This post has been republished via RSS; it originally appeared at: Azure Data Explorer Blog articles.
Azure Data Explorer offers customers 3 ways to manage their cluster's instances count:
While in the Manual Scale/Custom Autoscale the customer should be involved in controlling the cluster size, the Optimized Autoscale is being managed by the Azure Data Explorer service and thus provide a hassle free solution for auto scaling the cluster - the customer just needs to set up lower and upper bounds for the cluster instances count and ADX will optimize its instances count between these bounds. The Optimized Autoscale usually achieves better cost optimization than the other methods while keeping the cluster performance at its best.
Until now the Optimized Autoscale was implemented using a Reactive model with smart rules on a variety of metrics, it is called “Reactive” since it acts in reaction to historical data, for example, if the CPU load was high during the last hour, it scales out.
We would like to introduce to you a significant improvement to the Optimized Autoscale solution that is currently in private preview, where we are using a Predictive model that tackles few issues of the Reactive model to perform better.
One of the major constraints of the solution is not scale the cluster frequently because after scaling a cluster, a process of data reorganization starts, it last up to 2 hours, and it is impairs the clusters performance. This process optimize the cached data on all nodes and it impairs the clusters performance by increasing the CPU load and by causing a situation where queries data might not yet be in the clusters cache. Taking into consideration the data reorganization process, we would like optimally that:
- Scale-out will occur 2 hours ahead of load increase - to finish the data reorganization before the load increase.
- Scale-in will occur right after the load decrease - to optimize cost.
The Reactive model works great when there is a non cyclic usage, for example, when new workloads that significantly increases the CPU usage are added, it scales-out the cluster to withstand the higher load, and when workloads are removed, it scales-in the cluster to save cost, however it is always acts in reaction to historical data and does not account for cyclic usage where the resource demand follows a pattern, thus, can't follow the optimal rules above.
Let us look on the accumulated CPU metric of a typical cluster over 3 weeks:
We noticed that in many clusters the metric follows a seasonal pattern - might be weekly (less load on the weekend), daily (night vs daily business hours) or other seasonality's due to recurring automatic workflows. So, we leveraged series_decompose_forecast, ADX native forecasting function, to predict the next day metrics. Based on this prediction, we schedule scale operations for the next day, considering the data reorganization process.
Continuing with the above example, let’s review the forecast of its CPU:
In the blue line, you can see the actual accumulated CPU, and in the orange line its forecasting. To the right, you can see that the actual CPU is flat for the future, but the prediction follows the same historical pattern.
Let us zoom in on our forecast (the orange line above appears hear in blue) for January 30th:
When we identify increase in cluster load, in the time stamp marked by the red line, we look for the maximum load in subsequent hours, which is marked by the yellow line, and schedule a scale-out operation to the number of instances that will handle this maximum-load. Doing so, we avoid consecutive scale-out operations. Also, we schedule the operations to occur 2 hours before the load-increase, as marked by the green line, so the cluster will be ready for the load increase after it has already finished the data reorganization.
When we identify decrease in cluster load, like in the red line, we avoid consecutive scale-in/scale-out operations by:
- Scheduling the scale-in operation only if we see that the low load will last long enough, as seen by the yellow arrowed segment
- Determining the target instances count of the scale-in operation by the maximum load of subsequent hours, which is marked by the orange line.
Furthermore, we avoid consecutive scale-in operations by scheduling only one scale-in operation for the consecutive hours. Note that we schedule the scale-in operation to occur right on time when the load drops, as marked by the green line (compared to the Reactive model, that acts in about 3 hours delay).
The Predictive model might not always forecast accurately (e.g. when the CPU does not have a pattern or when there is a change of the CPU pattern), so we created a general hybrid solution that integrates both the Reactive model and the Predictive model in the following way:
For each cluster:
- Every 12 hours, we forecast the metrics for the next day based on 3 weeks data.
- We add a confidence measurement to the forecast.
- Based on the forecast, we plan scale operations for the next 12 hours.
- For each of the clusters whose confidence is above a threshold, we run a light validation function to verify the forecast still holds.
- The clusters that have high confidence and passed the validation will be scaled by the Predictive model scheduled operations, while the rest will fall back to the Reactive model until next forecast.
Here is an example cluster manifesting the improvement when changing from the Reactive model to the Predictive model:
In the blue line, you can see the cluster CPU, and in the red line the instances count of the cluster.
- To the left of the green line, the Reactive Autoscale was active - the instances count was on average 25 machines and the scale operations are not aligned with the data because of the delayed response.
- To the right of the green line, the Predictive Autoscale is active - most of the time the cluster has 10 instances and only for predicted times of high loads it scales out to 25 instances. Note that the scale operations are aligned with the data - scale-out occurs around 2 hours before the load increase, making sure the cluster is ready for the high load ahead of time, and scale-in takes place just in time when the high load ends.
Overall, in this case the new Predictive Autoscale saved about 50% of the cluster cost while even improving the performance compared to the Reactive model.
To summarize, ADX built a new innovative Predictive Autoscale model, based on ML and Time Series Analysis, that guarantees the best performance while optimizing cluster cost. This Predictive Autoscale is currently in a private preview, to be generally available in the next few months. Meanwhile, you are welcome to open a support request to join this preview and start saving $$$.