A container elastic scaling method based on asymmetric risk perception and double-channel space-time feature fusion

Abstract

The present application relates to the technical field of cloud computing container orchestration and artificial intelligence operation and maintenance, and particularly relates to a container elastic scaling method based on asymmetric risk perception and double-channel space-time feature fusion. The method comprises the following steps: S1: constructing a real-time load data stream based on multi-order differential enhancement, S2: constructing a load prediction model based on double-channel space-time feature fusion, S3: model training based on an asymmetric weighted loss function, S4: constructing an elastic scaling decision system based on prediction perception, and S5: active resource scheduling execution and anti-jitter design. The method overcomes the symmetry assumption defects of traditional prediction models, significantly reduces the default risk caused by resource under-provisioning, constructs an active defense mechanism for business risk at the algorithm level, and meets the stringent requirements of enterprise-level production environments for high stability. The method has comprehensive feature capture, strong model generalization capability, and can eliminate the cold start time difference, achieving zero-delay response.

Description

Technical Field

[0001] This invention relates to the technical fields of cloud computing container orchestration and artificial intelligence operation and maintenance, and in particular to a container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion. Background Technology

[0002] With the widespread adoption of cloud-native technologies, Kubernetes has become the de facto standard for container orchestration. In scenarios with extremely high real-time requirements, such as e-commerce promotions, high-frequency financial trading, and the industrial internet, application workloads typically exhibit characteristics of drastic instantaneous fluctuations and high-frequency noise hidden within periodic patterns. Elastic scaling, as a core capability of cloud-native technologies, aims to minimize resource costs while ensuring service level agreements (SLAs).

[0003] However, existing container elastic stretching technologies still have the following three shortcomings:

[0004] First, there's the lag in scaling strategies. Currently, mainstream Kubernetes HPA (Horizontal PodAutoscaler) employs a threshold-based reactive strategy, triggering scaling only when resource utilization exceeds a preset value. From monitoring data collection and aggregation to triggering the decision, there's typically a delay of minutes. This passive response mechanism cannot handle millisecond-level bursts of traffic, easily leading to service avalanche.

[0005] Second, there is a flaw in the symmetry assumption of the prediction model. Existing proactive scaling solutions mostly employ time-series models such as Long Short-Term Memory (LSTM) networks and Autoregressive Differential Moving Average (ARIMA) models, whose training process is typically based on symmetric loss functions such as mean squared error (MSE). This mechanism implicitly assumes that the cost of overestimating or underestimating is equal. However, in production environments, the cost of SLA default due to resource under-allocation far outweighs the waste caused by resource over-allocation. Current technologies lack algorithmic modeling for this asymmetric risk, resulting in models that cannot output predictions with safety redundancy. This makes the system highly susceptible to resource shortages when facing load fluctuations, making it difficult to meet enterprise-level high availability requirements.

[0006] Third, there is a lack of dynamic compensation for end-to-end startup latency. Actual business loads often combine long-term tidal patterns with high-frequency instantaneous noise, making it difficult for a single model to accurately capture both spatiotemporal characteristics simultaneously. Furthermore, existing methods often only predict when the load will increase, neglecting the end-to-end startup latency required for containers to go from scheduling instructions to readiness. Without considering dynamically changing startup time, even if load prediction is accurate, resources cannot be ready in time before traffic peaks arrive, resulting in a resource mismatch problem where scaling operations are completed but business requests have already timed out.

[0007] Therefore, how to provide an effective method for elastic scaling of containers to solve the problems of response lag, asymmetric risk perception, single feature capture, and resource mismatch caused by cold start delay in existing work has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] In view of the shortcomings of the prior art described above, the purpose of this invention is to propose a container elastic scaling method and system based on asymmetric risk perception and dual-channel spatiotemporal feature fusion.

[0009] First, this invention constructs a real-time load data stream based on multi-order difference enhancement. Unlike traditional monitoring methods that only collect values ​​at the current moment, this invention introduces a time-series feature engineering method. It calculates the first-order difference, representing the rate of load change, and the second-order difference, representing the acceleration of change, and combines this with an adaptive hierarchical interpolation strategy for data cleaning. This method constructs a time-series dataset that not only reflects the current load amplitude but also sensitively characterizes the trend and intensity of load changes.

[0010] Secondly, this invention constructs a load prediction model based on the fusion of dual-channel spatiotemporal features. This model employs a dual-channel parallel feature extraction architecture: the local mutation channel utilizes a feature-enhanced temporal convolutional network combined with dilated causal convolution to extract high-frequency traffic spike features; the global memory channel utilizes a self-attention-gated recurrent unit to capture long-term regular features. The two channels are adaptively fused through a dynamic gating mechanism, significantly improving the model's generalization ability under complex mixed traffic patterns.

[0011] Furthermore, this invention designs a risk-aware end-to-end latency compensation strategy. At the algorithm level, an asymmetric weighted loss function is introduced, which guides the model to build a prediction safety buffer by increasing the gradient penalty weight for underestimating the load. At the engineering level, the end-to-end startup latency profile of the container is calculated in real time, and the advance amount of capacity expansion triggering is dynamically adjusted to complete resource preheating before the traffic peak arrives, effectively avoiding the risk of service level agreement (SLA) default due to cold start delay.

[0012] Technical solution

[0013] A container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion includes the following steps:

[0014] S1: Construct a real-time load data stream based on multi-order difference enhancement; collect historical load time-series data of container clusters, introduce time-series feature engineering methods, calculate the first-order difference representing the rate of load change and the second-order difference representing the acceleration of change, and combine an adaptive hierarchical interpolation strategy to clean missing values, thereby constructing a time-series dataset that not only reflects the current load amplitude, but also represents the trend and intensity of load change.

