Oct 14, 2026

Scalable Metrics Anomaly Detection and Forecasting

8views

Scalable Metrics Anomaly Detection and Forecasting

Discover how intelligent metric correlation and cohort-based machine learning make regression modeling and anomaly detection scalable for millions of network device performance metrics.

Scalable Metrics Anomaly Detection and Forecasting

Discover how intelligent metric correlation and cohort-based machine learning make regression modeling and anomaly detection scalable for millions of network device performance metrics.

Chapters

00:00

Introduction

Introduction to analyzing large volumes of time series performance data

00:21

Trend and Anomaly Detection

01:00

Intelligent Cohorting

01:26

Device Analysis Example

02:27

Scaling Performance Analytics

Transcript

00:00

In this video, you'll see how we analyze large volumes of time series performance data from many different network devices.

00:06

Using intelligent metric correlation and cohort based machine learning.

00:10

We make regression modeling and anomaly detection scalable.

00:14

In this example, we're analyzing around 440 network devices with 32 metrics per device.

00:22

First, we detect metric trends.

00:24

Here you can see devices with increasing trends in data receive rate, write latency and transmit rate.

00:31

Also identify devices with decreasing metric trends including data receive rate, write latency and transmit rate.

00:39

Next, we identify anomalies in the performance data including network data, transmit rate, receive rate and virtual disk write IUPs.

00:48

This view reveals a significant number of anomalies.

00:51

We also show overall accuracy with CPU utilization achieving the highest accuracy at 81% alongside accuracy for other metrics.

01:00

Intelligent cohorting is our key differentiator.

01:03

instead of running separate regressions across hundreds of thousands of metrics, we intelligently group correlated metrics into cohorts shown here as correlation groups.

01:13

For example, 100,000 metric streams may require only 10,000 regression models.

01:20

This dramatically reduces computation while preserving analysis at the individual device level.

01:26

Let's take a closer look at NetApp storage device CPU utilization localization.

01:37

Here, the baseline captures normal workload patterns and seasonality.

01:41

A recurring CPU spike is not necessarily an anomaly, so we compare actual measurements against the learned baseline.

01:49

The red dots represent anomalies while the chart displays the actual data along with the upper and lower bounds.

01:56

Overall CPU utilization shows a decreasing trend.

01:59

We also provide forecasting.

02:01

Everything to the right of the yellow bar represents forecasted and predicted future data.

02:06

You can also view technical details including accuracy, percentage, seasonality and anomaly percentage.

02:13

Additional details include the training duration, which is seven days, and a future prediction period of seven days.

02:20

The overall trend is decreasing.

02:22

There are eight high anomalies, 11 medium anomalies and seven low anomalies.

02:27

Now let's look at another example.

02:30

Here we have nfsiops, which shows an increasing trend.

02:35

You can also see anomalies and predicted future data.

02:39

The same intelligence works across cpu, memory, network and storage metrics, automatically adapting to different workload patterns.

02:47

That's the power of this approach.

02:49

You can analyze millions of performance metrics without running millions of machine learning models.

02:54

By intelligently grouping correlated metrics, we dramatically reduce computing cost and processing time while still delivering accurate baselines.

03:03

Anomaly detection and forecasting for every device.

03:07

This is how we make AIPowered performance analytics practical at massive scale.

Scalable Metrics Anomaly Detection and Forecasting — Tella