duration less than the first preset duration that is preset by those skilled in the art according to actual needs, for example: 1 hour, which will not be elaborated here.

[0029] Specifically, the j-th preset time point in T2 is earlier than the (j+1)-th preset time point in T2, and the interval between two adjacent preset time points in T2 is consistent with the duration corresponding to the preset sampling period.

[0030] Specifically, after step S1 and before step S2, the following steps are also included:

[0031] S01, based on A1, A2, ..., A i A n Obtain the average correlation score R corresponding to A.

[0032] In one specific embodiment, step S01 includes the following steps:

[0033] S011. Obtain the actual Pod count sequence H = (H1, H2, ..., H1) corresponding to T1. x H p ), H x Let p be the number of Pods corresponding to the target service at the x-th historical sampling time point in T1; 1≤x≤p, where p is the number of historical sampling time points in T1.

[0034] S012, If H1, H2, ..., H x H p If all are the same, then according to the preset splicing order corresponding to the preset performance indicators, A will be... i The feature vectors corresponding to each time series data list in the array are concatenated to form A. i The corresponding data feature vector L i Those skilled in the art will know that any method in the prior art for obtaining the feature vector corresponding to a time series data list is within the protection scope of this invention. For example, using an LSTM encoder to encode the time series data list to obtain the feature vector; this will not be elaborated further here.

[0035] S013, set L1, L2, ..., L i , ..., L n Any two different data feature vectors are used to form a data feature vector pair, and the cosine similarity between the two data feature vectors in the data feature vector pair is used as the vector similarity of the data feature vector pair.

[0036] S014. Take the average vector similarity of all data feature vector pairs as R.

[0037] Through the above steps, when the data in the actual Pod count sequence is completely consistent, it indicates that the number of Pods corresponding to the target service is constant within the preset historical time period, and the load borne by each Pod corresponding to the target service should be similar, and their behavior should be highly similar. At this time, the feature vectors corresponding to the time series data list of each preset performance indicator in the multidimensional performance indicator dataset are concatenated into a data feature vector, and any two different data feature vectors in all data feature vectors are combined into a data feature vector pair. The cosine similarity between the two data feature vectors in the data feature vector pair is used as the vector similarity of the data feature vector pair. The average value of the vector similarity of all data feature vector pairs is used as the average correlation score corresponding to the total set of the multidimensional indicator dataset. This enables the determination of whether Pod behavior is consistent based on the average correlation score, realizing a quantitative evaluation of the consistency of Pod behavior and providing reliable data quality assurance for subsequent high-precision prediction.

[0038] Specifically, after step S014, the following is also included:

[0039] S015. If H1, H2, ..., H x H p If they are not completely identical, then they are based on H1, H2, ..., H x H p Divide T1 into several consecutive time intervals to obtain a list of time intervals T3 corresponding to T1. 31 T 32 ,…,T 3y ,…,T 3q ), T 3y Let y be the y-th time interval corresponding to T1, 1≤y≤q, where q is the number of time intervals corresponding to T1.

[0040] Specifically, T 3y It includes several historical sampling time points, and in T 3y At any two historical sampling points, the number of Pods corresponding to the target service is the same.

[0041] S016, Obtain T 3y The corresponding set of multidimensional performance index datasets U 3y U 3y Including T 3y A multidimensional performance metric dataset for each Pod corresponding to the target service within the system.

[0042] S017, based on U 3y , get U 3y The corresponding average correlation score R 3y .

[0043] Specifically, steps S012-S014 are adopted based on A1, A2, ..., A i A n Obtain R in the same way, based on U 3y For each multidimensional performance metric dataset in the dataset, obtain R... 3y .

[0044] S018, R 31 R 32 , ..., R 3y , ..., R 3q The average value is taken as R.

[0045] Specifically, the higher the average relevance score for A, the higher the consistency of performance behavior among the Pods corresponding to the target service within T1.

[0046] Through the above steps, when the data in the actual Pod count sequence is not completely consistent, it indicates that the number of Pods corresponding to the target service has changed within the preset historical time period. In this case, the preset historical time period is divided into multiple time intervals according to the constant interval of Pod count. For each time interval, the set of multidimensional performance index datasets corresponding to the time interval is obtained, and its corresponding average correlation score is calculated. The average of the average correlation scores corresponding to the set of multidimensional performance index datasets corresponding to all time intervals is used as the average correlation score corresponding to the set of multidimensional performance index datasets for the whole. This allows for the calculation of the average correlation score in scenarios where Pods are dynamically scaled up or down. It avoids similarity distortion caused by the non-constant number of Pods, thereby improving the accuracy of the average correlation score. It ensures that an accurate average correlation score can be obtained regardless of whether the number of Pods changes, so that the consistency of Pod behavior can be judged based on the average correlation score, realizing a quantitative assessment of the consistency of Pod behavior, thus providing reliable data quality assurance for subsequent high-precision prediction.