[0015] S2: Construct a load prediction model that fuses spatiotemporal features in two channels; construct a dual-channel parallel deep learning model that includes a local channel for extracting local mutation features and a global channel for global long-range temporal memory; the local channel uses a feature-enhanced temporal convolutional network combined with dilated causal convolution to extract high-frequency traffic spike features; the global channel uses a self-attention gated recurrent unit to capture long-cycle regular features; the two are adaptively fused through a dynamic gating mechanism.

[0016] S3: Model training based on asymmetric weighted loss function; the processed dataset is divided into training and test sets, and a risk penalty coefficient is introduced during training to construct an asymmetric loss function; by increasing the gradient penalty weight for the underestimation of load region, the model is forced to establish a positive prediction safety buffer during gradient descent, thereby constructing an active defense mechanism against resource under-allocation risk at the algorithm level.

[0017] S4: Construct an elastic scaling decision system based on prediction awareness; monitor the entire process time from container scheduling to service readiness in real time, and update the startup latency dynamic profile using an exponentially weighted moving average algorithm; combine the load prediction value output in step S2, use a multi-index parallel evaluation strategy to calculate the target number of container replicas, and determine the advance triggering time of the expansion operation based on the startup latency profile, so as to complete proactive resource preheating before the traffic peak arrives.

[0018] S5: Performs proactive resource scheduling and anti-jitter processing; when the target number of replicas exceeds the current number of running replicas, it calls the cluster API to modify the replica field and performs seamless scaling in conjunction with the ready probe mechanism; at the same time, it introduces an asymmetric cooling window mechanism, which uses the time dimension hysteresis effect to filter instantaneous load jitter by setting a scaling-down judgment threshold and a cooling lock period after scaling-down execution, ensuring the high availability of the system.

[0019] Beneficial effects

[0020] (1) High risk controllability, effectively guaranteeing the service level agreement. This invention overcomes the symmetry assumption of traditional prediction models by introducing an asymmetric weighted loss function. By increasing the penalty weight for underestimating the load, this invention guides the model to build a positive prediction safety buffer, significantly reducing the risk of default due to resource under-allocation. At the algorithm level, it constructs an active defense mechanism against business risks, meeting the stringent requirements of high stability in enterprise-level production environments.

[0021] (2) Comprehensive feature capture and strong model generalization ability. This invention adopts a dual-channel parallel spatiotemporal feature fusion architecture, combined with multi-order differential feature inputs representing load rate and acceleration, effectively solving the technical problem that a single-structure model cannot take into account both local mutations and global dependencies. This architecture uses local channels to capture high-frequency burst features and global channels to capture long-range periodic patterns, resulting in higher prediction accuracy under complex and variable mixed flow patterns.

[0022] (3) Eliminate cold start latency and achieve zero-latency response. This invention avoids the service interruption risk caused by cold start latency by constructing a dynamic profile of startup latency and a full-link latency compensation strategy. The system dynamically adjusts the expansion lead based on the real-time updated startup time, realizing precise resource preheating and effectively solving the problem of resource supply and business demand mismatch caused by expansion lag. Attached Figure Description

[0023] Figure 1 is an overall flowchart of the present invention;

[0024] Figure 2 is a diagram of the overall architecture of the load prediction model based on dual-channel spatiotemporal feature fusion of the present invention.

[0025] Figure 3 is a detailed network structure design diagram of the load prediction model with dual-channel spatiotemporal feature fusion of the present invention;

[0026] Figure 4 is a comparison curve of the load prediction results of the embodiment of the present invention and the benchmark model. Detailed Implementation

[0027] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention.

[0028] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0029] A container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion is presented. Its core process is shown in Figure 1 and includes the following steps:

[0030] S1: Constructing a real-time load data stream based on multi-order differential enhancement

[0031] To address the issues of high noise and limited features in traditional monitoring data, this invention constructs a millisecond-level monitoring system and introduces a multi-dimensional feature derivation mechanism.

[0032] S11: Collection of Multi-Dimensional Monitoring Indicators

[0033] Deploy the Prometheus monitoring component and Node Exporter in a Kubernetes cluster and configure a high-frequency data collection cycle. Second. Raw indicator vectors collected at different times include:

[0034] Basic resource metrics: CPU utilization Memory usage .

[0035] Business performance metrics: Requests per second (QPS), P90 response latency .

[0036] Network metrics: Network I / O rate .

[0037] S12: Missing value filling and anomaly cleaning

[0038] To address potential query timeouts or packet loss issues during high-frequency data acquisition, this embodiment abandons the computationally complex Lagrange interpolation and adopts an adaptive hierarchical strategy to meet real-time requirements. (Short-term missing data - number of missing items...) (per cycle): Linear interpolation is used to fill in instantaneous packet loss.

[0039]

[0040] in, The valid observation value of the period preceding the missing time; For missing moments The first valid data point afterwards; For this valid data point relative to the current time The time step offset.

[0041] Long-term absence ( (Per cycle): The previous value preservation method is used, and data confidence markers are set to prevent misleading the model. For periods exceeding... Outliers are considered as monitoring noise and are smoothly replaced using a sliding window midpoint. This is the arithmetic mean of the load data within the current time sliding window. This represents the standard deviation of the data within the sliding window.

[0042] S13: Feature Derivation and Normalization

[0043] To enhance the model's perception of load variation trends, this invention introduces time-series difference features. Calculation Time observation value First-order difference and second-order difference :

[0044]

[0045]

[0046] Specifically, for the start time Since there is no historical data, the following settings are made: Similarly, set as well as .

[0047] S14: Dimensional expansion through concatenation of eigenvectors

[0048] Original collection of indicators for , dimension Calculate the first and second differences for each of the above indicators to obtain the velocity vector. and acceleration vector Constructed through vector concatenation operations Extended feature vector at time step :

[0049]

[0050] At this point, the feature dimension of the input data changes from... It is expanded to 15 dimensions. This dimensionality expansion operation allows the model to include not only the current static load value in the input at each time step, but also the dynamic evolution trend of the load.

