Log anomaly detection - Amazon CloudWatch Logs

Log anomaly detection

You can create a log anomaly detector for each log group. The anomaly detector scans the log events ingested into the log group and find anomalies in the log data. Anomaly detection uses machine-learning and pattern recognition to establish baselines of typical log content.

After you create an anomaly detector for a log group, it trains using the past two weeks of log events in the log group for training. The training period can take up to 15 minutes. After the training is complete, it begins to analyze incoming logs to identify anomalies, and the anomalies are displayed in the CloudWatch Logs console for you to examine.

CloudWatch Logs pattern recognition extracts log patterns by identifying static and dynamic content in your logs. Patterns are useful for analyzing large log sets because a large number of log events can often be compressed into a few patterns.

For example, see the following sample of three log events.

2023-01-01 19:00:01 [INFO] Calling DynamoDB to store for resource id 12342342k124-12345 2023-01-01 19:00:02 [INFO] Calling DynamoDB to store for resource id 324892398123-12345 2023-01-01 19:00:03 [INFO] Calling DynamoDB to store for resource id 3ff231242342-12345

In the previous sample, all three log events follow one pattern:

<*> <*> [INFO] Calling DynamoDB to store for resource id <*>

Fields within a pattern are called tokens. Fields that vary within a pattern, such as a request ID or timestamp, are referred to as dynamic tokens. Dynamic tokens are represented by <*> when CloudWatch Logs displays the pattern. Each different value found for a dynamic token is called a token value.

Common examples of dynamic tokens include error codes, timestamps, and request IDs.

Logs anomaly detection uses these patterns to find anomalies. After the anomaly detector model training period, logs are evaluated against known trends. The anomaly detector flags significant fluctuations as anomalies.

Creating log anomaly detectors doesn't incur charges.

Severity and priority of anomalies and patterns

Each anomaly found by a log anomaly detector is assigned a priority. Each pattern found is assigned a severity.

  • Priority is automatically computed, and is based on both the severity level of the pattern and the amount of deviation from expected values. For example, if a certain token value suddenly increases by 500%, that anomaly might be designated as HIGH priority even if its severity is NONE.

  • Severity is based only on keywords found in the patterns such as FATAL, ERROR, and WARN. If none of these keywords are found, the severity of a pattern is marked as NONE.

Anomaly visibility time

When you create an anomaly detector, you specify the maximum anomaly visibility period for it. This is the number of days that the anomaly is displayed in the console and is returned by the ListAnomalies API operation. After this time period has elapsed for an anomaly, if it continues to happen, it's automatically accepted as regular behavior and the anomaly detector model stops flagging it as an anomaly.

If you don't adjust the visibility time when you create an anomaly detector, 21 days is used as the default.

Suppressing an anomaly

After an anomaly has been found, you can choose to suppress it temporarily or permanently. Suppressing an anomaly causes the anomaly detector to stop flagging this occurrence as an anomaly for the amount of time that you specify. When you suppress an anomaly, you can choose to suppress only that specific anomaly, or suppress all anomalies related to the pattern that the anomaly was found in.

You can still view suppressed anomalies in the console. You can also choose to stop suppressing them.

Frequently asked questions

Does AWS use my data to train machine-learning algorithms for AWS use or for other customers?

No. The anomaly detection model created by the training is based on the log events in a log group and is used only within that log group and that AWS account.

What types of log events work well with anomaly detection?

Log anomaly detection is well-suited for: Application logs and other types of logs where most log entries fit typical patterns. Log groups with events that contain a log level or severity keywords such as INFO, ERROR, and DEBUG are especially well-suited to log anomaly detection.

Log anomaly detection is not suited for: Log events with extremely long JSON structures, such as CloudTrail Logs. Pattern analysis analyzes only up to the first 1500 characters of a log line, so any characters beyond that limit are skipped.

Audit or access logs, such as VPC flow logs, will also have less success with anomaly detection. Anomaly detection is meant to find application issues, so it might not be well-suited for network or access anomalies.

To help you determine whether an anomaly detector is suited to a certain log group, use CloudWatch Logs pattern analysis to find the number of patterns in the log events in the group. If the number of patterns is no more than about 300, anomaly detection might work well. For more information about pattern analysis, see Pattern analysis.

What gets flagged as an anomaly?

The following occurrences can cause a log event to be flagged as an anomaly:

  • A log event with a pattern not seen before in the log group.

  • A significant variation to a known pattern.

  • A new value for a dynamic token that has a discrete set of usual values.

  • A large change in the number of occurrences of a value for a dynamic token.

While all the preceding items might be flagged as anomalies, they don't all mean that the application is performing poorly. For example, a higher-than-usual number of 200 success values might be flagged as an anomaly. In cases like this, you might consider suppressing these anomalies that don't indicate problems.

What happens with sensitive data that is being masked?

Any parts of log events that are masked as sensitive data are not scanned for anomalies. For more information about masking sensitive data, see Help protect sensitive log data with masking.