An insight, as mentioned earlier, is knowledge gained from analyzing information. The goal of gathering data is to derive insights. How do you get insights from your data? In two steps:
Step 1. You need to build a knowledge base. Knowing what is expected behavior and what is erroneous behavior is critical when it comes to understanding data. This is where you, the expert, play a role. You define what is good and what is bad. In addition, new techniques can use the data itself to map what is expected behavior and what is not. We touch on these techniques later in this section.
Step 2. You apply the knowledge base to the data—hopefully automatically, as this is a network automation book.
As an example of this two-step process, say that you define that if a device is running at 99% CPU utilization, this is a bad sign (step 1). By monitoring your devices and gathering CPU utilization data, you identify that a device is running at 100% utilization (step 2) and replace it. The data insight here is that a device was not working as expected.
Raising an alarm is an action on the insight (in this case, the insight that it’s necessary to raise the alarm). You can trigger alarms based on different types of data, such as log data or metrics.
You can use multiple tools to generate alarms (refer to Chapter 1); the important parameter is what to alarm on. For example, say that you are gathering metrics from your network using telemetry, and among these metrics is memory utilization. What value for memory utilization should raise an alarm? There is no universal answer; it depends on the type of system and the load it processes. For some systems, running on 90% memory utilization is common; for others, 90% would indicate a problem. Defining the alarm value is referred to as defining a threshold. Most of the time, you use your experience to come up with a value.
You can also define a threshold without having to choose a critical value. This technique, called baselining, determines expected behavior based on historical data.
A very simple baselining technique could be using the average from a specific time frame. However, there are also very complex techniques, such as using neural networks. Some tools (for example, Cisco’s DNA Center) have incorporated baselining modules that help you set up thresholds.
If you are already using Grafana, creating an alarm is very simple. By editing the Figure 3-4 dashboard as shown in Figure 3-8, you can define the alarm metric. You can use simple metrics such as going over a determined value, or calculated metrics such as the average over a predefined time. Figure 3-8 shows a monitoring graph of disk I/O operations on several services; it is set to alarm if the value reaches 55 MBs.
Figure 3-8 Setting Up a Grafana Alarm for 55 MBs Disk I/O
Tools like Kibana also allow you to set up alerts based on log data. For example, Figure 3-9 shows the setup of an alarm based on receiving more than 75 Syslog messages from a database with the error severity in the past 5 minutes.
Figure 3-9 Setting Up a Kibana Alarm for Syslog Data
In addition to acting as alerts for humans, alarms can be automation triggers; that is, they can trigger automated preventive or corrective actions. Consider the earlier scenario of an abnormally high CPU utilization percentage. In this case, a possible alarm could be a webhook that triggers an Ansible playbook to clean up known CPU-consuming processes.
Figure 3-10 illustrates an attacker sending CPU punted packets to a router, which leads to a spike in CPU usage. This router is configured to send telemetry data to Telegraf, which stores it in an InfluxDB database. As Grafana ingests this data, it notices the unusual metric and triggers a configured alarm that uses an Ansible Tower webhook to run a playbook and configure an ACL that will drop packets from the attacker, mitigating the effects of the attack.
Figure 3-10 Automated Resolution Triggered by Webhook
Typically, deploying a combination of automated action and human notification is the best choice. Multiple targets are possible when you set up an alarm to alert humans, such as email, SMS, dashboards, and chatbots (for example, Slack or Webex Teams).
In summary, you can use an alarm to trigger actions based on a state or an event. These actions can be automated or can require manual intervention. The way you set up alarms and the actions available on alarms are tightly coupled with the tools you use in your environment. Grafana and Kibana are widely used in the industry, and others are available as well, including as Splunk, SolarWinds, and Cisco’s DNA Center.
In Chapter 1, we touched on the topic of configuration drift. If you have worked in networking, you know it happens. Very few companies have enough controls in place to completely prevent it, especially if they have long-running networks that have been in place more than 5 years. So, how do you address drift?
You can monitor device configurations and compare them to known good ones (templates). This tells you whether configuration drift has occurred. If it has, you can either replace the device configuration with the template or the other way around. The decision you make depends on what is changed.
You can apply templates manually, but if your network has a couple hundred or even thousands of devices, it will quickly become a burden. Ansible can help with this task, as shown in Example 3-8.
Example 3-8 Using Ansible to Identify Configuration Differences in a Switch
Example 3-8 shows a template file (template.txt) with the expected configuration, which is the same configuration initially used on the switch. The example shows how you can connect automatically to the host to verify whether its configuration matches the provided template. If it does not, you see the differences in the output (indicated with the + or – sign). The example shows that, on the second execution, after the template is modified, the output shows the differences.
You can also achieve configuration compliance checking by using the same type of tool and logic but, instead of checking for differences against a template, checking for deviations from the compliance standards (for example, using SHA-512 instead of SHA-256).
