docs/anomaly-detection/Presets.md
Please check the Quick Start Guide to install and run
vmanomaly
Preset{{% available_from "v1.13.0" anomaly %}} mode allows for simpler configuration and anomaly detection with vmanomaly on widely-recognized metrics, such as those generated by node_exporter, which are typically challenging to monitor using static threshold-based alerting rules.
This approach represents a paradigm shift from traditional static threshold-based alerting rules, focused on raw metric values, to static rules based on anomaly_scores. These scores offer a consistent, default threshold that remains stable over time, being adjusted for trends, seasonality, data scale, thus, reducing the engineering effort required for maintenance. Anomaly scores are produced by machine learning models, which are regularly retrained on varying time frames, ensuring alerts remain current and responsive to evolving data patterns.
Additionally, preset mode minimizes user input needed to run the service - vmanomaly service can be configured by specifying only the preset name and data sources in the reader and writer sections of the configuration file. All other parameters are already preconfigured.
Each preset, including the default, comes with premade downstream assets such as alerting rules or Grafana dashboards, making integration even more straightforward.
Available presets:
To enable preset mode, preset arg should be set to particular preset name:
preset: "chosen_preset_name" # i.e. "node-exporter", "ui", etc.
Also, additional minimal set of arguments may be required from user to run the preset. See corresponding preset sections below for the details.
Run a service using config file with one of the available options.
The UI preset is the easiest to start from and is designed for users who want to quickly experiment with vmanomaly in UI mode. It requires minimal configuration and provides a user-friendly interface for exploring anomaly detection capabilities:
preset: "ui"
Please refer to the UI documentation for detailed instructions on how to use the interface and backtest your anomaly detection configurations guide.
The default preset indicates that the vmanomaly service is running in its standard mode — either because the preset section is not specified in the configuration or is explicitly set to default or default:vX.Y.
Since this mode is designed to support any configuration, it requires the user to fully define all necessary configuration sections, including:
queries section)Although this mode is designed to be as flexible as possible, it includes a pre-built Grafana dashboard. This dashboard is based on the service output, specifically anomaly scores, and helps streamline the anomaly drill-down process.
For additional benefits, this dashboard is based on VictoriaMetrics datasource rather than on
Prometheusdatasource. Please follow these instructions to enable datasource in Grafana.
You can find the Grafana dashboard .json file either here or, for already running vmanomaly instance: http://localhost:8490/presets/dashboard.json.
Alternatively, you can download it from Grafana cloud or import directly in Grafana by ID 22922. Below is the navigation guide.
P.S. Expand the "Dashboard Hints" panel to view suggested usage tips.
preset, model_alias, scheduler_alias, which are the labels that all anomaly_score metrics have by defaultAd hoc filters ad hoc Grafana filter.Metrics prefix (by default - no prefix) according to output format, defined in writer.metric_format config section.
Example: __name__: 'vmanomaly_$VAR' will result in metric names 'vmanomaly_anomaly_score', 'vmanomaly_y', 'vmanomaly_yhat', 'vmanomaly_yhat_upper', etc. - put vmanomaly_ in the field for such metrics to be properly fetched in visualizationsGroup by dashboard filter allows you to segment visualizations based on selected label (applicable to Global Anomaly Score Statistics section).
model_alias will create separate sub-visualizations for each unique model alias.for label (which stands for query alias defined in reader.queries and can be also configured in writer.metric_format config section)Threshold dashboard constant (default: 1.0) determines anomaly classification:
anomaly_score ≤ Thresholdanomaly_score > ThresholdMin. Anomaly% dashboard constant (from [0,1] range) helps to filter the visualizations in Local Anomaly Score Statistics dashboard section, leaving only timeseries with % of anomalies over time > selected threshold.
Set it to 0 to disable filtering on anomaly percentage.Show series dashboard custom multi-valued list allows to specify what vmanomaly output series to plot on Local Anomaly Score Statistics dashboard section. Defaults to y - raw metric values model saw during prediction phase. Setting it to (y, yhat, yhat_lower, yhat_upper) may ease the debugging of produced anomaly scores.This section provides aggregated anomaly score statistics, grouped by the selected Group by filter label (e.g. by for):
A table summarizing key statistics over the selected data range, including:
Threshold filter)A time series graphs for the respective group, displaying:
q99) of anomaly scores over timeA histogram visualizing the distribution of individual anomaly scores (for scores exceeding the Threshold filter), helping to identify a suitable cutoff for alerting rules and distinguish real business anomalies from minor noise.
