TLDR Tech / DevOps
Scale before the spike: Predictive autoscaling for GPU workloads on Kubernetes
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.
Read the full story