Insights can come from artificial intelligent (AI) or machine learning (ML) techniques. AI and ML have been applied extensively in the past few years in many contexts (for example, for financial fraud detection, image recognition, and natural language processing). They can also play a role in networking.
ML involves constructing algorithms and models that can learn to make decisions/predictions directly from data without following predefined rules. Currently, ML algorithms can be divided into three major categories: supervised, unsupervised, and reinforcement learning. Here we focus on the first two categories because reinforcement learning is about training agents to take actions, and that type of ML is very rarely applied to networking use cases.
Supervised algorithms are typically used for classification and regression tasks, based on labeled data (where labels are the expected result). Classification involves predicting a result from a predefined set (for example, predicting malicious out of the possibilities malicious or benign). Regression involves predicting a value (for example, predicting how many battery cycles a router will live through).
Unsupervised algorithms try to group data into related clusters. For example, given a set of NetFlow log data, grouping it with the intent of trying to identify malicious flows would be considered unsupervised learning.
On the other hand, given a set of NetFlow log data where you have previously identified which subset of flows is malicious and using it to predict whether if future flows are malicious would be considered supervised learning.
One type is not better than the other; supervised and unsupervised learning aims to address different types of situations.
There are three major ways you can use machine learning:
Training your own models with automated machine learning (AutoML)
Training your models manually
Before you learn about each of these ways, you need to understand the steps involved in training and using a model:
Step 1. Define the problem (for example, prediction, regression, clustering). You need to define the problem you are trying to solve, the data that you need to solve it, and possible algorithms to use.
Step 2. Gather data (for example, API, Syslog, telemetry). Typically you need a lot of data, and gathering data can take weeks or months.
Step 3. Prepare the data (for example, parsing, aggregation). There is a whole discipline called data engineering that tries to accomplish the most with data in a machine learning context.
Step 4. Train the model. This is often a trial-and-error adventure, involving different algorithms and architectures.
Step 5. Test the model. The resulting models need to be tested with sample data sets to see how they perform at predicting the problem you defined previously.
Step 6. Deploy/use the model. Depending on the results of testing, you might have to repeat steps 4 and 5; when the results are satisfactory, you can finally deploy and use the model.
These steps can take a long time to work through, and the process can be expensive. The process also seems complicated, doesn’t it? What if someone already addressed the problem you are trying to solve? These is where pretrained models come in. You find these models inside products and in cloud marketplaces. These models have been trained and are ready to be used so you can skip directly to step 6. The major inconvenience with using pretrained models is that if your data is very different from what the model was trained with, the results will not be ideal. For example, a model for detecting whether people were wearing face masks was trained with images at 4K resolution. When predictions were made with CCTV footage (at very low resolution), there was a high false positive rate. Although it may provide inaccurate results, using pretrained models is typically the quickest way to get machine learning insights in a network.
If you are curious and want to try a pretrained model, check out Amazon Rekognition, which is a service that identifies objects within images. Another example is Amazon Monitron, a service with which you must install Amazon-provided sensors that analyze industrial machinery behavior and alert you when they detect anomalies.
AutoML goes a step beyond pretrained models. There are tools available that allow you to train your own model, using your own data but with minimal machine learning knowledge. You need to provide minimal inputs. You typically have to provide the data that you want to use and the problem you are trying to solve. AutoML tools prepare the data as they see fit and train several models. These tools present the best-performing models to you, and you can then use them to make predictions.
With AutoML, you skip steps 3 through 5, which is where the most AI/ML knowledge is required. AutoML is commonly available from cloud providers.
Finally, you can train models in-house. This option requires more knowledge than the other options. Python is the tool most commonly used for training models in-house. When you choose this option, all the steps apply.
An interesting use case of machine learning that may spark your imagination is in regard to log data. When something goes wrong—for example, an access switch reboots—you may receive many different logs from many different components, from routing neighborships going down to applications losing connectivity and so on. However, the real problem is that a switch is not active. Machine learning can detect that many of those logs are consequences of a problem and not the actual problem, and it can group them accordingly. This is part of a new discipline called AIOps (artificial intelligence operations). A tool not mentioned in Chapter 1 that tries to achieve AIOps is Moogsoft.
Here are a couple examples of machine learning applications for networks:
Malicious traffic classification based on k-means clustering.
Interface bandwidth saturation forecast based on recurrent neural networks.
Log correlation based on natural language processing.
Traffic steering based on time-series forecasting.
The code required to implement machine learning models is relatively simple, and we don’t cover it in this book. The majority of the complexity is in gathering data and parsing it. To see how simple using machine learning can be, assume that data has already been gathered and transformed and examine Example 3-9, where x is your data, and y is your labels (that is, the expected result). The first three lines create a linear regression model and train it to fit your data. The fourth line makes predictions. You can pass a value (new_value, in the same format as the training data, x), and the module will try to predict its label.