[0051] Because the original values ​​and the difference values ​​(which may be negative or decimals) have vastly different dimensions, the expanded feature matrix is ​​independently normalized column-wise to accelerate gradient descent convergence. For the ... Each feature dimension ( The value is mapped to the [0, 1] interval using Max-Min normalization:

[0052]

[0053] All normalized elements The recombined features result in the final model input feature matrix. .

[0054] S15: Temporal Tensor Construction

[0055] The sliding window method is used to construct the model input. In this embodiment, the window size is set to [size missing]. =32, prediction step size H=8. Feature matrix constructed in step S14 Based on (where) For the total time step, For any sampling time ( ), cut length is Constructing a local feature matrix from continuous feature fragments :

[0056]

[0057] in, Representation matrix In the The feature vector of the row. Then, by sliding sampling with a step size of 1 on the time axis, multiple local feature matrices are combined. Stacking along the batch dimension, the final constructed temporal tensor is represented as follows: , where B is the batch size.

[0058] S2: Construct a load prediction model that fuses dual-channel spatiotemporal features.

[0059] The dual-channel spatiotemporal feature fusion load prediction model includes: a local mutation feature extraction module, a global long-range temporal memory module, and a dynamic gating fusion module. Its overall architecture is shown in Figure 2.

[0060] Among them, the local mutation feature extraction module uses the feature-enhanced temporal convolutional network (FE-TCN) to capture the high-frequency fluctuation features of the load data; the global long-range temporal memory module uses the self-attention gated recurrent unit (SA-GRU) to capture the periodic patterns of the load data; and the dynamic gated fusion module is responsible for adaptively weighting and fusing the two features to finally output the load prediction value at future time.

[0061] The model is based on the high-dimensional temporal tensor generated in step S1. Parallel processing is performed, and the detailed processing flow is shown in Figure 3, explained below:

[0062] S21: Constructing a local mutation feature extraction module

[0063] This module aims to extract local features reflecting millisecond-level traffic spikes and instantaneous fluctuations from the input temporal tensor. Compared to traditional convolutional networks, this module introduces a feature channel attention mechanism in the residual block to achieve adaptive filtering of key load metrics (such as the rate of change of requests per second).

[0064] S211: Dilated Causal Convolution Processing

[0065] To expand the model's receptive field without revealing future information, this embodiment stacks four layers of residual blocks. The expansion coefficient of each layer increases exponentially (set to 1, 2, 4, 8).

[0066] For the Layer, input sequence is Time step Convolution output at the point Calculated using the following formula:

[0067]

[0068] In the formula, The kernel size is 3 in this embodiment. The position index within the convolution kernel ( ), used to traverse the convolution window; For the first The expansion coefficient of the layer, For the first Layer position The learnable weight matrix at that location, This is a bias term. This structure allows the top-level neurons to cover a sufficiently long history of mutations.

[0069] S212: Feature Channel Attention Filtering

[0070] To achieve adaptive feature selection (e.g., automatically increasing the weight of QPS features while suppressing memory features during traffic bursts), the convolution output... Post-embedded attention module:

[0071] Global information compression: Features are compressed along the time dimension using global average pooling, and the convolutional output is compressed. Mapped to channel descriptor . No. The formula for calculating each channel is:

[0072]

[0073] In the formula, The time dimension length (i.e., sequence length) of the current layer feature map. The total number of feature channels, i.e., the number of convolutional kernels (set in this embodiment). To match the dimensionality reduction ratio (divisibility requirement); For channel index ( ); Indicates the first Layer Each channel at time step The eigenvalue at that location.

[0074] Weight generation: By capturing the non-linear dependencies between channels through two fully connected layers, a normalized weight vector is generated. :

[0075]

[0076] in It is the ReLU activation function. This is the Sigmoid function. For the dimensionality reduction weight matrix (dimensionality reduction ratio) ), This is the weight matrix for increasing dimensionality.

[0077] Feature recalibration: The generated weight vector Treat it as a feature selector, and combine it with the convolution output. Perform channel-by-channel multiplication to obtain the weighted features. :

[0078]

[0079] Residual connections: To prevent network degradation and accelerate convergence, weighted features are used to perform residual connections. The original input of the residual block in this layer Perform element-wise addition and then output the final result after passing it through an activation function:

[0080]

[0081] After processing the stacked residual blocks, the output of the last layer is taken at the current time step. The feature vector, denoted as the local mutation feature vector. ( This serves as one of the inputs for the subsequent dual-channel fusion module.

[0082] S22: Construct a global long-term temporal memory module

[0083] This module aims to capture the macroscopic tidal patterns of load data on a daily basis, resolving the problem of long-sequence dependencies. This module works in parallel with the S21 module, processing the same input time-series tensors. .

[0084] S221: Gated Loop Unit Processing

[0085] The basic backbone utilizes a gated recurrent unit (GRU). Compared to Long Short-Term Memory (LSTM) networks, GRU combines the forget gate and input gate into an update gate, reducing the number of parameters by approximately 30%, making it more suitable for high-frequency real-time inference scenarios. For time steps... The input is The hidden state at the previous moment was The specific calculation process is as follows:

[0086] Reset door Used to control the degree of historical information discarded; the weight matrix is ​​as follows: The calculation formula is:

[0087]

[0088] Update Gate : Used to control the degree of retention of historical information, the weight matrix is ​​as follows The calculation formula is:

[0089]

[0090] Hidden state update: The weight matrix is Based on the above gating results, candidate hidden states are calculated. With the final hidden state :

[0091]

[0092]

[0093] Through the above recursive calculation, a hidden state sequence containing long-range dependency information is obtained.

[0094] S222: Masked Multi-Attention Enhancement

[0095] To further address the gradient vanishing or forgetting problem in recurrent networks under ultra-long sequences, a multi-head self-attention mechanism is introduced into the GRU output sequence to directly establish the dependency between the current time step and long historical time steps.