This section provides anomaly score statistics at the individual time series level:
A table displaying raw time series data, including:
A time series graph visualizing anomaly scores over time, allowing you to pinpoint when and where anomalies occurred.
A time series graph with respective raw metric values over time.
This panel has data if these conditions were met:
- univariate models were used under respective
model_aliaslabel- The
provide_seriesargument explicitly or implicitly includedy(input data of original timeseries, observed by the model at prediction time).- In
writer.metric_formatspecial labelforwas configured to point out to query alias used inreader.queriessection. Seemetric_formatargument description on writer page for the details.
Use Show series dashboard variable to specify what vmanomaly output series to plot on Local Anomaly Score Statistics dashboard section corresponding to raw metrics scale. Defaults to y - raw metric values model saw during prediction phase. Setting it to (y, yhat, yhat_lower, yhat_upper) may ease the debugging of produced anomaly scores.
To analyze anomalies effectively, start from a high-level overview and progressively narrow down to specific time series.
Select a time range:
Choose the relevant time window (e.g., the last 12h) on dashboard's time range filter.
Apply grouping:
Use an appropriate Group by label (e.g., for to group by reader.queries). See the filters section for details.
Sort groups:
q99 or the maximum anomaly percentage in descending order (if not already).Analyze visual trends:
context_switch metric in the anomaly percentage graph shows a steady increase over time.Zoom in on the most anomalous groups:
context_switch in this case).Refine the view:
for=context_switch) to exclude unrelated data.In the Local View section, identify the most anomalous time series by:
To focus on the most significant anomalies, set the Min anomaly % dashboard constant (e.g., 5% in this example). This will filter the table and graph to display only time series with anomaly percentages over selected range above the threshold.
To further refine the results, you can:
(Global effect) Filter table columns (e.g., by a specific instance).
This applies an additional ad-hoc filter at the dashboard level, so remember to remove it after your investigation.
(Local effect) Click on legend values to isolate specific series (CTRL or CMD + click to select multiple items).
For 1 or more filtered timeseries, explore the label set directly in the table or via the popup.
To determine whether the anomaly is a true or false positive, check the raw metric values:
(Preferred method) Directly within the dashboard, if univariate models were used and the provide_series argument explicitly or implicitly included y (i.e., the original values observed by the model at prediction time). Also, using Show series filter to add yhat, yhat_lower, yhat_upper to the visualizations (if respective columns were defined in provide_series) might help to understand model behavior and the magnitude of produced anomaly scores.
(Alternative method) Use the Explore tab in Grafana:
for query MetricsQL expression from the reader.queries section of the vmanomaly config (e.g., rate(node_context_switches_total[3m]) in this particular example).Metrics browser field.{label=value} filters to the query or by clicking on legend items as shown in the screenshot.Final Analysis: This anomaly appears to be a pattern shift, potentially caused by:
vmagent, a higher number of scrapers, or more remote write activity.vmagent.Further investigation is needed to determine the exact cause.
The Node-Exporter preset simplifies the monitoring and anomaly detection of key system metrics collected by node_exporter. This preset reduces the need for manual configuration and detects anomalies in metrics such as CPU usage, network errors, and disk latency, ensuring timely identification of potential issues. Below are detailed instructions on enabling and using the Node-Exporter preset, along with a list of included assets like alerting rules and Grafana dashboard.
Node-Exporter preset assets can be also found here
For enabling Node-Exporter in config file set the preset arg accordingly. Also, include at least datasource_url-s (and tenant_id if using cluster version of VictoriaMetrics) in reader and writer sections, like that:
preset: "node-exporter"
reader:
datasource_url: "http://victoriametrics:8428/" # source victoriametrics/prometheus
# tenant_id: '0:0' # specify for cluster version
writer:
datasource_url: "http://victoriametrics:8428/" # destination victoriametrics/prometheus
# tenant_id: '0:0' # specify for cluster version
Run a service using such config file with one of the available options.
Machine learning models will be fit for each timeseries, returned by underlying MetricsQL queries.
In addition to preset label, anomaly score metric labels will also contain model and scheduler aliases for labelset uniqueness (preset, model_alias, scheduler_alias labels, respectively).
