VictoriaMetrics Anomaly Detection, also known as `vmanomaly`, is a service for detecting unexpected changes in time series data. Utilizing machine learning models, it computes and pushes back an ["anomaly score"](/anomaly-detection/components/models#vmanomaly-output) for user-specified metrics. This hands-off approach to anomaly detection reduces the need for manual alert setup and can adapt to various metrics, improving your observability experience.
> **Note: `vmanomaly` is a part of [enterprise package](../enterprise.md). You need to get a [free trial license](https://victoriametrics.com/products/enterprise/trial/) for evaluation.**
Among the metrics produced by `vmanomaly` (as detailed in [vmanomaly output metrics](/anomaly-detection/components/models#vmanomaly-output)), `anomaly_score` is a pivotal one. It is **a continuous score > 0**, calculated in such a way that **scores ranging from 0.0 to 1.0 usually represent normal data**, while **scores exceeding 1.0 are typically classified as anomalous**. However, it's important to note that the threshold for anomaly detection can be customized in the alert configuration settings.
The decision to set the changepoint at `1.0` is made to ensure consistency across various models and alerting configurations, such that a score above `1.0` consistently signifies an anomaly, thus, alerting rules are maintained more easily.
> Note: `anomaly_score` is a metric itself, which preserves all labels found in input data and (optionally) appends [custom labels, specified in writer](/anomaly-detection/components/writer#metrics-formatting) - follow the link for detailed output example.
For most of the [univariate models](/anomaly-detection/components/models#univariate-models) that can generate `yhat`, `yhat_lower`, and `yhat_upper` time series in [their output](/anomaly-detection/components/models#vmanomaly-output) (such as [Prophet](/anomaly-detection/components/models#prophet) or [Z-score](/anomaly-detection/components/models#z-score)), the anomaly score is calculated as follows:
- If `yhat` (expected series behavior) equals `y` (actual value observed), then the anomaly score is 0.
- If `y` (actual value observed) falls within the `[yhat_lower, yhat_upper]` confidence interval, the anomaly score will gradually approach 1, the closer `y` is to the boundary.
- If `y` (actual value observed) strictly exceeds the `[yhat_lower, yhat_upper]` interval, the anomaly score will be greater than 1, increasing as the margin between the actual value and the expected range grows.
Please see example graph illustrating this logic below:
> p.s. please note that additional post-processing logic might be applied to produced anomaly scores, if common arguments like [`min_dev_from_expected`](https://docs.victoriametrics.com/anomaly-detection/components/models/#minimal-deviation-from-expected) or [`detection_direction`](https://docs.victoriametrics.com/anomaly-detection/components/models/#detection-direction) are enabled for a particular model. Follow the links above for the explanations.
`vmanomaly` operates on data fetched from VictoriaMetrics, where you can leverage full power of [MetricsQL](/metricsql) for data selection, sampling, and processing. Users can also [apply global filters](/#prometheus-querying-api-enhancements) for more targeted data analysis, enhancing scope limitation and tenant visibility.
`vmanomaly` operates on data fetched from VictoriaMetrics using [MetricsQL](/metricsql) queries, so the initial data quality can be fine-tuned with aggregation, grouping, and filtering to reduce noise and improve anomaly detection accuracy.
`vmanomaly` models generate [metrics](/anomaly-detection/components/models#vmanomaly-output) like `anomaly_score`, `yhat`, `yhat_lower`, `yhat_upper`, and `y`. These metrics provide a comprehensive view of the detected anomalies. The service also produces [health check metrics](/anomaly-detection/components/monitoring#metrics-generated-by-vmanomaly) for monitoring its performance.
Selecting the best model for `vmanomaly` depends on the data's nature and the [types of anomalies](https://victoriametrics.com/blog/victoriametrics-anomaly-detection-handbook-chapter-2/#categories-of-anomalies) to detect. For instance, [Z-score](/anomaly-detection/components/models#z-score) is suitable for data without trends or seasonality, while more complex patterns might require models like [Prophet](/anomaly-detection/components/models#prophet).
