SCALE BEFORE THE SPIKE: PREDICTIVE AUTOSCALING FOR GPU WORKLOADS ON KUBERNETES

Summary from the TLDR DevOps newsletter

Two Adobe platform engineers built a predictive Kubernetes autoscaler after a production incident left users seeing 15-20% error rates because reactive scaling fired too late to help. The system uses a Bi-LSTM model embedded in a Go controller binary to forecast demand 10 minutes ahead, running inference every 60 seconds against the past hour of Prometheus metrics. A graduated scaler caps scale-out at 20 pods per minute to avoid overwhelming the scheduler, and a burst detector adds a heuristic safety net for traffic spikes the model wasn't trained on. In hackathon validation, 23 out of 23 checks passed, with the predictor catching simulated spike patterns 11 minutes early.