[0047] S02, when R > R 0 When the time comes, proceed to step S2; R 0 The preset relevance score is set by those skilled in the art according to actual needs, such as 0.7 or 0.8, which will not be elaborated here.

[0048] Specifically, when R < R 0 When an abnormal Pod is detected, an error message is issued to alert the user that an abnormal Pod exists.

[0049] Through the above steps, the average correlation score is obtained for each multidimensional performance indicator dataset in the overall dataset of multidimensional performance indicators. When the average correlation score is greater than the preset correlation score, it indicates that the behavior of each Pod corresponding to the target service is highly consistent within the preset historical time period, the system is in a steady state, and the subsequent prediction process can proceed. Otherwise, it indicates that there are Pods with abnormal behavior among all Pods corresponding to the target service within the preset historical time period, such as deviations in indicators caused by faults, configuration errors, or uneven load. In this case, the subsequent prediction process is not proceeded and an abnormal prompt message is issued. This avoids making predictions based on data including abnormal Pods, thereby improving the input quality and prediction accuracy of the time series large model.

[0050] Specifically, the procedure before step S2 also includes:

[0051] S001, Based on A i And an indicator trend type identification model, which determines A from several preset indicator trend types. i The corresponding target indicator change trend type; where each preset indicator change trend type corresponds to a time series large model.

[0052] Specifically, the output of the indicator change trend type identification model includes a preset indicator change trend type and the confidence level corresponding to that preset indicator change trend type.

[0053] Furthermore, the target indicator change trend type is the preset indicator change trend type in the output of the indicator change trend type identification model.

[0054] Specifically, the preset indicator change trend type is the indicator change trend type preset by those skilled in the art according to actual needs, including but not limited to: continuous rise, periodic fluctuation, recovery after a sudden drop, stability, and step-like growth.

[0055] Specifically, the indicator change trend type identification model is obtained by weakly supervising learning through the first preset model and optimizing it in combination with an active learning strategy. The first preset model is a lightweight supervised learning model that supports probability output, that is, a classifier with small number of parameters, high inference efficiency, and the ability to output category prediction probability to evaluate confidence, such as XGBoost and Random Forest, which will not be elaborated here.

[0056] Specifically, the training of the indicator change trend type identification model is based on a first preset training sample set. The first preset training sample set includes several first preset training samples. Each first preset training sample includes a feature representation and a preset indicator change trend type corresponding to the feature representation, which is manually labeled. The feature representation is obtained by feature engineering through the multi-dimensional performance indicator dataset corresponding to a single Pod within a specific time period. The specific time period is a time period that is predetermined by those skilled in the art based on actual needs, with an end time earlier than the start time of a preset historical time period and a duration consistent with the duration of the preset historical time period. This will not be elaborated further here.

[0057] Furthermore, the training process employs a combination of weak supervision and active learning: in the initial stage, only a small number of feature representations are manually labeled to form the initial first preset training sample set, which is used to train the first preset model; based on the trained first preset model, predictions are made on the feature representations that have not been manually labeled, and feature representations with low classification confidence are selected and manually labeled by experts to obtain their corresponding preset indicator change trend types; the newly labeled feature representations and their corresponding preset indicator change trend types are used as new first training samples and added to the first preset training sample set for retraining the first preset model; through multiple rounds of closed-loop iteration of "prediction-screening-labeling-retraining", an indicator change trend type recognition model is obtained.

[0058] Through the above steps, a model for identifying the trend of indicator changes is trained by combining weakly supervised learning and active learning. This significantly reduces the reliance on large-scale manually labeled data, greatly reducing labeling costs and manpower while ensuring the model's recognition accuracy. At the same time, the model has a small number of parameters and efficient inference, making it suitable for online or near real-time scenarios.

[0059] Specifically, in step S001, for A i Perform feature engineering to obtain A i Corresponding feature representation; A i The corresponding feature representations are input into the indicator change trend type recognition model to obtain A. i The corresponding target indicator change trend type.

[0060] S002, If A1, A2, ..., A i A m If the corresponding target indicator change trend types are all consistent, then the time series model corresponding to the target indicator change trend type is taken as the time series model corresponding to A.

[0061] Through the above steps, for each multidimensional performance indicator dataset, based on the multidimensional performance indicator dataset and the indicator change trend type identification model, the target indicator change trend type corresponding to the multidimensional performance indicator dataset is determined from several preset indicator change trend types. If the target indicator change trend types corresponding to all multidimensional performance indicator datasets are consistent, it indicates that the behavior of each Pod corresponding to the target service is highly consistent within the preset historical time period. The time series model corresponding to the target indicator change trend type is then used as the time series model corresponding to the total set of multidimensional performance indicator datasets for indicator prediction. Since different trend types have significantly different requirements for time series modeling capabilities, using a dedicated time series model that matches the actual change pattern can more accurately capture the dynamic characteristics of the indicators compared to using a single general-purpose time series model, thereby significantly improving prediction accuracy.