Also, starting from [v1.12.0](/anomaly-detection/changelog/#v1120) it's possible to auto-tune the most important params of selected model class, find [the details here](/anomaly-detection/components/models#autotuned).
Please refer to [respective blogpost on anomaly types and alerting heuristics](https://victoriametrics.com/blog/victoriametrics-anomaly-detection-handbook-chapter-2/) for more details.
While `vmanomaly` detects anomalies and produces scores, it *does not directly generate alerts*. The anomaly scores are written back to VictoriaMetrics, where an external alerting tool, like [`vmalert`](/vmalert), can be used to create alerts based on these scores for integrating it with your alerting management system.
Produced anomaly scores are designed in such a way that values from 0.0 to 1.0 indicate non-anomalous data, while a value greater than 1.0 is generally classified as an anomaly. However, there are no perfect models for anomaly detection, that's why reasonable defaults expressions like `anomaly_score > 1` may not work 100% of the time. However, anomaly scores, produced by `vmanomaly` are written back as metrics to VictoriaMetrics, where tools like [`vmalert`](/vmalert) can use [MetricsQL](/MetricsQL) expressions to fine-tune alerting thresholds and conditions, balancing between avoiding [false negatives](https://victoriametrics.com/blog/victoriametrics-anomaly-detection-handbook-chapter-1/#false-negative) and reducing [false positives](https://victoriametrics.com/blog/victoriametrics-anomaly-detection-handbook-chapter-1/#false-positive).
Starting from [v1.7.2](/anomaly-detection/changelog#v172) you can produce (and write back to VictoriaMetrics TSDB) anomaly scores for historical (backtesting) period, using `BacktestingScheduler` [component](/anomaly-detection/components/scheduler#backtesting-scheduler) to imitate consecutive "production runs" of `PeriodicScheduler` [component](/anomaly-detection/components/scheduler#periodic-scheduler). Please find an example config below:
Configuration above will produce N intervals of full length (`fit_window`=14d + `fit_every`=1h) until `to_iso` timestamp is reached to run N consecutive `fit` calls to train models; Then these models will be used to produce `M = [fit_every / sampling_frequency]` infer datapoints for `fit_every` range at the end of each such interval, imitating M consecutive calls of `infer_every` in `PeriodicScheduler` [config](/anomaly-detection/components/scheduler#periodic-scheduler). These datapoints then will be written back to VictoriaMetrics TSDB, defined in `writer` [section](/anomaly-detection/components/writer#vm-writer) for further visualization (i.e. in VMUI or Grafana)
`vmanomaly` itself is a lightweight service, resource usage is primarily dependent on [scheduling](/anomaly-detection/components/scheduler) (how often and on what data to fit/infer your models), [# and size of timeseries returned by your queries](./components/reader.md#vm-reader), and the complexity of the employed [models](/anomaly-detection/components/models). Its resource usage is directly related to these factors, making it adaptable to various operational scales.
> **Note**: Starting from [v1.13.0](/anomaly-detection/changelog#v1130), there is a mode to save anomaly detection models on host filesystem after `fit` stage (instead of keeping them in-memory by default). **Resource-intensive setups** (many models, many metrics, bigger [`fit_window` arg](/anomaly-detection/components/scheduler#periodic-scheduler-config-example)) and/or 3rd-party models that store fit data (like [ProphetModel](/anomaly-detection/components/models#prophet) or [HoltWinters](/anomaly-detection/components/models#holt-winters)) will have RAM consumption greatly reduced at a cost of slightly slower `infer` stage. To enable it, you need to set environment variable `VMANOMALY_MODEL_DUMPS_DIR` to desired location. [Helm charts](https://github.com/VictoriaMetrics/helm-charts/blob/master/charts/victoria-metrics-anomaly/README.md) are being updated accordingly ([`StatefulSet`](https://kubernetes.io/docs/concepts/workloads/controllers/statefulset/) for persistent storage starting from chart version `1.3.0`).