[0096]

[0097] in, These represent the query vector, key vector, and value vector, respectively, all of which are generated by linear mapping from the hidden state sequence output by the GRU. This is a scaling factor used to adjust the numerical range of the dot product result to prevent gradient vanishing. This is an upper triangular mask matrix. In this embodiment, the elements of its upper triangular region (excluding the diagonal) are set to... This masking mechanism forces the model to perform calculations... When attention is distributed at any given moment, it can only converge. The system incorporates historical information up to and including the timeframe, thus strictly ensuring the causality of time-series modeling and preventing the model from utilizing future information.

[0098] After the above calculations, the module finally outputs a global feature vector containing long-range dependency information. ( ,in ), which serves as the second input for the subsequent dual-channel fusion module.

[0099] S23: Constructing a dynamic gating fusion module and predictive output

[0100] This module receives the local mutation feature vector from S21. and global feature vectors from S22 An adaptive gating mechanism is used to dynamically adjust the contribution ratio of the two features, ultimately generating a load prediction value.

[0101] S231: Generation of gating coefficients

[0102] To quantify whether the current load is more inclined towards a "mutational state" or a "periodic state," the model first concatenates the two feature vectors along the channel dimension to construct a joint feature vector. (in ).

[0103] Subsequently, the joint features are mapped to scalar gating coefficients using a multilayer perceptron. This process incorporates nonlinear transformations to enhance expressive power; the calculation formula is as follows:

[0104]

[0105] In the formula, It is a ReLU activation function; Use the Sigmoid activation function to ensure the output coefficients ; and The weight matrix is ​​a learnable matrix; This is a bias term.

[0106] S232: Adaptive Feature Fusion

[0107] Using the generated gating coefficients As a modulating factor, the two feature streams are soft-switched and weighted fused. When the value approaches 1, the model mainly relies on the local mutation features extracted by TCN; when... When the value approaches 0, the model primarily relies on long-range historical features extracted by GRU. The fused feature vector... The calculation is as follows:

[0108]

[0109] S233: Load Prediction Output

[0110] Finally, the fused feature vectors The input is fed into a linear regression layer, which maps the high-dimensional feature space back to the target payload space, and outputs the prediction results for future time moments. :

[0111]

[0112] In the formula, and These are the weights and biases for the output layer. Prediction results. It includes predicted values ​​for key metrics such as CPU utilization, memory usage, and QPS, which are used for subsequent elastic scaling decisions.

[0113] S3: Model Training Based on Asymmetric Weighted Loss Function

[0114] As shown in Figure 3, this embodiment introduces a risk-aware asymmetric loss function during the model training phase, aiming to address the shortcomings of the traditional mean squared error (MSE) in treating positive and negative errors equally, thereby constructing a safety buffer at the algorithm level.

[0115] S31: Construction of the Asymmetric Weighted Loss Function

[0116] In container elastic scaling scenarios, a predicted load lower than the actual load can lead to service avalanche, the cost of which is far greater than the resource idle cost caused by a predicted load slightly higher than the actual load. Therefore, this invention defines an asymmetric loss function. as follows:

[0117]

[0118] in Indicates the size of the training batch; Represents the actual load observation value. This represents the model's predicted output value; This is an indicator function that takes the value 1 when the condition inside the parentheses is met, and 0 otherwise; In this embodiment, the risk penalty coefficient is set as follows: .

[0119] When the actual load Greater than the predicted value When the indicator function is activated, the error gradient is amplified. This mechanism forces the model to learn a prediction curve slightly higher than the true load mean in order to minimize the total loss during gradient descent optimization. By introducing this asymmetric gradient penalty mechanism at the algorithmic level, the model implicitly establishes a safety buffer to cope with sudden traffic surges, thus eliminating the need to artificially add hard threshold rules in subsequent decision-making stages.

[0120] S32: Optimizer Configuration and Learning Rate Scheduling

[0121] To ensure the stable convergence of the dual-channel models (FE-TCN and SA-GRU), this embodiment designs a specific training strategy. This strategy utilizes a collaborative configuration optimization algorithm, a learning rate scheduling mechanism, and an early stopping criterion to achieve rapid fitting and global optimization of complex non-stationary objective functions. Specifically, it includes the following:

[0122] Parameter optimization uses the Adam optimizer for parameter updates. The first moment is used to estimate the exponential decay rate. Second-order moment estimation of exponential decay rate To accommodate sparse gradients and non-stationary objective functions.

[0123] Learning rate scheduling incorporates a cosine annealing strategy to escape local optima. Initial learning rate. Minimum learning rate Learning rate With the current Epoch The formula for change is:

[0124]

[0125] Compared to traditional step-wise decay, cosine annealing utilizes the smoothness of the cosine function to achieve a non-linear, seamless decay of the learning rate. In the early stages of training, a higher learning rate allows the model to quickly traverse the flat region of the loss function, accelerating convergence. As training nears its end, the learning rate gradually decays to a minimum with a smooth curve, enabling the optimizer to perform fine-tuning searches near the global optimum. This mechanism effectively balances training speed and accuracy, preventing the model from drastically oscillating near the optimal solution.

[0126] Monitor the asymmetric loss value on the validation set. If the validation set loss does not decrease for 10 consecutive epochs, terminate training early and save the current optimal weights to prevent overfitting.

[0127] S4: Building a resilient and scalable decision-making system based on prediction awareness

[0128] This step aims to convert the load forecast value output from step S2 into a more accurate representation of the load forecast. Transforming this into specific resource scheduling instructions hinges on resolving two key issues: the amount of computing capacity expansion and the timing of expansion.

