VictoriaMetrics/docs/anomaly-detection/components/models.md
Fred Navruzov 78114e2cd3
docs/vmanomaly - release 1.18.2 (#7533)
### Describe Your Changes

docs update for `vmanomaly` release 1.18.2

### Checklist

The following checks are **mandatory**:

- [x] My change adheres [VictoriaMetrics contributing
guidelines](https://docs.victoriametrics.com/contributing/).
2024-11-14 17:21:33 +01:00

65 KiB
Raw Blame History

title weight menu aliases
Models 1
docs
identifier parent weight
vmanomaly-models vmanomaly-components 1
/anomaly-detection/components/models.html
/anomaly-detection/components/models/custom_model.html
/anomaly-detection/components/models/models.html

This section covers the Models component of VictoriaMetrics Anomaly Detection (commonly referred to as vmanomaly) and provides a guide on how to configure the service.

Note: Starting from v1.13.0, models can be dumped to disk instead of being stored in RAM. This option slightly reduces inference speed but significantly decreases RAM usage, particularly useful for larger setups. For more details, see the relevant FAQ section.

Note: Starting from v1.10.0 model section in config supports multiple models via aliasing.
Also, vmanomaly expects model section to be named models. Using old (flat) format with model key is deprecated and will be removed in future versions. Having model and models sections simultaneously in a config will result in only models being used:

models:
  model_univariate_1:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 2.5
    queries: ['query_alias2']  # referencing queries defined in `reader` section
  model_multivariate_1:
    class: 'isolation_forest_multivariate'  # or model.isolation_forest.IsolationForestMultivariateModel until v1.13.0
    contamination: 'auto'
    args:
      n_estimators: 100
      # i.e. to assure reproducibility of produced results each time model is fit on the same input
      random_state: 42
    # if there is no explicit `queries` arg, then the model will be run on ALL queries found in reader section
# ...

Old-style configs (< 1.10.0)

model:
    class: "zscore"  # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3.0
    # no explicit `queries` arg is provided
# ...

will be implicitly converted to

models:
  default_model:  # default model alias, backward compatibility
    class: "model.zscore.ZscoreModel"
    z_threshold: 3.0
    # queries arg is created and propagated with all query aliases found in `queries` arg of `reader` section
    queries: ['q1', 'q2', 'q3']  # i.e., if your `queries` in `reader` section has exactly q1, q2, q3 aliases
# ...

Common args

From 1.10.0, common args, supported by every model (and model type) were introduced.

Queries

Introduced in 1.10.0, as a part to support multi-model configs, queries arg is meant to define queries from VmReader particular model should be run on (meaning, all the series returned by each of these queries will be used in such model for fitting and inferencing).

queries arg is supported for all the built-in (as well as for custom) models.

This arg is backward compatible - if there is no explicit queries arg, then the model, defined in a config, will be run on ALL queries found in reader section:

models:
  model_alias_1:
    # ...
    # no explicit `queries` arg is provided

will be implicitly converted to

models:
  model_alias_1:
    # ...
    # if not set, `queries` arg is created and propagated with all query aliases found in `queries` arg of `reader` section
    queries: ['q1', 'q2', 'q3']  # i.e., if your `queries` in `reader` section has exactly q1, q2, q3 aliases

Schedulers

Introduced in 1.11.0, as a part to support multi-scheduler configs, schedulers arg is meant to define schedulers particular model should be attached to.

schedulers arg is supported for all the built-in (as well as for custom) models.

This arg is backward compatible - if there is no explicit schedulers arg, then the model, defined in a config, will be attached to ALL the schedulers found in scheduler section:

models:
  model_alias_1:
    # ...
    # no explicit `schedulers` arg is provided

will be implicitly converted to

models:
  model_alias_1:
    # ...
    # if not set, `schedulers` arg is created and propagated with all scheduler aliases found in `schedulers` section
    schedulers: ['s1', 's2', 's3']  # i.e., if your `schedulers` section has exactly s1, s2, s3 aliases

Provide series

Introduced in 1.12.0, provide_series arg limit the output generated by vmanomaly for writing. I.e. if the model produces default output series ['anomaly_score', 'yhat', 'yhat_lower', 'yhat_upper'] by specifying provide_series section as below, you limit the data being written to only ['anomaly_score'] for each metric received as a subject to anomaly detection.

models:
  model_alias_1:
    # ...
    provide_series: ['anomaly_score']  # only `anomaly_score` metric will be available for writing back to the database

Note

: If provide_series is not specified in model config, the model will produce its default model-dependent output. The output can't be less than ['anomaly_score']. Even if timestamp column is omitted, it will be implicitly added to provide_series list, as it's required for metrics to be properly written.

Detection direction

Introduced in 1.13.0, detection_direction arg can help in reducing the number of false positives and increasing the accuracy, when domain knowledge suggest to identify anomalies occurring when actual values (y) are above, below, or in both directions relative to the expected values (yhat). Available choices are: both, above_expected, below_expected.

Here's how default (backward-compatible) behavior looks like - anomalies will be tracked in both directions (y > yhat or y < yhat). This is useful when there is no domain expertise to filter the required direction.

schema_detection_direction=both

When set to above_expected, anomalies are tracked only when y > yhat.

Example metrics: Error rate, response time, page load time, number of failed transactions - metrics where lower values are better, so higher values are typically tracked.

schema_detection_direction=above_expected

When set to below_expected, anomalies are tracked only when y < yhat.