Here's an example of produced metrics:
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="system", model_alias="prophet", scheduler_alias="1d_1m"} 0.23451242720277776
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="user", model_alias="prophet", scheduler_alias="1d_1m"} 0.2637952255694444
anomaly_score{for="page_faults", instance="node-exporter:9100", job="node-exporter", preset="node-exporter", model_alias="prophet", scheduler_alias="1d_1m"} 0.00593712535
anomaly_score{for="read_latency", instance="node-exporter:9100", preset="node-exporter", model_alias="mad", scheduler_alias="1d_1m"} 0.27773362795333334
anomaly_score{for="receive_bytes", instance="node-exporter:9100", preset="node-exporter", model_alias="mad", scheduler_alias="1d_1m"} 0.037753486136666674
anomaly_score{for="transmit_bytes", instance="node-exporter:9100", preset="node-exporter", model_alias="mad", scheduler_alias="1d_1m"} 0.17633085235
anomaly_score{for="write_latency", instance="node-exporter:9100", preset="node-exporter", model_alias="mad", scheduler_alias="1d_1m"} 0.019314370926666668
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="idle", model_alias="mad", scheduler_alias="1d_1m"} 4.2323617935
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="idle", model_alias="mad", scheduler_alias="2w_1m"} 1.5261359215
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="idle", model_alias="prophet", scheduler_alias="2w_1m"} 0.5850743651
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="idle", model_alias="z-score", scheduler_alias="1d_1m"} 1.6496064663
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="idle", model_alias="z-score", scheduler_alias="2w_1m"} 0.924392581
anomaly_score{for="cpu_seconds_total", instance="node-exporter:9100", preset="node-exporter", mode="iowait", model_alias="mad", scheduler_alias="1d_1m"} 0.8571428657
...
For optimal alerting experience, we include Awesome alerts to cover indicators not addressed by the preset, as static thresholds can effectively complement our machine learning approach.
Provided
vmanomalyalerts are set to fire only if all anomaly detection models vote that the datapoint is anomalous.
You can find corresponding alerting rules here:
vmanomaly Anomaly Detection alerts: http://localhost:8490/presets/vmanomaly_alerts.ymlhttp://localhost:8490/presets/awesome_alerts.ymlGrafana dashboard .json file can be found here: http://localhost:8490/presets/dashboard.json
The produced anomaly scores will have a label for containing the name of corresponding indicator.
page_faults
</td>
<td>
node_vmstat_pgmajfault
</td>
<td>
Number of major faults that have occurred since the last update. Major faults occur when a process tries to access a page in memory that is not currently mapped in the process's address space, and it requires loading data from the disk. </td> </tr> <tr> <td>
context_switch
</td>
<td>
node_context_switches_total
</td>
<td>
This metric represents the total number of context switches across all CPUs. </td> </tr> <tr> <td>
cpu_seconds_total
</td>
<td>
node_cpu_seconds_total
</td>
<td>
Total amount of CPU time consumed by the system in seconds by CPU processing mode (e.g., user, system, idle). </td> </tr> <tr> <td>
<span style="white-space: nowrap;">host_network_transmit_errors</span>
& host_network_receive_errors
</td>
<td>
node_network_receive_errs_total,
node_network_receive_packets_total,
node_network_transmit_errs_total,
<span style="white-space: nowrap;">node_network_transmit_packets_total</span>
<td>
Total number of errors encountered while receiving/transmitting packets on the network interfaces of a node. </td> </tr> <tr> <td>
receive_bytes & transmit_bytes
</td>
<td>
node_network_receive_bytes_total,
node_network_transmit_bytes_total
</td>
<td>
Total number of bytes received/transmitted on network interfaces of a node. </td> </tr> <tr> <td>
read_latency & write_latency
</td>
<td>
node_disk_read_time_seconds_total,
node_disk_reads_completed_total,
node_disk_write_time_seconds_total,
node_disk_writes_completed_total
</td>
<td>
Disk latency. The total read/write time spent in seconds. / The total number of reads/writes completed successfully. </td> </tr> </tbody>
</table>Here's how attached Grafana dashboard can be used to drill down anomalies:
On the (global) graph 'Percentage of Anomalies', you can see a spike 8.75% of anomalies at the timestamp '2024-06-03 10:35:00'. The (global) graph 'Anomalies per Indicator' shows the indicators that were anomalous at the corresponding time.
At this timestamp on the 'Number of Anomalous Indicators by Node' graph we can identify the node that had the most anomalies: 10.142.0.27
Now you can select anomalous node to drill down further (local):
For this node from the timestamp 2024-06-03 10:35:00 CPU time spent handling software interrupts started to grow.
(cpu_seconds_total{mode="softirq"})
At the same time cpu_seconds_total for steal mode started to grow as well.