[0129] S41: Target Replica Count Calculation

[0130] After obtaining the load prediction results, this embodiment extracts three metrics directly related to container physical resource quotas—CPU utilization, memory usage, and QPS—as baseline dimensions. Other features, such as network performance and latency, serve as auxiliary references to improve prediction accuracy. For each load dimension... According to the model's prediction of the future Time-of-flight load value Based on the current resource utilization rate and the corresponding target threshold Calculate the required number of replicas respectively. To eliminate the uncertainty risk of model predictions, this embodiment introduces a safety redundancy coefficient. The deterministic resource buffer pool is constructed using the following formula:

[0131]

[0132] In the formula, This represents the number of currently running replicas. This represents the floor function; the range of values ​​in this embodiment is... Based on the probabilistic prediction in step S3, this coefficient introduces regular hard constraints to construct a deterministic resource buffer pool. This compensates for resource gaps that may arise due to model prediction bias or sudden traffic noise, thereby ensuring the system's service quality. To ensure overall system high availability, the final target number of replicas... Take the maximum value of the results calculated for all dimensions, that is:

[0133]

[0134] S42: End-to-End Delay Compensation Decision

[0135] To achieve proactive resource preheating and minimize response latency, the system must calculate the advance trigger time window. This is to offset the physical time required for resource startup. The calculation formula is as follows:

[0136]

[0137] In the formula The cold start time of a Pod is not fixed. This invention continuously monitors the actual time taken for a Pod to go from scheduling to the ready state using the Prometheus open-source monitoring and alerting system, and dynamically updates it using an exponentially weighted moving average algorithm. The calculation formula is as follows:

[0138]

[0139] in, For the current moment Updated startup latency estimate; This represents the actual physical time taken for a Pod to go from scheduling to readiness, as monitored at the current moment. For the previous moment The estimated startup latency; The smoothing factor (0.3 in this example) is used to balance the weights of historical trends and current observations to adapt to fluctuations in mirror pull speed or changes in node load. For system safety buffer time, this embodiment takes 5 to 10 seconds to cover Kubernetes API scheduling latency and service registration and discovery synchronization latency.

[0140] Based on the aforementioned time window, the system selects the sequence with the closest time span from the future load prediction sequence output in step S3. Prediction points This serves as a basis for decision-making. If the predicted load at a given moment exceeds the current capacity, then immediately... This triggers a scaling command. This mechanism ensures that when the future traffic peak actually arrives, the scaling Pods will have already started and be ready.

[0141] Step S5: Proactive Resource Scheduling Execution and Anti-Jitter Design

[0142] S51: Seamless Capacity Expansion Execution

[0143] The final target replica count output by the system real-time monitoring step S41 .when Greater than the number of currently running replicas This indicates that, according to model predictions, the load at a future time will cause resource utilization to exceed the preset target threshold. (In this embodiment, 70% is used), and the system immediately triggers an expansion command.

[0144] To prevent resource exhaustion due to extreme prediction bias, this embodiment sets a limit on the maximum number of replicas in the cluster. (For example, set to 50 or the cluster quota limit). The actual number of target replicas allocated for expansion. Take the smaller of the calculated value and the upper limit value, that is:

[0145]

[0146] Sure Then, the system directly calls the Kubernetes Client-go API to modify the Replicas field of the target Deployment; at the same time, it uses the Kubernetes Readiness Probe mechanism to ensure that the newly expanded Pod is officially registered to the Service's endpoint list and takes over traffic only after it has completed initialization and warm-up, thus achieving seamless traffic switching.

[0147] S52: Volume Reduction Hysteresis Cooling Strategy

[0148] To prevent frequent Pod creation and destruction due to minor load fluctuations around a threshold, this embodiment introduces an asymmetric cooling window mechanism. For scaling up operations, the cooldown time is set to 0 seconds, meaning scaling up commands have the highest priority and are executed immediately upon detecting a risk of insufficient resources, thus ensuring high service availability. For scaling down operations, this embodiment sets the scaling down threshold to 30%. A scaling down command is triggered only when the model-predicted future load value is consistently lower than this threshold, after which the system enters a 300-second lockout state. During this period, new scaling down commands are prohibited, but scaling up commands are allowed to preemptively execute.

[0149] This strategy effectively filters out erroneous operations caused by instantaneous load fluctuations through the hysteresis effect in the time dimension.

[0150] Example

[0151] This invention comprehensively verifies its beneficial effects through multiple comparative experiments conducted in a simulation environment, focusing on five dimensions: evaluation index definition, model training convergence, risk perception and prediction performance, elastic scaling theory effectiveness, and ablation experiments. The specific implementation process and analysis are as follows:

[0152] (1) Experimental environment and dataset construction

[0153] The experimental hardware platform uses a 13th Gen Intel(R) Core(TM) i7-13700KF (3.40 GHz) processor, equipped with 32.0 GB of memory and an NVIDIA GeForce RTX 4060 (8GB) graphics card; the software environment is based on a Kubernetes v1.24 cluster, the operating system is Windows 11 (for model training and simulation), and the deep learning framework is PyTorch 1.12.

[0154] The dataset used was the publicly available and authoritative Alibaba Cluster-trace-v2018. Fifty online service container data with typical tidal fluctuations and sudden bursts were selected and divided into training and testing sets in an 8:2 ratio. Real HTTP request pressure was simulated using a traffic replay tool.

[0155] (2) Definition of evaluation indicators

[0156] To quantitatively measure the beneficial effects of the method of the present invention, the following indicators are used for evaluation in this embodiment:

[0157] Basic error metrics: Mean squared error (MSE) and mean absolute error (MAE) are used to measure the overall deviation between the predicted value and the actual load.

[0158] Risk perception capability: To measure the model's ability to defend against resource shortage risks, an underestimation rate metric is defined:

[0159]

[0160] in For real load, To predict load, This is an indicator function. The lower the value of this indicator, the lower the probability of resource underallocation.

[0161] Safety buffer index: Defines the average deviation Positive values ​​indicate that the model has positive safety redundancy, while negative values ​​indicate that the model tends to underestimate resource requirements.