Example metrics: Service Level Agreement (SLA) compliance, conversion rate, Customer Satisfaction Score (CSAT) - metrics where higher values are better, so lower values are typically tracked.

schema_detection_direction=below_expected

Config with a split example:

models:
  model_above_expected:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3.0
    # track only cases when y > yhat, otherwise anomaly_score would be explicitly set to 0
    detection_direction: 'above_expected'
    # for this query we do not need to track lower values, thus, set anomaly detection tracking for y > yhat (above_expected)
    queries: ['query_values_the_lower_the_better']
  model_below_expected:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3.0
    # track only cases when y < yhat, otherwise anomaly_score would be explicitly set to 0
    detection_direction: 'below_expected'
    # for this query we do not need to track higher values, thus, set anomaly detection tracking for y < yhat (above_expected)
    queries: ['query_values_the_higher_the_better']
  model_bidirectional_default:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3.0
    # track in both direction, same backward-compatible behavior in case this arg is missing
    detection_direction: 'both'
    # for this query both directions can be equally important for anomaly detection, thus, setting it bidirectional (both)
    queries: ['query_values_both_direction_matters']
reader:
  # ...
  queries:
    query_values_the_lower_the_better: metricsql_expression1
    query_values_the_higher_the_better: metricsql_expression2
    query_values_both_direction_matters: metricsql_expression3
# other components like writer, schedule, monitoring

Minimal deviation from expected

Introduced in v1.13.0, the min_dev_from_expected argument is designed to reduce false positives in scenarios where deviations between the actual value (y) and the expected value (yhat) are relatively high. Such deviations can cause models to generate high anomaly scores. However, these deviations may not be significant enough in absolute values from a business perspective to be considered anomalies. This parameter ensures that anomaly scores for data points where |y - yhat| < min_dev_from_expected are explicitly set to 0. By default, if this parameter is not set, it behaves as min_dev_from_expected=0 to maintain backward compatibility.

Note

: min_dev_from_expected must be >= 0. The higher the value of min_dev_from_expected, the fewer data points will be available for anomaly detection, and vice versa.

Example: Consider a scenario where CPU utilization is low and oscillates around 0.3% (0.003). A sudden spike to 1.3% (0.013) represents a +333% increase in relative terms, but only a +1 percentage point (0.01) increase in absolute terms, which may be negligible and not warrant an alert. Setting the min_dev_from_expected argument to 0.01 (1%) will ensure that all anomaly scores for deviations <= 0.01 are set to 0.

Visualizations below demonstrate this concept; the green zone defined as the [yhat - min_dev_from_expected, yhat + min_dev_from_expected] range excludes actual data points (y) from generating anomaly scores if they fall within that range.

min_dev_from_expected-default

min_dev_from_expected-small

min_dev_from_expected-big

Example config of how to use this param based on query results:

# other components like writer, schedulers, monitoring ...
reader:
  # ...
  queries:
    # the usage of min_dev should reduce false positives here
    need_to_include_min_dev: small_abs_values_metricsql_expression
    # min_dev is not really needed here
    normal_behavior: no_need_to_exclude_small_deviations_metricsql_expression
models:
  zscore_with_min_dev:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3
    min_dev_from_expected: 5.0
    queries: ['need_to_include_min_dev']  # use such models on queries where domain experience confirm usefulness
  zscore_wo_min_dev:
    class: 'zscore' # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3
    # if not set, equals to setting min_dev_from_expected == 0
    queries: ['normal_behavior']  # use the default where it's not needed

Group By

Note

: The groupby argument works only in combination with multivariate models.

Introduced in v1.16.0, the groupby argument (list[string]) enables logical grouping within multivariate models. When specified, a separate multivariate model is trained for each unique combination of label values present in the groupby columns.

For example, to perform multivariate anomaly detection at the machine level while avoiding interference between different entities, you can set groupby: [host] or groupby: [instance]. This ensures that a separate multivariate model is trained for each individual entity (e.g., per host). Below is a simplified example illustrating how to track multivariate anomalies using CPU, RAM, and network data for each host.

# other config sections ...
reader:
  # other reader params ...
  # assume there are M unique hosts identified by the `host` label
  queries:
    # return one timeseries for each CPU mode per host, total = N*M timeseries
    cpu: sum(rate(node_cpu_seconds_total[5m])) by (host, mode)
    # return one timeseries per host, total = 1*M timeseries
    ram: | 
      (
       (node_memory_MemTotal_bytes - node_memory_MemAvailable_bytes) 
       / node_memory_MemTotal_bytes
      ) * 100 by (host)
    # return one timeseries per host for both network receive and transmit data, total = 1*M timeseries
    network: |
      sum(rate(node_network_receive_bytes_total[5m])) by (host) 
      + sum(rate(node_network_transmit_bytes_total[5m])) by (host)      

models:
  iforest: # alias for the model
    class: isolation_forest_multivariate
    contamination: 0.01
    # the multivariate model can be trained on 2+ timeseries returned by 1+ queries
    queries: [cpu, ram, network]
    # train a distinct multivariate model for each unique value found in the `host` label
    # a single multivariate model will be trained on (N + 1 + 1) timeseries, total = M models
    groupby: [host]

Model types

There are 2 model types, supported in vmanomaly, resulting in 4 possible combinations:

Each of these models can be of type

Moreover, starting from v1.15.0, there exist online (incremental) models subclass. Please refer to the correspondent section for more details.

Univariate Models

For a univariate type, one separate model is fit/used for inference per each time series, defined in its queries arg.

For example, if you have some univariate model, defined to use 3 MetricQL queries, each returning 5 time series, there will be 3*5=15 models created in total. Each such model produce individual output for each of time series.

If during an inference, you got a series having new labelset (not present in any of fitted models), the inference will be skipped until you get a model, trained particularly for such labelset during forthcoming re-fit step.