[0062] Specifically, if A1, A2, ..., A i A m If the corresponding target indicator change trends are not completely consistent, then A1, A2, ..., A i A mIn the corresponding Pod, the Pod that runs continuously within T1 is used as the benchmark Pod. If the target indicator change trend type corresponding to the multidimensional performance indicator dataset of all benchmark Pods is consistent, then the time series large model corresponding to the target indicator change trend type is used as the time series large model corresponding to A. Otherwise, the time series large model corresponding to the default indicator change trend type is used as the time series large model corresponding to A. The default indicator change trend type is determined by those skilled in the art from several preset indicator change trend types according to actual needs. For example, the preset indicator change trend type "stable" is used as the default indicator change trend type, which will not be elaborated here.

[0063] Through the above steps, if the target metric change trend types corresponding to all multidimensional performance metric datasets are not completely consistent, the Pod that has been running continuously within the preset historical time period will be used as the benchmark Pod. If the target metric change trend types corresponding to the multidimensional performance metric datasets of all benchmark Pods are consistent, the time series model corresponding to the target metric change trend type will be used as the time series model corresponding to A. Otherwise, the time series model corresponding to the default metric change trend type will be used as the time series model corresponding to A. This effectively eliminates the impact of temporary scaling up or down or short-lifetime Pods, while ensuring that subsequent metric predictions can still be performed even when the time series models with strong correlations are uncertain.

[0064] In another specific embodiment, if the target metric change trend types corresponding to the multidimensional performance metric datasets of all benchmark Pods are not completely consistent, an abnormal prompt message is issued. At this time, it may indicate that among all Pods corresponding to the target service within a preset historical time period, there are Pods with abnormal behavior, such as metric deviations caused by faults, configuration errors, or uneven loads, and therefore an abnormal prompt message is issued. This avoids making predictions based on data including abnormal Pods, thereby improving the input quality and prediction accuracy of the time series large model.

[0065] Specifically, step S2 includes the following sub-steps:

[0066] S21. Based on A, obtain the average dataset F; F includes the average data list corresponding to each preset performance index, and the average data list includes the average sampled value corresponding to each historical sampling time point in T1.

[0067] Specifically, when calculating the average sample value in F, for each historical sampling time point in T1, only the sample values ​​corresponding to Pods that have been created and not terminated at that time are averaged. Pods that have not been created or have been terminated are not included in the calculation of the average sample value corresponding to that historical sampling time point.

[0068] S22. Input F and C into the time series model corresponding to A to obtain the output of the time series model and use the output as D.

[0069] Through the above steps, based on the Pods created and not terminated at each historical sampling point in the preset historical time period, the sampled value corresponding to each preset performance indicator is calculated, and then the average dataset is constructed, which effectively avoids the underestimation of load caused by including non-running Pods. Furthermore, the average dataset and the list of initial Pod counts corresponding to the preset future time period are input into the corresponding time series large model, so that the time series large model can simultaneously consider the historical load evolution pattern and the impact of future Pod count changes on the single instance load when making indicator predictions, which is beneficial to the accuracy of the prediction results output by the time series large model.

[0070] Specifically, the time series model corresponding to the preset indicator change trend type is obtained by fine-tuning the second preset model based on the second preset training sample set corresponding to the preset indicator change trend type; the second preset model is a multivariate time series prediction model; those skilled in the art know that any multivariate time series prediction model in the prior art that can perform the corresponding processing is within the protection scope of this invention, such as: Moirai model, Chronos-2 model, TemporalFusion Transformer model, which will not be elaborated here.

[0071] Furthermore, the second preset training sample set corresponding to the preset indicator change trend type includes several second preset training samples; wherein, the second preset training samples are constructed through the following steps:

[0072] Select multiple sampling time periods with a duration of t1 and an end time earlier than the start time of a preset time period.

[0073] For each service within each sampling time period, obtain the total set of the preset multidimensional performance index dataset corresponding to that service within that sampling time period; wherein, the method of obtaining the total set of the preset multidimensional performance index dataset is the same as the method of obtaining A.

[0074] If the target indicator change trend type corresponding to each preset multidimensional indicator dataset in the set of preset multidimensional performance indicator datasets is consistent with the same preset indicator change trend type, then the set of preset multidimensional performance indicator datasets shall be regarded as the set of specific multidimensional performance indicator datasets corresponding to the preset indicator change trend.

