Editorial Director, Technical & Educational Resources
Beginning with v1.27, the open-source Netdata Agent is capable of unsupervised anomaly detection with machine learning (ML). As with all things Netdata, the anomalies collector comes with preconfigured alarms and instant visualizations that require no query languages or organizing metrics. You configure the collector to look at specific charts, and it handles the rest.
Netdata's implementation uses a handful of functions in the Python Outlier Detection (PyOD)
library, which periodically runs a
train function that learns what
"normal" looks like on your node and creates an ML model for each chart, then utilizes the
predict() PyOD functions to
quantify how anomalous certain charts are.
All these metrics and alarms are available for centralized monitoring in Netdata Cloud. If you choose to sign up for Netdata Cloud and claim your nodes, you will have the ability to run tailored anomaly detection on every node in your infrastructure, regardless of its purpose or workload.
In this guide, you'll learn how to set up the anomalies collector to instantly detect anomalies in an Nginx web server and/or the node that hosts it, which will give you the tools to configure parallel unsupervised monitors for any application in your infrastructure. Let's get started.
- A node running the Netdata Agent. If you don't yet have that, get Netdata.
- A Netdata Cloud account. Sign up if you don't have one already.
- Familiarity with configuring the Netdata Agent with
- Optional: An Nginx web server running on the same node to follow the example configuration steps.
The anomalies collector uses a few Python packages, available with
pip3, to run ML training. It requires
pyod, in addition to
netdata-pandas, which is a package built by the Netdata team to pull data
from a Netdata Agent's API into a Pandas. Read more about
netdata-pandas on its package
repo or in Netdata's community
pip3command fails, you need to install it. For example, on an Ubuntu system, use
sudo apt install python3-pip.
exit to become your normal user again.
Navigate to your Netdata config directory and use
to open the
python.d.conf file, search for the
anomalies line. If the line exists, set the value to
yes. Add the line
yourself if it doesn't already exist. Either way, the final result should look like:
Restart the Agent with
sudo systemctl restart netdata to start up the
anomalies collector. By default, the model training process runs every 30 minutes, and uses the previous 4 hours of
metrics to establish a baseline for health and performance across the default included charts.
💡 The anomaly collector may need 30-60 seconds to finish its initial training and have enough data to start generating anomaly scores. You may need to refresh your browser tab for the Anomalies section to appear in menus on both the local Agent dashboard or Netdata Cloud.
The file contains many user-configurable settings with sane defaults. Here are some important settings that don't involve tweaking the behavior of the ML training itself.
charts_regex: Which charts to train models for and run anomaly detection on, with each chart getting a separate model.
charts_to_exclude: Specific charts, selected by the regex in
charts_regex, to exclude.
train_every_n: How often to train the ML models.
train_n_secs: The number of historical observations to train each model on. The default is 4 hours, but if your node doesn't have historical metrics going back that far, consider changing the metrics retention policy or reducing this window.
custom_models: A way to define custom models that you want anomaly probabilities for, including multi-node or streaming setups. More on custom models in part 3 of this guide series.
charts_regexwith many charts or
train_n_secsto a very large number will have an impact on the resources and time required to train a model for every chart. The actual performance implications depend on the resources available on your node. If you plan on changing these settings beyond the default, or what's mentioned in this guide, make incremental changes to observe the performance impact. Considering
train_max_nto cap the number of observations actually used to train on.
charts_regex allows for some basic regex, such as wildcards (
*) to match all contexts with a certain pattern. For
system\..* matches with any chart wit ha context that begins with
system., and ends in any number of other
.*). Note the escape character (
\) around the first period to capture a period character exactly, and
not any character.
anomalies.conf to the following:
This value tells the anomaly collector to train against every
system. chart, every
nginx_local chart, every
web_log_nginx chart, and specifically the
As you can see in the above screenshot, this node is now looking for anomalies in many places. The result is a single
anomalies_local.probability chart with more than twenty dimensions, some of which the dashboard hides at the bottom of
a scroll-able area. In addition, training and analyzing the anomaly collector on many charts might require more CPU
utilization that you're willing to give.
First, explicitly declare which
system. charts to monitor rather than of all of them using regex (
Next, remove some charts with the
charts_to_exclude setting. For this example, using an Nginx web server, focus on the
volume of requests/responses, not, for example, which type of 4xx response a user might receive.
Now that you know how to set up unsupervised anomaly detection in the Netdata Agent, using an Nginx web server as an example, it's time to apply that knowledge to other mission-critical parts of your infrastructure. If you're not sure what to monitor next, check out our list of collectors to see what kind of metrics Netdata can collect from your systems, containers, and applications.
For a more user-friendly anomaly detection experience, try out the Metric Correlations feature in Netdata Cloud. Metric Correlations runs only at your requests, removing unrelated charts from the dashboard to help you focus on root cause analysis.
Stay tuned for the next two parts of this guide, which provide more real-world context for the anomalies collector. First, maximize the immediate value you get from anomaly detection by tracking preconfigured alarms, visualizing anomalies in charts, and building a new dashboard tailored to your applications. Then, learn about creating custom ML models, which help you holistically monitor an application or service by monitoring anomalies across a cluster of charts.