> **Note**: Starting from [v1.15.0](https://docs.victoriametrics.com/anomaly-detection/changelog#v1150) with the introduction of [online models](https://docs.victoriametrics.com/anomaly-detection/components/models/#online-models), you can additionally reduce resource consumption (e.g., flatten `fit` stage peaks by querying less data from VictoriaMetrics at once).
**Additional Benefits of Switching to Online Models**:
- **Reduced Latency**: Online models update incrementally, which can lead to faster response times for anomaly detection since the model continuously adapts to new data without waiting for a batch `fit`.
- **Scalability**: Handling smaller data chunks at a time reduces memory and computational overhead, making it easier to scale the anomaly detection system.
- **Improved Resource Utilization**: By spreading the computational load over time and reducing peak demands, online models make more efficient use of system resources, potentially lowering operational costs.
Here's an example of how we can switch from (offline) [Z-score model](https://docs.victoriametrics.com/anomaly-detection/components/models/#z-score) to [Online Z-score model](https://docs.victoriametrics.com/anomaly-detection/components/models/#online-z-score):
```yaml
schedulers:
periodic:
class: 'periodic'
fit_every: '1h'
fit_window: '2d'
infer_every: '1m'
# other schedulers ...
models:
zscore_example:
class: 'zscore'
schedulers: ['periodic']
# other model params ...
# other config sections ...
```
to something like
```yaml
schedulers:
periodic:
class: 'periodic'
fit_every: '180d' # we need only initial fit to start
fit_window: '4h' # reduced window, especially if the data doesn't have strong seasonality
infer_every: '1m' # the model will be updated during each infer call
# other schedulers ...
models:
zscore_example:
class: 'zscore_online'
min_n_samples_seen: 120 # i.e. minimal relevant seasonality or (initial) fit_window / sampling_frequency
schedulers: ['periodic']
# other model params ...
# other config sections ...
```
As a result, switching from the offline Z-score model to the Online Z-score model results in significant data volume reduction, i.e. over one week:
**Old Configuration**:
-`fit_window`: 2 days
-`fit_every`: 1 hour
**New Configuration**:
-`fit_window`: 4 hours
-`fit_every`: 180 days ( >1 week)
The old configuration would perform 168 (hours in a week) `fit` calls, each using 2 days (48 hours) of data, totaling 168 * 48 = 8064 hours of data for each timeseries returned.
The new configuration performs only 1 `fit` call in 180 days, using 4 hours of data initially, totaling 4 hours of data, which is **magnitutes smaller**.
P.s. `infer` data volume will remain the same for both models, so it does not affect the overall calculations.
> **Note:** As of latest release we do not support cluster or auto-scaled version yet (though, it's in our roadmap for - better backends, more parallelization, etc.), so proposed workarounds should be addressed *manually*.
`vmanomaly` can be scaled horizontally by launching multiple independent instances, each with its own [MetricsQL](/MetricsQL) queries and [configurations](/anomaly-detection/components/):
- By splitting **queries**, [defined in reader section](/anomaly-detection/components/reader#vm-reader) and spawn separate service around it. Also in case you have *only 1 query returning huge amount of timeseries*, you can further split it by applying MetricsQL filters, i.e. using "extra_filters" [param in reader](/anomaly-detection/components/reader?highlight=extra_filters#vm-reader). See the example below.
- or **models** (in case you decide to run several models for each timeseries received i.e. for averaging anomaly scores in your alerting rules of `vmalert` or using a vote approach to reduce false positives) - see `queries` arg in [model config](/anomaly-detection/components/models#queries)
- or **schedulers** (in case you want the same models to be trained under several schedules) - see `schedulers` arg [model section](/anomaly-detection/components/models#schedulers) and `scheduler` [component itself](/anomaly-detection/components/scheduler)