[0075] Specifically, A is obtained from S001. i Using the same method as the target indicator change trend type, obtain the target indicator change trend type corresponding to the preset multidimensional indicator dataset.

[0076] For the set of datasets corresponding to each specific multidimensional performance indicator based on the preset indicator change trend:

[0077] Calculate the average dataset corresponding to the total set of a specific multidimensional performance index dataset; wherein, the average dataset corresponding to the total set of the specific multidimensional performance index dataset is obtained in the same way as F in step S21.

[0078] Starting from the end time of the sampling period corresponding to the total set of the specific multidimensional performance index dataset, the time period with a duration of t2 is taken as the future time period corresponding to the average dataset.

[0079] According to the preset sampling period, determine the sampling time point corresponding to the future time period, and record the number of Pods corresponding to the service corresponding to the total set of the specific multidimensional performance index dataset at each sampling time point, so as to obtain the actual Pod number sequence corresponding to the average dataset.

[0080] For each preset performance metric, at each sampling time point, the average value of the corresponding metric value of all Pods corresponding to the service corresponding to the total set of the specific multidimensional performance metric dataset is calculated to obtain the time series data list corresponding to each preset performance metric, and then the actual multidimensional performance metric dataset corresponding to the average dataset is constructed.

[0081] The average dataset, the actual Pod count sequence corresponding to the average dataset, and the actual multidimensional performance metric dataset corresponding to the average dataset are combined into a second preset training sample; each second preset training sample uniquely corresponds to a preset metric change trend type.

[0082] Through the above steps, a dedicated time-series model is customized for each preset indicator trend type. The training process is highly targeted and data consistent: First, by strictly screening historical data, only multi-dimensional performance indicator datasets in which all Pods belong to the same trend type are retained to ensure that the training samples are highly matched with the target trend; Second, when constructing training samples, the same averaging strategy as in the inference stage is used to generate input features, and the actual Pod quantity sequence and the corresponding multi-dimensional indicator mean are recorded simultaneously in future time periods to construct a second preset training sample; This allows the model to learn the dynamic mapping relationship between Pod quantity changes and performance indicators under specific load patterns during the training stage, which not only improves the model's ability to fit various trend scenarios, but also ensures the consistency between training and inference.

[0083] In another specific embodiment, in step S2, A and C are input into the time series large model corresponding to A to obtain the output result of the time series large model and use the output result as D; in this embodiment, the second preset training sample includes the set of specific multidimensional performance index datasets, the actual Pod number sequence corresponding to the set of specific multidimensional performance index datasets, and the actual multidimensional performance index dataset corresponding to the set of specific multidimensional performance index datasets.

[0084] S3. Input D into the self-healing system to obtain the predicted Pod count list E corresponding to T2; E includes the predicted Pod count corresponding to each preset time point in T2.

[0085] Specifically, the self-healing system incorporates multiple types of operation and maintenance rules, including resource constraint rules (such as minimum / maximum replica count limits), anomaly response rules (such as automatic reconstruction of Pods in abnormal states), and elastic scaling rules (scaling-down strategies based on performance metric thresholds). Combined with a data-driven predictive model, it comprehensively analyzes data from various preset performance metrics over a period of time, automatically completes anomaly detection, capacity demand prediction, and elastic scheduling decisions, and finally outputs the planned number of Pods required for the corresponding time period.

[0086] Specifically, in step S3, while obtaining E, count is set to count+1 to update count; count is a pre-designed value and count is initially 0.

[0087] S4. When C and E are completely consistent, obtain the target performance metric prediction dataset for each Pod corresponding to the target service based on D; when C and E are not completely consistent, update D based on C and E, take E as C, and proceed to step S3.

[0088] Through the above steps, in the multidimensional performance indicator prediction process, an initial Pod count list corresponding to a preset future time period is used as input. This list, combined with a time-series large-scale model, generates an initial performance indicator prediction dataset. This dataset is then input into a self-healing system, which generates a predicted Pod count list. Further, when the predicted Pod count list and the initial Pod count list are not completely consistent, both are updated, and a predicted Pod count sequence is generated again based on the self-healing system for the next round of evaluation. When the predicted Pod count list and the initial Pod count list are completely consistent, the target performance indicator prediction dataset is obtained based on the current initial performance indicator prediction dataset. This approach considers the impact of dynamic changes in the number of Pods on the load and performance indicators of individual Pods, ensuring that the prediction results more accurately reflect the true load distribution and improving the accuracy of the prediction results.

[0089] Specifically, in step S4, based on D, the target performance metric prediction dataset corresponding to each Pod of the target service is obtained, including:

[0090] S41. When count=1, use D as the target performance metric prediction dataset for each Pod corresponding to the target service.

