Attention-Enhanced Multivariate Forecasting for Intelligent Microservice Autoscaling

Proactive autoscaling in cloud-native microservices requires forecasts that are accurate enough to support reliable scaling decisions rather than merely low regression error. This study evaluates a leakage-aware forecasting-to-decision pipeline for CPU and memory utilization using a retained Alibaba microservice trace subset containing 48 service instances and 47 nodes. The experimental design applies chronological per-service train-validation-test splits, train-only preprocessing, one-step multivariate forecasting, fixed-backbone attention ablation, and NoPred/PredOnly/WithPred feature-set attribution for downstream action classification. The forecasting stage compares recurrent baselines, a vanilla Transformer Encoder, and BiLSTM encoder-decoder variants with Bahdanau, Luong, and multi-head attention. The best configuration, a BiLSTM encoder-decoder with Bahdanau attention and residual connection, achieves the lowest joint CPU-memory test RMSE of 0.017188, improving over the BiLSTM baseline (0.020649) and Transformer Encoder (0.019534). Target-specific results show CPU RMSE of 0.028924 and memory RMSE of 0.005452, while burst analysis indicates lower spike-segment RMSE and MAE despite stronger Transformer behavior on selected peak-alignment indicators. In the decision stage, tuned XGBoost with forecast-augmented features obtains 0.950602 accuracy and 0.951026 macro F1, outperforming NoPred and PredOnly settings. A read-only downstream simulation further reduces service-level objective violation compared with a horizontal autoscaling policy under Normal and Stress×3 scenarios while maintaining zero simulated downtime under the in-place update assumption. Overall, the findings indicate that forecast-derived signals can improve proactive autoscaling decisions when evaluated through strict temporal controls, explicit attribution, and bounded simulation.