[0162] (3) Comparison of load timing prediction performance

[0163] The model's predictive performance on CPU utilization is, for example... Figure 4 As shown in the figure, the actual load observations exhibit significant volatility and non-stationarity, accompanied by multiple instantaneous peak abrupt changes. The baseline models (LSTM and Transformer, dashed lines) often show obvious lag at the peaks, and their predicted curves are mostly below the actual curves, failing to respond promptly to sudden load spikes and demonstrating insufficient ability to capture high-frequency abrupt changes. In contrast, the load prediction model based on dual-channel spatiotemporal feature fusion of this invention (solid red line) shows a higher degree of fit between its predicted curve and the actual load compared to the baseline model, closely following the load's rise and fall trends. Particularly in the high-frequency fluctuation range from time step 150 to 200, the model effectively covers the peaks of the actual load, demonstrating stronger feature capture and trend following capabilities. This intuitively verifies the effectiveness of the dual-channel architecture in extracting local abrupt changes and global temporal features.

[0164] (4) Comparison of multidimensional evaluation indicators and analysis of risk defense capabilities

[0165] The quantitative comparison results of the model on specific evaluation indicators are shown in Table 1.

[0166] Table 1

[0167]

[0168] The experimental results show that the dual-channel load prediction model proposed in this invention outperforms or matches the comparative model in all key evaluation indicators, demonstrating significant technical advantages.

[0169] Specifically, in terms of the mean absolute error (MAE) metric, the model of this invention achieves a result of 0.1534, which is lower than the Transformer model's 0.1573 and the LSTM model's 0.1731. This indicates that the model of this invention has reached an optimal level in terms of overall prediction accuracy and can provide a more accurate load prediction benchmark.

[0170] Meanwhile, the mean squared error (MSE) of this invention's model is 0.0412, which is only slightly different from the Transformer model's 0.0391. This indicates that the model of this invention achieves a level comparable to the existing best benchmark model in ensuring the stability of prediction results.

[0171] In terms of risk defense-related indicators, the model of this invention demonstrates significant advantages. Specifically, the resource underestimation rate of the model is reduced to 28.33%, significantly lower than the 40.67% of the Transformer model; simultaneously, the average deviation of the model is +0.0465, exhibiting a positive numerical characteristic. This positive deviation verifies that the model possesses algorithmic-level security defense capabilities, effectively addressing the risk of resource shortages caused by sudden traffic surges.

[0172] In summary, the model of this invention not only achieves industry-leading accuracy in basic fitting, but also significantly improves business security and risk avoidance capabilities, effectively balancing prediction accuracy and service high availability.

[0173] (5) Ablation experiment analysis

[0174] This embodiment uses ablation experiments to gradually remove key components in order to verify the contribution of each module to the model performance. The results are shown in Table 2.

[0175] Table 2

[0176]

[0177] Experiments show that removing the SA-GRU module has the greatest impact on prediction accuracy, with the MAE error increasing significantly from 0.1177 to 0.2375, confirming that capturing long-range temporal dependencies is the key to ensuring basic accuracy. Removing the FE-TCN module also increases the MAE to 0.1891, demonstrating the necessity of this module for extracting high-frequency mutation features.

[0178] Furthermore, experiments with the asymmetric loss function show that when it is replaced with the standard MSE, although the resource shortage rate fluctuates less, the average deviation flips from a positive value (+0.0590) to a negative value (-0.011). This sign reversal indicates that the standard loss function tends to underestimate the load, while this invention successfully guides the model to build a positive prediction safety buffer through an asymmetric penalty mechanism, effectively avoiding the risk of resource under-allocation. In summary, the modules in the model work synergistically and complementarily, balancing high prediction accuracy with high business security.

[0179] In summary, the container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion described in this invention addresses the shortcomings of existing prediction models that neglect asymmetric risks by introducing an asymmetric loss function to construct a safety buffer at the algorithm level, effectively avoiding the service avalanche risk caused by resource under-allocation. Furthermore, it proposes a dual-channel spatiotemporal feature fusion architecture, combined with an end-to-end latency compensation strategy, achieving zero-latency response to sudden traffic surges while ensuring high prediction accuracy. Therefore, this invention effectively overcomes the shortcomings of existing container elastic scaling technologies, such as slow response and weak risk defense capabilities, and has high industrial application value.

[0180] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. A container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion, characterized in that, Includes the following steps: S1: Construct a real-time load data stream based on multi-order difference enhancement; collect historical load time-series data of container clusters, introduce time-series feature engineering methods, calculate the first-order difference representing the rate of load change and the second-order difference representing the acceleration of change, and combine adaptive hierarchical interpolation strategy to clean missing values, thereby constructing a time-series dataset that not only reflects the current load amplitude, but also represents the trend and intensity of load change. S2: Construct a load prediction model that fuses spatiotemporal features in two channels; construct a dual-channel parallel deep learning model that includes a local channel for extracting local mutation features and a global channel for global long-range temporal memory; the local channel uses a feature-enhanced temporal convolutional network combined with dilated causal convolution to extract high-frequency traffic spike features; the global channel uses a self-attention gated recurrent unit to capture long-cycle regular features; the two are adaptively fused through a dynamic gating mechanism. S3: Model training based on asymmetric weighted loss function; the processed dataset is divided into training and test sets, and a risk penalty coefficient is introduced during training to construct an asymmetric loss function; by increasing the gradient penalty weight for the underestimation of load region, the model is forced to establish a positive prediction safety buffer during gradient descent, thereby constructing an active defense mechanism against resource under-allocation risk at the algorithm level. S4: Construct an elastic scaling decision system based on prediction awareness; monitor the entire process time from container scheduling to service readiness in real time, and update the startup latency dynamic profile using an exponentially weighted moving average algorithm; combine the load prediction value output in step S2, use a multi-index parallel evaluation strategy to calculate the target number of container replicas, and determine the advance triggering time of the expansion operation based on the startup latency profile, so as to complete proactive resource preheating before the traffic peak arrives; S5: Performs proactive resource scheduling and anti-jitter processing; when the target number of replicas exceeds the current number of running replicas, it calls the cluster API to modify the replica field and performs seamless scaling in conjunction with the ready probe mechanism; at the same time, it introduces an asymmetric cooling window mechanism, which uses the time dimension hysteresis effect to filter instantaneous load jitter by setting a scaling down judgment threshold and a cooling lock period after scaling down execution, thus ensuring the high availability of the system.

2. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, Step S1 is as follows: S11: Collection of Multi-Dimensional Monitoring Indicators Deploy monitoring components in a Kubernetes cluster and configure high-frequency data collection cycles. Second; Raw indicator vectors collected at different times include: Basic resource metrics: CPU utilization Memory usage ; Business performance metrics: Requests per second (QPS), P90 response latency ; Network metrics: Network I / O rate ; S12: Missing value filling and anomaly cleaning An adaptive hierarchical strategy is adopted to meet real-time requirements; for short-term missing data, linear interpolation is used to fill in instantaneous packet loss. For long-term missing values, the previous value preservation method is used, and data confidence labels are set to prevent incorrect model guidance; for values ​​exceeding [a certain threshold], [further details are needed]. Outliers are considered monitoring noise and are smoothly replaced using a sliding window midpoint; where, This is the arithmetic mean of the load data within the current time sliding window. The standard deviation of the data within the sliding window; S13: Feature Derivation and Normalization Introducing temporal difference features to calculate observations First-order difference and second-order difference : Specifically, for the start time Since there is no historical data, the following settings are made: , as well as ; S14: Dimensional expansion through concatenation of eigenvectors Original collection of indicators for , dimension Calculate the first and second differences for each of the above indicators to obtain the velocity vector. and acceleration vector ; Constructed through vector concatenation operations Extended feature vector at time step : At this point, the feature dimension of the input data changes from... Expanded to 15 dimensions; Perform column-independent normalization on the expanded feature matrix; for the ... Each feature dimension ( The value is mapped to the [0, 1] interval using Max-Min normalization: All normalized elements The recombined features result in the final model input feature matrix. ; S15: Temporal Tensor Construction The sliding window method is used to construct the model input; the window size is set. The prediction step size H is the feature matrix constructed in step S14. Based on, among which For the total time step, For any condition satisfying sampling time The cut length is Constructing a local feature matrix from continuous feature fragments : in, Representation matrix In the The feature vectors of the row are then sampled by sliding along the time axis with a step size of 1, and multiple local feature matrices are combined. Stacking along the batch dimension, the final constructed temporal tensor is represented as follows: , where B is the batch size.

3. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, In step S2, The load prediction model based on dual-channel spatiotemporal feature fusion includes: a local mutation feature extraction module, a global long-range temporal memory module, and a dynamic gating fusion module; The load prediction model based on dual-channel spatiotemporal feature fusion takes as input the high-dimensional temporal tensor generated in step S1. ; The local mutation feature extraction module uses the feature-enhanced temporal convolutional network FE-TCN to capture the high-frequency fluctuation features of the load data; the global long-range temporal memory module uses the self-attention gated recurrent unit SA-GRU to capture the periodic patterns of the load data; the dynamic gated fusion module is responsible for adaptively weighting and fusing the two features, and finally outputting the load prediction value for future time.

4. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 3, characterized in that, The local mutation feature extraction module aims to extract local features reflecting millisecond-level flow spikes and instantaneous fluctuations from the input time-series tensor; by introducing a feature channel attention mechanism in the residual block, adaptive filtering of key load indicators is achieved. S211: Dilated Causal Convolution Processing Four layers of residual blocks are stacked, and the expansion coefficient of each layer increases exponentially. For the Layer, input sequence is Time step Convolution output at the point Calculated using the following formula: In the formula, The kernel size; The position index within the convolution kernel ( ), used to traverse the convolution window; For the first The expansion coefficient of the layer, For the first Layer position The learnable weight matrix at that location, For bias terms; S212: Feature Channel Attention Filtering To achieve adaptive feature selection, in the convolution output... Post-embedded attention module: Global information compression: Features are compressed along the time dimension using global average pooling, and the convolutional output is compressed. Mapped to channel descriptor ;No. The formula for calculating each channel is: In the formula, The time dimension length of the current layer feature map; This represents the total number of feature channels, i.e., the number of convolutional kernels. For channel index ( ); Indicates the first Layer Each channel at time step Eigenvalues ​​at; Weight generation: By capturing the non-linear dependencies between channels through two fully connected layers, a normalized weight vector is generated. : in It is the ReLU activation function. For the Sigmoid function; For the dimension reduction weight matrix, To reduce the dimensionality ratio, This is an upgraded weight matrix; Feature recalibration: The generated weight vector Treat it as a feature selector, and combine it with the convolution output. Perform channel-by-channel multiplication to obtain the weighted features. : Residual connections: To prevent network degradation and accelerate convergence, weighted features are used to perform residual connections. The original input of the residual block in this layer Perform element-wise addition and then output the final result after passing it through an activation function: After processing the stacked residual blocks, the output of the last layer is taken at the current time step. The feature vector, denoted as the local mutation feature vector. ( This serves as one of the inputs for the subsequent dual-channel fusion module.

5. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 4, characterized in that, The global long-range temporal memory module aims to capture the macroscopic tidal patterns of load data on a daily basis, thus solving the problem of long-sequence dependencies. This module works in parallel with the S21 module, processing the same input temporal tensor. ; S221: Gated Loop Unit Processing Using a gated recurrent unit (GRU) as the basic backbone, for time steps The input is The hidden state at the previous moment was The specific calculation process is as follows: Reset door Used to control the degree of historical information discarded; the weight matrix is ​​as follows: The calculation formula is: Update Gate : Used to control the degree of retention of historical information, the weight matrix is ​​as follows The calculation formula is: Hidden state update: The weight matrix is Based on the above gating results, candidate hidden states are calculated. With the final hidden state : Through the above recursive calculation, a hidden state sequence containing long-range dependency information is obtained; S222: Masked Multi-Attention Enhancement A multi-head self-attention mechanism is introduced onto the GRU output sequence to directly establish the dependency between the current time step and long historical time steps: in, These represent the query vector, key vector, and value vector, respectively, all of which are generated by linear mapping from the hidden state sequence output by the GRU. This is a scaling factor used to adjust the numerical range of the dot product result to prevent gradient vanishing. It is an upper triangular mask matrix; The final output is a global feature vector containing long-range dependency information. , ,in It serves as the second input for the subsequent dual-channel fusion module.

6. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 5, characterized in that, The dynamic gating fusion module receives the local mutation feature vector from S21. and global feature vectors from S22 An adaptive gating mechanism is used to dynamically adjust the contribution ratio of the two features, and finally the load prediction value is generated. S231: Generation of gating coefficients First, the two feature vectors are concatenated along the channel dimension to construct a joint feature vector. ; Subsequently, the joint features are mapped to scalar gating coefficients using a multilayer perceptron. This process includes nonlinear transformations to enhance expressive power; the calculation formula is as follows: In the formula, It is a ReLU activation function; Use the Sigmoid activation function to ensure the output coefficients ; and The weight matrix is ​​a learnable matrix; For bias terms; S232: Adaptive Feature Fusion Using the generated gating coefficients As a regulating factor, the two feature paths are soft-switched and weighted, resulting in a fused feature vector. The calculation is as follows: S233: Load Prediction Output Finally, the fused feature vectors The input is fed into a linear regression layer, which maps the high-dimensional feature space back to the target payload space, and outputs the prediction results for future time moments. : In the formula, and Weights and biases for the output layer; prediction results It includes predicted values ​​for key metrics such as CPU utilization, memory usage, and QPS, which are used for subsequent elastic scaling decisions.

7. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, In step S3, the asymmetric loss function The definition is as follows: in Indicates the size of the training batch; Represents the actual load observation value. This represents the model's predicted output value; This is an indicator function that takes the value 1 when the condition inside the parentheses is met, and 0 otherwise; This represents the risk penalty coefficient.

8. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, In step S3, a collaborative configuration optimization algorithm, learning rate scheduling mechanism, and model early stopping criterion are used to achieve rapid fitting and global optimization of complex non-stationary objective functions. Specifically, this includes the following: Parameter optimization uses the Adam optimizer for parameter updates; the first moment is set to estimate the exponential decay rate. Second-order moment estimation of exponential decay rate To accommodate sparse gradients and non-stationary objective functions; Learning rate scheduling introduces a cosine annealing strategy to escape local optima; initial learning rate Minimum learning rate Learning rate With the current Epoch The formula for change is: Monitor the asymmetric loss value on the validation set. If the validation set loss does not decrease for 10 consecutive epochs, terminate training early and save the current optimal weights to prevent overfitting.

9. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, In step S4, The target number of replicas is calculated as follows: After obtaining the load forecast results, CPU utilization, memory usage, and QPS—three metrics directly related to container physical resource quotas—are extracted from the forecast results as baseline dimensions. For each load dimension... According to the model's prediction of the future Time-of-flight load value Based on the current resource utilization rate and the corresponding target threshold Calculate the required number of replicas respectively. To eliminate the uncertainty risk of model predictions, a safety redundancy coefficient is introduced. The deterministic resource buffer pool is constructed using the following formula: In the formula, This represents the number of currently running replicas. This represents the floor function; Final target number of copies Take the maximum value of the results calculated for all dimensions, that is: End-to-end latency compensation decision: To achieve proactive resource preheating and minimize response latency, the system must calculate the advance trigger time window. This is to offset the physical time required to start up resources; The calculation formula is as follows: In the formula The cold start time of a Pod is not fixed. The Prometheus open-source monitoring and alerting system continuously monitors the actual time taken for a Pod to go from scheduling to the ready state, and dynamically updates it using an exponentially weighted moving average algorithm. The calculation formula is as follows: in, For the current moment Updated startup latency estimate; This represents the actual physical time taken for a Pod to go from scheduling to readiness, as monitored at the current moment. For the previous moment The estimated startup latency; As a smoothing factor, it is used to balance the weights of historical trends and current observations to adapt to fluctuations in mirror pull speed or changes in node load; This serves as a system safety buffer time to cover Kubernetes API scheduling latency and service registration and discovery synchronization delays. Based on the aforementioned time window, the system selects the sequence with the closest time span from the future load prediction sequence output in step S3. Prediction points As a basis for decision-making; if the predicted load at that moment exceeds the current capacity, then immediately at that moment... Trigger expansion command.

10. The container elastic scaling method based on asymmetric risk perception and dual-channel spatiotemporal feature fusion as described in claim 1, characterized in that, In step S5, S51: Seamless Capacity Expansion Execution The final target replica count output by the system real-time monitoring step S41 ;when Greater than the number of currently running replicas Upon that, the system immediately triggers an expansion command; To prevent resource exhaustion due to extreme prediction bias, a maximum replica limit is set for the cluster. The actual number of expansion target replicas issued Take the smaller of the calculated value and the upper limit value, that is: S52: Volume Reduction Hysteresis Cooling Strategy To prevent frequent Pod creation and destruction caused by slight load fluctuations around the threshold, an asymmetric cooling window mechanism is introduced: For scaling up operations, the cooldown time is set to 0 seconds, meaning that scaling up commands have the highest priority and are executed immediately once a risk of insufficient resources is detected, thus ensuring high service availability; for scaling down operations, the scaling down threshold is set to 30%; a scaling down command is only triggered when the model predicts that the future load value is consistently lower than this threshold. After execution, the system enters a 300-second lockout state, during which new scaling down commands are prohibited, but scaling up commands are allowed to preemptively execute.

Images