[0091] S42. When count ≠ 1, input A and C into the time series model corresponding to A to obtain the updated D and use the updated D as the target performance metric prediction dataset for each Pod corresponding to the target service.

[0092] Through the above steps, when count=1, it means that the initial prediction has met the Pod count consistency requirement, and no iteration is needed. The initial prediction result D is directly used as the target performance indicator prediction dataset for each Pod, avoiding redundant calculations. When count≠1, it means that the initial prediction does not meet the Pod count consistency requirement. It takes multiple rounds of "prediction-feedback" iterations to meet the Pod count consistency requirement. In this case, it is necessary to combine the latest initial Pod count sequence and the total set of multi-dimensional performance indicator datasets, and re-call the corresponding time series model. The output of the time series model is used as the target performance indicator prediction dataset for each Pod corresponding to the target service. This ensures efficiency in steady-state scenarios and accuracy and consistency of prediction results in dynamic adjustment scenarios. It also eliminates the accumulated error or state residue in the intermediate state of iteration, improving the accuracy and reliability of the final prediction results.

[0093] In one specific embodiment, prior to step S42, the method further includes:

[0094] When count > count 0 When count ≤ count, output predicted anomaly information; when count ≤ count 0 When the prediction error occurs, proceed to step S42; the prediction error information is used to prompt the user that the number of Pods cannot be predicted further; count 0 This is the preset maximum count value.

[0095] Through the above steps, when count > count 0 At that time, output the predicted anomaly information to avoid infinite iteration.

[0096] Specifically, in step S4, D is updated based on C and E, including:

[0097] S401. Obtain C = (C1, C2, ..., C...). j C n E = (E1, E2, ..., E) j , ..., E n ), D = {D1, D2, ..., D} e D f}, D e =(D e1 D e2 D ej D en ); where C jLet E be the initial number of Pods corresponding to the j-th preset time point in T2. j D represents the number of predicted Pods corresponding to the j-th preset time point in T2, where 1 ≤ j ≤ n, and n is the number of preset time points in T2; e This is a list of initial performance index prediction values ​​corresponding to the e-th preset performance index, where 1 ≤ e ≤ f, and f is the number of preset performance indices; D ej D e The corresponding preset performance index is the initial index prediction value at the j-th preset time point in T2.

[0098] S402, Iterate through C and E, when C... j ≠E j At that time, let D ej =D ej / C j ×E j This is to enable updating of D.

[0099] Through the above steps, during the iteration process, when the predicted number of Pods output by the self-healing system is inconsistent with the current initial number of Pods at a certain preset time point, the initial performance index prediction data set of each preset performance index at that preset time point is scaled proportionally to achieve rapid update. This eliminates the need to rerun complex time-series models, minimizes computational overhead, and effectively reflects the impact of changes in the number of Pods on the performance index of a single instance, significantly improving the rationality of intermediate prediction results and accelerating closed-loop convergence.

[0100] The present invention also provides a specific embodiment, which differs from the above embodiment in that, after step S1, the following steps are included:

[0101] S10. Based on A, obtain the average dataset F.

[0102] S20. Based on the frequency of occurrence of associated flight events corresponding to the target service within each historical sub-time period corresponding to T1, obtain the first comprehensive event feature vector G1.

[0103] Specifically, the associated flight events corresponding to the target service are the flight dynamic events that the target service needs to respond to in the business process; for example, when the target service is a flight status management service, its associated flight events include: actual takeoff, actual arrival, flight delay, flight cancellation, etc.

[0104] Specifically, T1 is divided according to a preset interval to obtain several historical sub-time periods corresponding to T1; any two historical sub-time periods do not overlap; those skilled in the art know that the preset interval is an interval determined in advance by those skilled in the art according to actual needs, such as 5 minutes or 10 minutes, which will not be elaborated here.

[0105] Specifically, step S20 includes the following sub-steps:

[0106] S201. For each historical sub-time period, based on the occurrence frequency of associated flight events corresponding to the target service in the historical sub-time period, obtain a list of flight event occurrence frequency values ​​corresponding to that historical sub-time period.

[0107] Specifically, the list of flight event frequency values ​​includes the cumulative frequency value of each associated flight event corresponding to the target service within its corresponding historical sub-time period.

[0108] S202. Based on the list of frequency values ​​of flight events, obtain the feature vector of flight events corresponding to the historical sub-time period corresponding to the list of frequency values ​​of flight events.

[0109] Specifically, feature extraction is performed on the list of flight event frequency values ​​to obtain flight event feature vectors. As those skilled in the art know, any existing method for extracting features from a list of flight event frequency values ​​to obtain flight event feature vectors is within the scope of protection of this invention. For example, using an LSTM encoder to encode the list of flight event frequency values ​​to obtain flight event feature vectors; this will not be elaborated further here.

