MongoDB's engineering team describes a 2023 research experiment that led to the production rollout of predictive auto-scaling in MongoDB Atlas. The prototype used three components: a Long-Term Forecaster (MSTL decomposition for seasonal/trend patterns), a Short-Term Forecaster (trend interpolation for non-seasonal workloads), and an Estimator (boosted decision tree regression trained on 25 million samples) to predict CPU utilization across instance tiers. A Planner then selects the cheapest instance size to handle forecasted demand. The experiment showed the predictive scaler would have saved an average of 9 cents per hour per replica set versus the reactive scaler. The production version, rolled out in November 2025, conservatively scales replica sets up before predicted load spikes while relying on the existing reactive algorithm to scale down afterward.
Table of contents
MongoDB AtlasPredictive auto-scalingPredictive scaling experimentPredictive scaling: PlannerPredictive scaling: Long-term forecasterPredictive scaling: Short-term forecasterPredictive scaling: estimatorPredictive scaling: Putting it all togetherPredictive auto-scaling in productionSort: