Hi,Community: I wonder if there any plan to develop flow model based autoscaling feature? A possible Flow model based autoscaling mechanism may work as follow steps: 1.take flow metrics samples from operators,maybe other features. 2.infer operators to be tuned and associated operator parallelism based on pre-trained operator flow analysis model??maybe algorithms like holt-winters is an option?? 3.autoscale target operator with inferred operator parallelism. The key points I consider to implement sunch an mechanism lies in three aspects,one is at what time point should we start the infer action and what precision can be achieved at the inferring time point,the second is how to accumulate/or update the online flow metrics to construct real time features and thus achieving real time model interation,the third is how to deal with sudden drastic flow fluctuations caused by manual operation like new flow data source introduced in. may be these three aspects can all be solved just by using real time in-box model training and predicting,a self contained way. Thanks.Any replies and discussion wiil be wellcomed.