[0110] In another specific embodiment, feature vectors are extracted from the list of frequency values ​​of flight events. The feature vectors are then multiplied by the business weights corresponding to the historical sub-time periods of the list of frequency values ​​of flight events to obtain the flight event feature vectors.

[0111] Furthermore, the business weight W corresponding to the historical sub-time period meets the following condition:

[0112] W = α × V1 + β × V2, where α is the weighting coefficient corresponding to the preset key flights; β is the weighting coefficient corresponding to the non-preset key flights; V1 is the cumulative occurrence frequency of events triggered by preset key flights among all related flight events corresponding to the target service within the historical sub-period; V2 is the cumulative occurrence frequency of events triggered by non-preset key flights among all related flight events corresponding to the target service within the historical sub-period; wherein, preset key flights are flights predetermined by those skilled in the art based on actual needs.

[0113] Specifically, β < α, and the specific values ​​of α and β are set by those skilled in the art according to actual needs, for example: α = 1, β = 0.2, which will not be elaborated here.

[0114] Through the above steps, feature vectors are extracted from the list of flight event frequency values. These feature vectors are then multiplied by the business weights corresponding to the historical sub-time periods of the flight event frequency value list to obtain the flight event feature vector. The business weights corresponding to the historical sub-time periods are based on the weight coefficients corresponding to preset key flights and non-preset key flights, as well as the cumulative occurrence frequency of related flight events triggered by non-preset key flights and the cumulative occurrence frequency of related flight events triggered by preset key flights. This ensures that the flight event feature vectors can reflect the differences in the business impact of the same event on different flights, providing richer contextual information for the time series model.

[0115] S203. Following the order of historical sub-time periods from morning to night, concatenate the flight event feature vectors corresponding to all historical sub-time periods to obtain G1.

[0116] Through the above steps, the preset historical time period is divided into multiple historical sub-time periods, and the occurrence frequency of each associated flight event corresponding to the target service within each historical sub-time period is counted to form a fine-grained list of flight event occurrence frequency values. Based on the list of flight event occurrence frequency values, the flight event feature vector corresponding to the historical sub-time period corresponding to the list of flight event occurrence frequency values ​​is obtained. In the order of historical sub-time periods from morning to night, the flight event feature vectors corresponding to all historical sub-time periods are concatenated to obtain the first comprehensive event feature vector. This preserves the dynamic changes of business load in the time dimension, enabling the model to perceive "when what type of business activity occurred", thereby more accurately associating business events with changes in performance indicators, and thus enhancing the accuracy of performance prediction and business relevance.

[0117] S30. Based on the frequency of occurrence of associated flight events corresponding to the target service within each future sub-time period corresponding to T2, obtain the second comprehensive event feature vector G2.

[0118] Specifically, based on the flight schedule within T2, the frequency of occurrence of associated flight times for the target service within each future sub-time period corresponding to T2 is counted.

[0119] Specifically, in steps S201-S203, G1 is obtained based on the frequency of occurrence of associated flight events corresponding to the target service within each historical sub-time period corresponding to T1, and G2 is obtained based on the frequency of occurrence of associated flight events corresponding to the target service within each future sub-time period corresponding to T2.

[0120] Specifically, T2 is divided according to a preset interval to obtain several future sub-time periods corresponding to T2; any two future sub-time periods do not overlap.

[0121] S40. Input F, H, G1 and G2 into the time series large model corresponding to A to obtain the target performance index prediction dataset for each Pod corresponding to the target service; wherein, the output of the time series large model is used as the target performance index prediction dataset for each Pod corresponding to the target service.

[0122] Through the above steps, a first comprehensive event feature vector is obtained based on the frequency of occurrence of related flight events corresponding to the target service within each historical sub-time period corresponding to a preset historical time period; a second comprehensive event feature vector is obtained based on the frequency of occurrence of related flight events corresponding to the target service within each future sub-time period corresponding to a preset future time period; the average dataset, the actual Pod quantity sequence corresponding to the preset historical time period, the first comprehensive event feature vector, and the second comprehensive event feature vector are input into the corresponding time-series large model to obtain the target performance indicator prediction dataset; in the multi-dimensional performance indicator prediction process, the average dataset, the actual Pod quantity sequence corresponding to the preset historical time period, the first comprehensive event feature vector, and the second comprehensive event feature vector are input together into the time-series large model for joint modeling, which not only considers the impact of changes in the number of Pods on the load of a single instance, but also introduces flight event information that can be known in advance; this enables the time-series large model to perceive changes in resource demand caused by business activities in advance, effectively enhancing the ability to predict sudden or periodic loads, thereby achieving more accurate and business-semantic Pod-level performance prediction and significantly improving the accuracy of prediction results.

[0123] Specifically, in this embodiment, the second preset training sample includes: an average dataset, a historical Pod count sequence corresponding to the average dataset, a first comprehensive event feature vector corresponding to the average dataset, a second comprehensive event feature vector corresponding to the average dataset, and an actual multidimensional performance index dataset corresponding to the average dataset.

[0124] Furthermore, the historical Pod count sequence corresponding to the average dataset includes the number of Pods corresponding to the services corresponding to the total set of specific multidimensional performance index datasets corresponding to the average dataset at each sampling time point in the sampling period corresponding to the average dataset; the sampling time points in the sampling period are determined based on a preset sampling period.

[0125] Furthermore, in step S20, the same method is used to obtain G1 based on the frequency of occurrence of associated flight events corresponding to the target service within each historical sub-time period corresponding to T1; the first comprehensive event feature vector corresponding to the average dataset is obtained based on the frequency of occurrence of associated flight events corresponding to the service within each sub-time period corresponding to the sampling time period corresponding to the average dataset; the method for dividing the sub-time periods corresponding to the sampling time period is consistent with the method for dividing the historical sub-time periods corresponding to the preset historical time period.

[0126] Furthermore, in step S20, the same method is used to obtain G1 based on the frequency of occurrence of associated flight events corresponding to the target service within each historical sub-time period corresponding to T1; based on the frequency of occurrence of associated flight events corresponding to the service in each sub-time period corresponding to the future time period corresponding to the average dataset, the second comprehensive event feature vector corresponding to the average dataset is obtained; the method for dividing the sub-time periods corresponding to the future time period is consistent with the method for dividing the historical sub-time periods corresponding to the preset historical time period.

[0127] The present invention also provides a specific embodiment, which differs from the above embodiment in that the following steps are included after step S30:

[0128] S100. Input F, H, G1, G2 and C into the time series model corresponding to A to obtain D.

[0129] In this embodiment, the second preset training sample includes: an average dataset, a historical Pod count sequence corresponding to the average dataset, a first comprehensive event feature vector corresponding to the average dataset, a second comprehensive event feature vector corresponding to the average dataset, an actual Pod count sequence corresponding to the average dataset, and an actual multidimensional performance index dataset corresponding to the average dataset.

[0130] S200. Input D into the self-healing system to obtain the predicted Pod count list E corresponding to T2. At the same time, let count = count + 1 to update count.

[0131] S300. When C and E are completely identical and count = 1, use D as the target performance metric prediction dataset for each Pod corresponding to the target service; when C and E are completely identical and count ≠ 1, if count > count 0 When C and E are completely identical and count ≠ 1, output the predicted anomaly information; when count ≤ count 0 When F, H, G1, G2 and C are input into the time series large model corresponding to A to obtain the updated D and use the updated D as the target performance index prediction dataset for each Pod corresponding to the target service; when C and E are not completely consistent, D is updated based on C and E, E is used as C and the process proceeds to step S200.

[0132] Through the above steps, a list of initial Pod counts corresponding to a preset future time period is used as input. This list is then combined with a time-series large-scale model to generate an initial performance metric prediction dataset. This dataset is then input into a self-healing system, which generates a predicted Pod count list. Further, when the predicted Pod count list and the initial Pod count list are not completely identical, the initial performance metric prediction dataset and the initial Pod count list are updated, and a predicted Pod count sequence is generated again based on the self-healing system for the next round of evaluation. When the predicted Pod count list and the initial Pod count list are completely identical, the target performance metric prediction dataset is obtained based on the current initial performance metric prediction dataset. This approach considers the impact of dynamic changes in the number of Pods on the load and performance metrics of individual Pods, ensuring that the prediction results more accurately reflect the actual load distribution and improving the accuracy of the prediction results.

[0133] Embodiments of the present invention also provide a non-transitory computer-readable storage medium that can be disposed in an electronic device to store a computer program related to implementing a method in the method embodiments, the computer program being loaded and executed by the processor to implement the method provided in the above embodiments.

[0134] Embodiments of the present invention also provide an electronic device, including: a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method provided in the above embodiments.

[0135] Embodiments of the present invention also provide a computer program product including program code, which, when the program product is run on an electronic device, causes the electronic device to perform the steps of the methods described above in various exemplary embodiments of the present invention.

[0136] While specific embodiments of the invention have been described in detail by way of examples, those skilled in the art should understand that the examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art should also understand that various modifications can be made to the embodiments without departing from the scope and spirit of the invention.

Claims

1. A method for predicting performance metrics for Pods, characterized in that, The method includes the following steps: S1. Obtain the set A = {A1, A2, ..., A...} of the multidimensional performance index dataset. i A n };A i A is a multi-dimensional performance metric dataset for the i-th Pod corresponding to the target service within a predefined historical time period T1, where 1 ≤ i ≤ n, and n is the number of Pods corresponding to the target service within T1; i It includes a time series data list corresponding to each preset performance indicator, and the time series data list includes the continuous sampled values ​​of the corresponding preset performance indicator in T1. The end time of T1 is the current time, and the duration is the first preset duration t1; S2. Based on A, the initial Pod count list C corresponding to the preset future time period T2, and the time series model corresponding to A, obtain the initial performance metric prediction dataset D corresponding to A; the starting time of T2 is the current time, and the duration is the second preset duration t2; C includes the initial Pod count corresponding to each preset time point in T2; D includes the initial metric prediction value list corresponding to each preset performance metric, and the initial metric prediction value list includes the initial metric prediction value corresponding to each preset time point in T2; t2 < t1; the initial Pod count is consistent with the current Pod count corresponding to the target service; S3. Input D into the self-healing system to obtain the predicted Pod count list E corresponding to T2; E includes the predicted Pod count corresponding to each preset time point in T2; wherein, while obtaining E, set count = count + 1 to update count; count is a pre-designed value and count is initially 0; S4. When C and E are completely consistent, if count=1, then D is used as the target performance metric prediction dataset for each Pod corresponding to the target service. If count≠1, then A and C are input into the time series large model corresponding to A to obtain the updated D and the updated D is used as the target performance metric prediction dataset for each Pod corresponding to the target service. When C and E are not completely consistent, D is updated based on C and E, E is used as C, and the process proceeds to step S3. Following step S1, the following steps are also included: S40. Input F, H, G1, and G2 into the time series model corresponding to A to obtain the target performance index prediction dataset for each Pod corresponding to the target service. Here, F is the average dataset obtained based on A; H is the actual Pod count sequence corresponding to T1, including the number of Pods corresponding to the target service at each historical sampling time point in T1; G1 is the first comprehensive event feature vector obtained based on the occurrence frequency of related flight events corresponding to the target service in each historical sub-time period corresponding to T1; G2 is the second comprehensive event feature vector obtained based on the occurrence frequency of related flight events corresponding to the target service in each future sub-time period corresponding to T2.

2. The method for predicting performance metrics for Pods according to claim 1, characterized in that, The sampled values ​​in the time series data list correspond one-to-one with the historical sampling time points in T1, and are arranged in order from earliest to latest according to the corresponding historical sampling time points; the historical sampling time points are determined based on the preset sampling period.

3. The method for predicting performance metrics for Pods according to claim 2, characterized in that, In step S4, D is updated based on C and E, including: S401. Obtain C = (C1, C2, ..., C...). j C n E = (E1, E2, ..., E) j , ..., E n ), D = {D1, D2, ..., D} e D f }, D e =(D e1 D e2 D ej D en ); where C j Let E be the initial number of Pods corresponding to the j-th preset time point in T2. j D represents the number of predicted Pods corresponding to the j-th preset time point in T2, where 1 ≤ j ≤ n, and n is the number of preset time points in T2; e This is a list of initial performance index prediction values ​​corresponding to the e-th preset performance index, where 1 ≤ e ≤ f, and f is the number of preset performance indices; D ej D e The corresponding preset performance index is the initial index prediction value at the j-th preset time point in T2; S402, Iterate through C and E, when C... j ≠E j At that time, let D ej =D ej / C j ×E j This is to enable updating of D.

4. The method for predicting performance metrics for Pods according to claim 3, characterized in that, The j-th preset time point in T2 is earlier than the (j+1)-th preset time point in T2, and the interval between two adjacent preset time points in T2 is consistent with the duration corresponding to the preset sampling period.

5. The method for predicting performance metrics for Pods according to claim 1, characterized in that, After step S1 and before step S2, the following steps are also included: S01, based on A1, A2, ..., A i A n Obtain the average correlation score R corresponding to A; S02, when R > R 0 When the time comes, proceed to step S2; R 0 The preset relevance score.

6. The method for predicting performance metrics for Pods according to claim 1, characterized in that, The steps preceding step S2 also include: S001, Based on A i And an indicator trend type identification model, which determines A from several preset indicator trend types. i The corresponding target indicator change trend type; where each preset indicator change trend type corresponds to a time series large model; S002, If A1, A2, ..., A i A m If the corresponding target indicator change trend types are all consistent, then the time series model corresponding to the target indicator change trend type is taken as the time series model corresponding to A.

7. A non-transitory computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by a processor to implement the performance metric prediction method for Pods as described in any one of claims 1-6.

8. An electronic device, comprising: A processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the computer program, it implements the performance metric prediction method for a Pod as described in any one of claims 1-6.

Images