Cloud Native Computing Foundation PrometheusDay NA 2022, 28 Oct 2022

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Lightning Talk: Real Time Prediction for Autoscaling with Prometheus Metrics - Ravi Hari, Intuit

Description

Don’t miss out! Join us at our upcoming event: KubeCon + CloudNativeCon Europe 2023 in Amsterdam, The Netherlands from April 17-21. Learn more at https://kubecon.io. The conference features presentations from developers and end users of Kubernetes, Prometheus, Envoy, and all of the other CNCF-hosted projects.

Lightning Talk: Real Time Prediction for Autoscaling with Prometheus Metrics - Ravi Hari, Intuit

Prometheus is a defacto standard for monitoring in kubernetes. Prometheus metrics are consumed by multiple addons in Kubernetes for monitoring, alerting, autoscaling etc., This talk will walkthrough how the prometheus metrics can be used for real time prediction so the application pods can be scaled horizontally or vertically. Getting the right sizing of pods is an important task especially when running kubernetes in cloud environments so we can optimise the cost. This talk will walk through the framework used for predicting the right sizing of the pods based on prometheus metric and how we can leverage a custom pod-autoscaler to scale the pods dynamically.

A

The topic is real-time prediction for auto scaling using Primitives metrics.

A

Briefly, we are interviewed started in 1983 and we have about 100 million customers, and we have found many you can subscribed.

A

Prometheus is a Defector monitoring tool in kubernetes and Prometheus. Metrics are consumed for various use cases, for example, monitoring, alerting, anomaly, detection and auto scaling Etc time series forecasting can be leveraged to make predictions, and in order to make a good prediction, uh we can follow some rules uh that we can collect. Aggregated data to avoid noise in the data and the data used for forecasting should be updated frequently and the Horizon in which we predict this data.

A

It should be over a short duration to make this predictions uh more accurate.

A

Some of the time series prediction, algorithms or models that are already existing today- are some Auto regressive models like arima and, if you add seasonality into it, this is called sarima and uh some of the other two other models like lstm, exponential smoking and other things also available. A couple of them are also uh contributed by some popular companies that Facebook and Google into tools like profit and temporal.

A

So we can see how we can use this time series data, so this is the data our namespace have continuous GP utilization that we want to create predictions for and then see how we can use that in a custom, auto scaler.

A

So the data that we have is given here and let's try to use that with the arima. Let's make a quick run of arima model here and plot the graph you can see that graph is given here. The red line is actual data. uh The trend actually is replicated with the arima model.

A

However, there is slight delay in here now: let's try with the Facebook profit, let's execute all these things, so here we are using the same data and we are creating the chart and then we are creating a model by considering some seasonality into this, and then here is the data with the profit to rent. Here a blue line is that is actually the predicted data that you can see and data lines, as you can see up to here.

A

These are the ones that's actual data, and you see there is a shade of low here, upper and above this blue line. That indicates there is uncertainty in this data and it gives a lower burn and upper Bound for these values. Now, let's use this information and see if we can and predict the auto scaling on severe utilization right. So so we are getting this uh data points. Let's now try to see. So this is a deployment in which there are five replicas and based on the prediction of the city usage.

A

We want to reduce our increase, our replica, so that is a controller already running here. Let me go ahead and submit a CR uh to briefly go over this year. This has a namespace and deployment on which we want to act upon and lower point and upper one CPU. If the value is in between these values, it doesn't do anything what if the value is above or lesser than the values that we are giving it it will either add or reduce the number of replicas right.

A

So looking up like this, so I created this one and it has acted upon this and we are watching the deployment type here so earlier. It was like uh five uh replicas now that uh there's reduced by one, because the prediction for this is 18 because and the threshold we have is uh 20 if it is less than 20. If you go ahead and reduce it.

A

uh Here is a simple report for that, and uh uh here we are calling the prediction and prediction is nothing but a service of the perfect thing that we have deployed in the cluster and we are calling that service and getting the value recommendation and if it is uh uh in between the boundaries, it doesn't do anything if it is a lesser than lower power, it will reduce by one correspondingly values.

A

uh So so we can leverage these things and then optimize our resources and reduce costs uh for the kubernetes Clusters. If you're running a provided right. uh That is a demo that what it issue. Thank you.