Implications: Univariate models are a go-to default, when your queries returns changing amount of individual time series of different magnitude, trend or seasonality, so you won't be mixing incompatible data with different behavior within a single fit model (context isolation).

Examples: Prophet, Holt-Winters

vmanomaly-model-type-univariate

Multivariate Models

For a multivariate type, one shared model is fit/used for inference on all time series simultaneously, defined in its queries arg.

For example, if you have some multivariate model to use 3 MetricQL queries, each returning 5 time series, there will be one shared model created in total. Once fit, this model will expect exactly 15 time series with exact same labelsets as an input. This model will produce one shared output.

Note: Starting from v1.16.0, N models — one for each unique combination of label values specified in the groupby common argument — can be trained. This allows for context separation (e.g., one model per host, region, or other relevant grouping label), leading to improved accuracy and faster training. See an example here.

If during an inference, you got a different amount of series or some series having a new labelset (not present in any of fitted models), the inference will be skipped until you get a model, trained particularly for such labelset during forthcoming re-fit step.

Implications: Multivariate models are a go-to default, when your queries returns fixed amount of individual time series (say, some aggregations), to be used for adding cross-series (and cross-query) context, useful for catching collective anomalies or novelties (expanded to multi-input scenario). For example, you may set it up for anomaly detection of CPU usage in different modes (idle, user, system, etc.) and use its cross-dependencies to detect unseen (in fit data) behavior.

Examples: IsolationForest

vmanomaly-model-type-multivariate

Rolling Models

A rolling model is a model that, once trained, cannot be (naturally) used to make inference on data, not seen during its fit phase.

An instance of rolling model is simultaneously fit and used for inference during its infer method call.

As a result, such model instances are not stored between consecutive re-fit calls (defined by fit_every arg in PeriodicScheduler), leading to lower RAM consumption.

Such models put more pressure on your reader's source, i.e. if your model should be fit on large amount of data (say, 14 days with 1-minute resolution) and at the same time you have frequent inference (say, once per minute) on new chunks of data - that's because such models require (fit + infer) window of data to be fit first to be used later in each inference call.

Note

: Rolling models require fit_every either to be missing or explicitly set equal to infer_every in your PeriodicScheduler.

Examples: RollingQuantile

vmanomaly-model-type-rolling

Non-Rolling Models

Everything that is not classified as rolling.

Produced models can be explicitly used to infer on data, not seen during its fit phase, thus, it doesn't require re-fit procedure.

Such models put less pressure on your reader's source, i.e. if you fit on large amount of data (say, 14 days with 1-minute resolution) but do it occasionally (say, once per day), at the same time you have frequent inference(say, once per minute) on new chunks of data

Note

: However, it's still highly recommended, to keep your model up-to-date with tendencies found in your data as it evolves in time.

Produced model instances are stored in-memory between consecutive re-fit calls (defined by fit_every arg in PeriodicScheduler), leading to higher RAM consumption.

Examples: Prophet

vmanomaly-model-type-non-rolling

Online Models

Introduced in v1.15.0, online (incremental) models allow defining a smaller frame fit_window and less frequent fit calls to reduce the data burden from VictoriaMetrics. They make incremental updates to model parameters during each infer_every call, even on a single datapoint. If the model doesn't support online mode, it's called offline (its parameters are only updated during fit calls).

Main differences between offline and online:

Fit stage

  • Both types have a fit stage, run on the fit_window data frame.
  • For offline models, fit_window should contain enough data to train the model (e.g., 2 seasonal periods).
  • For online models, training can start gradually from smaller chunks (e.g., 1 hour).

Infer stage

  • Both types have an infer stage, run on new datapoints (timestamps > last seen timestamp of the previous infer call).
  • Offline models use a pre-trained (during fit call) static model to make every infer call until the next fit call, when the model is completely re-trained.
  • Online models use a pre-trained (during fit call) dynamic model, which is gradually updated during each infer call with new datapoints. However, to prevent the model from accumulating outdated behavior, each fit call resets the model from scratch.

Strengths:

  • The ability to distribute the data load evenly between the initial fit and subsequent infer calls. For example, an online model can be fit on 10 1m datapoints during the initial fit stage once per month and then be gradually updated on the same 10 1m datapoints during each infer call each 10 minutes.
  • The model can adapt to new data patterns (gradually updating itself during each infer call) without needing to wait for the next fit call and one big re-training.
  • Slightly faster training/updating times compared to similar offline models.
  • Please refer to additional benefits for data-intensive setups in correspondent FAQ section.

Limitations:

  • Until the online model sees enough data (especially if the data shows strong seasonality), its predictions might be unstable, producing more false positives (anomaly_score > 1) or making false negative predictions, skipping real anomalies.
  • Not all models (e.g., complex ones like Prophet) have a direct online alternative, thus their applicability can be somewhat limited.

Each of the (built-in or custom) online models (like OnlineZscoreModel) shares the following common parameters and properties:

  • n_samples_seen_ (int) - this model property refers to the number of datapoints the model was trained on and increases from 0 (before the first fit) with each consecutive infer call.
  • min_n_samples_seen (int), optional - this parameter defines the minimum number of samples to be seen before reliably computing the anomaly score. Otherwise, the anomaly score will be 0 until n_samples_seen_ > min_n_samples_seen, as there is not enough data to trust the model's predictions. For example, if your data has hourly seasonality and '1m' frequency, setting min_n_samples_seen_ to 288 (1440 minutes in a day / 5 minutes) should be sufficient.

Offline models

Every other model that isn't online. Offline models are completely re-trained during fit call and aren't updated during consecutive infer calls.

Built-in Models

Overview

VictoriaMetrics Anomaly Detection models support 2 groups of parameters:

  • vmanomaly-specific arguments - please refer to Parameters specific for vmanomaly and Default model parameters subsections for each of the models below.
  • Arguments to inner model (say, Facebook's Prophet), passed in a args argument as key-value pairs, that will be directly given to the model during initialization to allow granular control. Optional.

Note

: For users who may not be familiar with Python data types such as list[dict], a dictionary in Python is a data structure that stores data values in key-value pairs. This structure allows for efficient data retrieval and management.

Models:

AutoTuned

Tuning hyperparameters of a model can be tricky and often requires in-depth knowledge of Machine Learning. AutoTunedModel is designed specifically to take the cognitive load off the user - specify as little as anomaly_percentage param from (0, 0.5) interval and tuned_model_class (i.e. model.zscore.ZscoreModel) to get it working with best settings that match your data.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.auto.AutoTunedModel" (or auto starting from v1.13.0 with class alias support)
  • tuned_class_name (string) - Built-in model class to tune, i.e. model.zscore.ZscoreModel (or zscore starting from v1.13.0 with class alias support).
  • optimization_params (dict) - Optimization parameters for unsupervised model tuning. Control % of found anomalies, as well as a tradeoff between time spent and the accuracy. The more timeout and n_trials are, the better model configuration can be found for tuned_class_name, but the longer it takes and vice versa. Set n_jobs to -1 to use all the CPUs available, it makes sense if only you have a big dataset to train on during fit calls, otherwise overhead isn't worth it.
    • anomaly_percentage (float) - Expected percentage of anomalies that can be seen in training data, from (0, 0.5) interval.
    • optimized_business_params (list[string]) - Starting from v1.15.0 this argument allows particular business-specific parameters such as detection_direction or min_dev_from_expected to remain unchanged during optimizations, retaining their default values. I.e. setting optimized_business_params to ['detection_direction'] will allow to optimize only detection_direction business-specific arg, while min_dev_from_expected will retain its default value (0.0). By default and if not set, will be equal to [] (empty list), meaning no business params will be optimized. A recommended option is to leave it empty for more stable results and increased convergence (less iterations needed for a good result).
    • seed (int) - Random seed for reproducibility and deterministic nature of underlying optimizations.
    • n_splits (int) - How many folds to create for hyperparameter tuning out of your data. The higher, the longer it takes but the better the results can be. Defaults to 3.
    • n_trials (int) - How many trials to sample from hyperparameter search space. The higher, the longer it takes but the better the results can be. Defaults to 128.
    • timeout (float) - How many seconds in total can be spent on each model to tune hyperparameters. The higher, the longer it takes, allowing to test more trials out of defined n_trials, but the better the results can be.

vmanomaly-autotune-schema

# ...
models:
  your_desired_alias_for_a_model:
    class: 'auto'  # or 'model.auto.AutoTunedModel' until v1.13.0
    tuned_class_name: 'zscore'  # or 'model.zscore.ZscoreModel' until v1.13.0
    optimization_params:
      anomaly_percentage: 0.004  # required. i.e. we expect <= 0.4% of anomalies to be present in training data
      seed: 42  # fix reproducibility & determinism
      n_splits: 4  # how much folds are created for internal cross-validation
      n_trials: 128  # how many configurations to sample from search space during optimization
      timeout: 10  # how many seconds to spend on optimization for each trained model during `fit` phase call
      n_jobs: 1  # how many jobs in parallel to launch. Consider making it > 1 only if you have fit window containing > 10000 datapoints for each series
      optimized_business_params: []  # business-specific params to include in optimization, if not set is empty list
  # ...

Note

: There are some expected limitations of Autotune mode:

  • It can't be made on your custom model.
  • It can't be applied to itself (like tuned_class_name: 'model.auto.AutoTunedModel')
  • AutoTunedModel can't be used on rolling models like RollingQuantile in combination with on-disk model storage mode, as the rolling models exists only during infer calls and aren't persisted neither in RAM, nor on disk.

Prophet

vmanomaly uses the Facebook Prophet implementation for time series forecasting, with detailed usage provided in the Prophet library documentation. All Prophet parameters are supported and can be directly passed to the model via args argument.

Note

: ProphetModel is a univariate, non-rolling, offline model.

Note

: Starting with v1.18.2, the format for tz_seasonalities has been updated to enhance flexibility. Previously, it accepted a list of strings (e.g., ['hod', 'minute']). Now, it follows the same structure as custom seasonalities defined in the seasonalities argument (e.g., {"name": "hod", "fourier_order": 5, "mode": "additive"}). This change is backward-compatible, so older configurations will be automatically converted to the new format using default values.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.prophet.ProphetModel" (or prophet starting from v1.13.0 with class alias support)
  • seasonalities (list[dict], optional): Additional seasonal components to include in Prophet. See Prophets add_seasonality() documentation for details.
  • tz_aware (bool): (Available since v1.18.0) Enables handling of timezone-aware timestamps. Default is False. Should be used with tz_seasonalities and tz_use_cyclical_encoding parameters.
  • tz_seasonalities (list[dict]): (Available since v1.18.0) Specifies timezone-aware seasonal components. Requires tz_aware=True. Supported options include minute, hod (hour of day), dow (day of week), and month (month of year). Starting with v1.18.2, users can configure additional parameters for each seasonality, such as fourier_order, prior_scale, and mode. For more details, please refer to the Timezone-unaware configuration example below.
  • tz_use_cyclical_encoding (bool): (Available since v1.18.0) If set to True, applies cyclical encoding technique to timezone-aware seasonalities. Should be used with tz_aware=True and tz_seasonalities.

Note

: Apart from standard vmanomaly output, Prophet model can provide additional metrics.

Additional output metrics produced by FB Prophet Depending on chosen seasonality parameter FB Prophet can return additional metrics such as:

  • trend, trend_lower, trend_upper
  • additive_terms, additive_terms_lower, additive_terms_upper,
  • multiplicative_terms, multiplicative_terms_lower, multiplicative_terms_upper,
  • daily, daily_lower, daily_upper,
  • hourly, hourly_lower, hourly_upper,
  • holidays, holidays_lower, holidays_upper,
  • and a number of columns for each holiday if holidays param is set

Config Example

Timezone-unaware example:

models:
  your_desired_alias_for_a_model:
    class: 'prophet'  # or 'model.prophet.ProphetModel' until v1.13.0
    provide_series: ['anomaly_score', 'yhat', 'yhat_lower', 'yhat_upper', 'trend']
    seasonalities:
      - name: 'hourly'
        period: 0.04166666666
        fourier_order: 30
        prior_scale: 20
    # inner model args (key-value pairs) accepted by
    # https://facebook.github.io/prophet/docs/quick_start.html#python-api
    args:
      interval_width: 0.98  # see https://facebook.github.io/prophet/docs/uncertainty_intervals.html
      country_holidays: 'US'

Timezone-aware example:

models:
  your_desired_alias_for_a_model:
    class: 'prophet'  # or 'model.prophet.ProphetModel' until v1.13.0
    provide_series: ['anomaly_score', 'yhat', 'yhat_lower', 'yhat_upper', 'trend']
    tz_aware: True
    tz_use_cyclical_encoding: True
    tz_seasonalities: # intra-day + intra-week seasonality, no intra-year / sub-hour seasonality
      - name: 'hod'  # intra-day seasonality, hour of the day
        fourier_order: 5  # keep it 3-8 based on intraday pattern complexity
        prior_scale: 10
      - name: 'dow'  # intra-week seasonality, time of the week
        fourier_order: 2  # keep it 2-4, as dependencies are learned separately for each weekday
    # inner model args (key-value pairs) accepted by
    # https://facebook.github.io/prophet/docs/quick_start.html#python-api
    args:
      interval_width: 0.98  # see https://facebook.github.io/prophet/docs/uncertainty_intervals.html
      country_holidays: 'US'

Resulting metrics of the model are described here

Z-score

Note

: ZScoreModel is a univariate, non-rolling, offline model.

Model is useful for initial testing and for simpler data (de-trended data without strict seasonality and with anomalies of similar magnitude as your "normal" data).

Parameters specific for vmanomaly:

  • class (string) - model class name "model.zscore.ZscoreModel" (or zscore starting from v1.13.0 with class alias support)
  • z_threshold (float, optional) - standard score for calculation boundaries and anomaly score. Defaults to 2.5.

Config Example

models:
  your_desired_alias_for_a_model:
    class: "zscore"  # or 'model.zscore.ZscoreModel' until v1.13.0
    z_threshold: 3.5

Resulting metrics of the model are described here.

Online Z-score

Note

: OnlineZScoreModel is a univariate, non-rolling, online model.

Online version of existing Z-score implementation with the same exact behavior and implications. Introduced in v1.15.0

Parameters specific for vmanomaly:

  • class (string) - model class name "model.online.OnlineZscoreModel" (or zscore_online starting from v1.15.0 with class alias support)
  • z_threshold (float, optional) - standard score for calculation boundaries and anomaly score. Defaults to 2.5.
  • min_n_samples_seen (int, optional) - the minimum number of samples to be seen (n_samples_seen_ property) before computing the anomaly score. Otherwise, the anomaly score will be 0, as there is not enough data to trust the model's predictions. Defaults to 16.

Config Example

models:
  your_desired_alias_for_a_model:
    class: "zscore_online"  # or 'model.online.OnlineZscoreModel'
    z_threshold: 3.5
    min_n_samples_seen: 128  # i.e. calculate it as full seasonality / data freq
    provide_series: ['anomaly_score', 'yhat']  # common arg example

Resulting metrics of the model are described here.

Holt-Winters

Note

: HoltWinters is a univariate, non-rolling, offline model.

Here we use Holt-Winters Exponential Smoothing implementation from statsmodels library. All parameters from this library can be passed to the model.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.holtwinters.HoltWinters" (or holtwinters starting from v1.13.0 with class alias support)

  • frequency (string) - Must be set equal to sampling_period. Model needs to know expected data-points frequency (e.g. '10m'). If omitted, frequency is guessed during fitting as the median of intervals between fitting data timestamps. During inference, if incoming data doesn't have the same frequency, then it will be interpolated. E.g. data comes at 15 seconds resolution, and our resample_freq is '1m'. Then fitting data will be downsampled to '1m' and internal model is trained at '1m' intervals. So, during inference, prediction data would be produced at '1m' intervals, but interpolated to "15s" to match with expected output, as output data must have the same timestamps. As accepted by pandas.Timedelta (e.g. '5m').

  • seasonality (string, optional) - As accepted by pandas.Timedelta.

  • If seasonal_periods is not specified, it is calculated as seasonality / frequency Used to compute "seasonal_periods" param for the model (e.g. '1D' or '1W').

  • z_threshold (float, optional) - standard score for calculating boundaries to define anomaly score. Defaults to 2.5.

Default model parameters:

  • If parameter seasonal is not specified, default value will be add.

  • If parameter initialization_method is not specified, default value will be estimated.

  • args (dict, optional) - Inner model args (key-value pairs). See accepted params in model documentation. Defaults to empty (not provided). Example: {"seasonal": "add", "initialization_method": "estimated"}

Config Example

models:
  your_desired_alias_for_a_model:
    class: "holtwinters"  # or 'model.holtwinters.HoltWinters' until v1.13.0
    seasonality: '1d'
    frequency: '1h'
    # Inner model args (key-value pairs) accepted by statsmodels.tsa.holtwinters.ExponentialSmoothing
    args:
      seasonal: 'add'
      initialization_method: 'estimated'

Resulting metrics of the model are described here.

MAD (Median Absolute Deviation)

Note

: MADModel is a univariate, non-rolling, offline model.

The MAD model is a robust method for anomaly detection that is less sensitive to outliers in data compared to standard deviation-based models. It considers a point as an anomaly if the absolute deviation from the median is significantly large.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.mad.MADModel" (or mad starting from v1.13.0 with class alias support)
  • threshold (float, optional) - The threshold multiplier for the MAD to determine anomalies. Defaults to 2.5. Higher values will identify fewer points as anomalies.

Config Example

models:
  your_desired_alias_for_a_model:
    class: "mad"  # or 'model.mad.MADModel' until v1.13.0
    threshold: 2.5

Resulting metrics of the model are described here.

Online MAD

Note

: OnlineMADModel is a univariate, non-rolling, online model.

The MAD model is a robust method for anomaly detection that is less sensitive to outliers in data compared to standard deviation-based models. It considers a point as an anomaly if the absolute deviation from the median is significantly large. This is the online approximate version, based on t-digests for online quantile estimation. introduced in v1.15.0

Parameters specific for vmanomaly:

  • class (string) - model class name "model.online.OnlineMADModel" (or mad_online starting from v1.13.0 with class alias support)
  • threshold (float, optional) - The threshold multiplier for the MAD to determine anomalies. Defaults to 2.5. Higher values will identify fewer points as anomalies.
  • min_n_samples_seen (int, optional) - the minimum number of samples to be seen (n_samples_seen_ property) before computing the anomaly score. Otherwise, the anomaly score will be 0, as there is not enough data to trust the model's predictions. Defaults to 16.
  • compression (int, optional) - the compression parameter for underlying t-digest. Higher values mean higher accuracy but higher memory usage. By default 100.

Config Example

models:
  your_desired_alias_for_a_model:
    class: "mad_online"  # or 'model.online.OnlineMADModel'
    threshold: 2.5
    min_n_samples_seen: 128  # i.e. calculate it as full seasonality / data freq
    compression: 100  # higher values mean higher accuracy but higher memory usage
    provide_series: ['anomaly_score', 'yhat']  # common arg example

Resulting metrics of the model are described here.

Rolling Quantile

Note

: RollingQuantileModel is a univariate, rolling, offline model.

This model is best used on data with short evolving patterns (i.e. 10-100 datapoints of particular frequency), as it adapts to changes over a rolling window.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.rolling_quantile.RollingQuantileModel" (or rolling_quantile starting from v1.13.0 with class alias support)
  • quantile (float) - quantile value, from 0.5 to 1.0. This constraint is implied by 2-sided confidence interval.
  • window_steps (integer) - size of the moving window. (see 'sampling_period')

Config Example

models:
  your_desired_alias_for_a_model:
    class: "rolling_quantile" # or 'model.rolling_quantile.RollingQuantileModel' until v1.13.0
    quantile: 0.9
    window_steps: 96

Resulting metrics of the model are described here.

Online Seasonal Quantile

Note

: OnlineQuantileModel is a univariate, non-rolling, online model.

Online (seasonal) quantile utilizes a set of approximate distributions, based on t-digests for online quantile estimation. Introduced in v1.15.0.

Best used on de-trended data with strong (possibly multiple) seasonalities. Can act as a (slightly less powerful) replacement to ProphetModel.

It uses the quantiles triplet to calculate yhat_lower, yhat, and yhat_upper output, respectively, for each of the min_subseasons sub-intervals contained in seasonal_interval. For example, with '4d' + '2h' seasonality patterns (multiple), it will hold and update 24*4 / 2 = 48 consecutive estimates (each 2 hours long).

Parameters specific for vmanomaly:

  • class (string) - model class name "model.online.OnlineQuantileModel" (or quantile_online starting from v1.13.0 with class alias support)
  • quantiles (list[float], optional) - The quantiles to estimate. yhat_lower, yhat, yhat_upper are the quantile order. By default (0.01, 0.5, 0.99).
  • seasonal_interval (string, optional) - the interval for the seasonal adjustment. If not set, the model will equal to a simple online quantile model. By default not set.
  • min_subseason (str, optional) - the minimum interval to estimate quantiles for. By default not set. Note that the minimum interval should be a multiple of the seasonal interval, i.e. if seasonal_interval='2h', then min_subseason='15m' is valid, but '37m' is not.
  • use_transform (bool, optional) - whether to internally apply a log1p(abs(x)) * sign(x) transformation to the data to stabilize internal quantile estimation. Does not affect the scale of produced output (i.e. yhat) By default False.
  • global_smoothing (float, optional) - the smoothing parameter for the global quantiles. i.e. the output is a weighted average of the global and seasonal quantiles (if seasonal_interval and min_subseason args are set). Should be from [0, 1] interval, where 0 means no smoothing and 1 means using only global quantile values.
  • scale (float, optional) - the scaling factor for the yhat_lower and yhat_upper quantiles. By default 1.0 (no scaling). if > 1, increases the boundaries [yhat_lower, yhat_upper] that define "non-anomalous" points. Should be > 0.
  • season_starts_from (str, optional) - the start date for the seasonal adjustment, as a reference point to start counting the intervals. By default '1970-01-01'.
  • min_n_samples_seen (int, optional) - the minimum number of samples to be seen (n_samples_seen_ property) before computing the anomaly score. Otherwise, the anomaly score will be 0, as there is not enough data to trust the model's predictions. Defaults to 16.
  • compression (int, optional) - the compression parameter for the underlying t-digests. Higher values mean higher accuracy but higher memory usage. By default 100.

Config Example

Suppose we have a data with strong intra-day (hourly) and intra-week (daily) seasonality, data granularity is '5m' with up to 5% expected outliers present in data. Then you can apply similar config:

models:
  your_desired_alias_for_a_model:
    class: "quantile_online"  # or 'model.online.OnlineQuantileModel'
    quantiles: [0.025, 0.5, 0.975]  # lowered to exclude anomalous edges, can be compensated by `scale` param > 1
    seasonal_interval: '7d'  # longest seasonality (week, day) = week, starting from `season_starts_from`
    min_subseason: '1h'  # smallest seasonality (week, day, hour) = hour, will have its own quantile estimates
    min_n_samples_seen: 288 # 1440 / 5 - at least 1 full day, ideal = 1440 / 5 * 7 - one full week (seasonal_interval)
    scale: 1.1  # to compensate lowered quantile boundaries with wider intervals
    season_starts_from: '2024-01-01'  # interval calculation starting point, especially for uncommon seasonalities like '36h' or '12d'
    compression: 100  # higher values mean higher accuracy but higher memory usage
    provide_series: ['anomaly_score', 'yhat']  # common arg example

Resulting metrics of the model are described here.

Seasonal Trend Decomposition

Note

: StdModel is a univariate, rolling, offline model.

Here we use Seasonal Decompose implementation from statsmodels library. Parameters from this library can be passed to the model. Some parameters are specifically predefined in vmanomaly and can't be changed by user(model='additive', two_sided=False).

Parameters specific for vmanomaly:

  • class (string) - model class name "model.std.StdModel" (or std starting from v1.13.0 with class alias support)
  • period (integer) - Number of datapoints in one season.
  • z_threshold (float, optional) - standard score for calculating boundaries to define anomaly score. Defaults to 2.5.

Config Example

models:
  your_desired_alias_for_a_model:
    class: "std"  # or 'model.std.StdModel' starting from v1.13.0
    period: 2

Resulting metrics of the model are described here.

Additional output metrics produced by Seasonal Trend Decomposition model

  • resid - The residual component of the data series.
  • trend - The trend component of the data series.
  • seasonal - The seasonal component of the data series.

Isolation forest (Multivariate)

Note

: IsolationForestModel is a univariate, non-rolling, offline model.

Note

: IsolationForestMultivariateModel is a multivariate, non-rolling, offline model.

Detects anomalies using binary trees. The algorithm has a linear time complexity and a low memory requirement, which works well with high-volume data. It can be used on both univariate and multivariate data, but it is more effective in multivariate case.

Important: Be aware of the curse of dimensionality. Don't use single multivariate model if you expect your queries to return many time series of less datapoints that the number of metrics. In such case it is hard for a model to learn meaningful dependencies from too sparse data hypercube.

Here we use Isolation Forest implementation from scikit-learn library. All parameters from this library can be passed to the model.

Parameters specific for vmanomaly:

  • class (string) - model class name "model.isolation_forest.IsolationForestMultivariateModel" (or isolation_forest_multivariate starting from v1.13.0 with class alias support)

  • contamination (float or string, optional) - The amount of contamination of the data set, i.e. the proportion of outliers in the data set. Used when fitting to define the threshold on the scores of the samples. Default value - "auto". Should be either "auto" or be in the range (0.0, 0.5].

  • seasonal_features (list of string) - List of seasonality to encode through cyclical encoding, i.e. dow (day of week). Introduced in 1.12.0.

    • Empty by default for backward compatibility.
    • Example: seasonal_features: ['dow', 'hod'].
    • Supported seasonalities:
      • "minute" - minute of hour (0-59)
      • "hod" - hour of day (0-23)
      • "dow" - day of week (1-7)
      • "month" - month of year (1-12)
  • args (dict, optional) - Inner model args (key-value pairs). See accepted params in model documentation. Defaults to empty (not provided). Example: {"random_state": 42, "n_estimators": 100}

Config Example

models:
  your_desired_alias_for_a_model:
    # To use univariate model, substitute class argument with "model.isolation_forest.IsolationForestModel".
    class: "isolation_forest_multivariate" # or 'model.isolation_forest.IsolationForestMultivariateModel' until v1.13.0
    contamination: "0.01"
    provide_series: ['anomaly_score']
    seasonal_features: ['dow', 'hod']
    args:
      n_estimators: 100
      # i.e. to assure reproducibility of produced results each time model is fit on the same input
      random_state: 42

Resulting metrics of the model are described here.

vmanomaly output

When vmanomaly is executed, it generates various metrics, the specifics of which depend on the model employed. These metrics can be renamed in the writer's section.

The default metrics produced by vmanomaly include:

  • anomaly_score: This is the primary metric.

    • It is designed in such a way that values from 0.0 to 1.0 indicate non-anomalous data.
    • A value greater than 1.0 is generally classified as an anomaly, although this threshold can be adjusted in the alerting configuration.
    • The decision to set the changepoint at 1 was made to ensure consistency across various models and alerting configurations, such that a score above 1 consistently signifies an anomaly.
  • yhat: This represents the predicted expected value.

  • yhat_lower: This indicates the predicted lower boundary.

  • yhat_upper: This refers to the predicted upper boundary.

  • y: This is the original value obtained from the query result.

Important: Be aware that if NaN (Not a Number) or Inf (Infinity) values are present in the input data during infer model calls, the model will produce NaN as the anomaly_score for these particular instances.

vmanomaly monitoring metrics

Each model exposes several monitoring metrics to its health_path endpoint:

Custom Model Guide

Apart from vmanomaly built-in models, users can create their own custom models for anomaly detection.

Here in this guide, we will

  • Make a file containing our custom model definition
  • Define VictoriaMetrics Anomaly Detection config file to use our custom model
  • Run service

Note

: The file containing the model should be written in Python language (3.11+)

1. Custom model

Note

: By default, each custom model is created as univariate / non-rolling model. If you want to override this behavior, define models inherited from RollingModel (to get a rolling model), or having is_multivariate class arg set to True (please refer to the code example below).

We'll create custom_model.py file with CustomModel class that will inherit from vmanomaly's Model base class. In the CustomModel class, the following methods are required: - __init__, fit, infer, serialize and deserialize:

  • __init__ method should initiate parameters for the model.

    Note: if your model relies on configs that have arg key-value pair argument, like Prophet, do not forget to use Python's **kwargs in method's signature and to explicitly call

    super().__init__(**kwargs)
    

    to initialize the base class each model derives from

  • fit method should contain the model training process. Please be aware that for RollingModel defining fit method is not needed, as the whole fit/infer process should be defined completely in infer method.

  • infer should return Pandas.DataFrame object with model's inferences.

  • serialize method that saves the model on disk.

  • deserialize load the saved model from disk.

For the sake of simplicity, the model in this example will return one of two values of anomaly_score - 0 or 1 depending on input parameter percentage.

import numpy as np
import pandas as pd
import scipy.stats as st
import logging
from pickle import dumps

from model.model import (
  PICKLE_PROTOCOL,
  Model,
  deserialize_basic
)
# from model.model import RollingModel  # inherit from it for your model to be of rolling type
logger = logging.getLogger(__name__)


class CustomModel(Model):
  """
  Custom model implementation.
  """
  # by default, each `Model` will be created as a univariate one
  # uncomment line below for it to be of multivariate type
  # is_multivariate = True
  # by default, each `Model` will be created as offline
  # uncomment line below for it to be of type online
  # is_online = True
  
  def __init__(self, percentage: float = 0.95, **kwargs):
    super().__init__(**kwargs)
    self.percentage = percentage
    self._mean = np.nan
    self._std = np.nan

  def fit(self, df: pd.DataFrame):
    # Model fit process:
    y = df['y']
    self._mean = np.mean(y)
    self._std = np.std(y)
    if self._std == 0.0:
      self._std = 1 / 65536

  def infer(self, df: pd.DataFrame) -> np.array:
    # Inference process:
    y = df['y']
    zscores = (y - self._mean) / self._std
    anomaly_score_cdf = st.norm.cdf(np.abs(zscores))
    df_pred = df[['timestamp', 'y']].copy()
    df_pred['anomaly_score'] = anomaly_score_cdf > self.percentage
    df_pred['anomaly_score'] = df_pred['anomaly_score'].astype('int32', errors='ignore')

    return df_pred

    def serialize(self) -> None:
      return dumps(self, protocol=PICKLE_PROTOCOL)

    @staticmethod
    def deserialize(model: str | bytes) -> 'CustomModel':
      return deserialize_basic(model)

2. Configuration file

Next, we need to create config.yaml file with vmanomaly configuration and model input parameters. In the config file's models section we need to set our model class to model.custom.CustomModel (or custom starting from v1.13.0 with class alias support) and define all parameters used in __init__ method. You can find out more about configuration parameters in vmanomaly config docs.

schedulers:
  s1:
    infer_every: "1m"
    fit_every: "1m"
    fit_window: "1d"

models:
  custom_model:
    class: "custom"  # or 'model.model.CustomModel' until v1.13.0
    percentage: 0.9


reader:
  datasource_url: "http://victoriametrics:8428/"
  sampling_period: '1m'
  queries:
    ingestion_rate: 'sum(rate(vm_rows_inserted_total)) by (type)'
    churn_rate: 'sum(rate(vm_new_timeseries_created_total[5m]))'

writer:
  datasource_url: "http://victoriametrics:8428/"
  metric_format:
    __name__: "custom_$VAR"
    for: "$QUERY_KEY"
    run: "test-format"

monitoring:
  # /metrics server.
  pull:
    port: 8080
  push:
    url: "http://victoriametrics:8428/"
    extra_labels:
      job: "vmanomaly-develop"
      config: "custom.yaml"

3. Running custom model

Let's pull the docker image for vmanomaly:

docker pull victoriametrics/vmanomaly:v1.18.2

Now we can run the docker container putting as volumes both config and model file:

Note

: place the model file to /model/custom.py path when copying

./custom_model.py:/vmanomaly/model/custom.py

docker run -it \
-v $(PWD)/license:/license \
-v $(PWD)/custom_model.py:/vmanomaly/model/custom.py \
-v $(PWD)/custom.yaml:/config.yaml \
victoriametrics/vmanomaly:v1.18.2 /config.yaml \
--licenseFile=/license

Please find more detailed instructions (license, etc.) here

Output

As the result, this model will return metric with labels, configured previously in config.yaml. In this particular example, 2 metrics will be produced. Also, there will be added other metrics from input query result.

{__name__="custom_anomaly_score", for="ingestion_rate", model_alias="custom_model", scheduler_alias="s1", run="test-format"},
{__name__="custom_anomaly_score", for="churn_rate",     model_alias="custom_model", scheduler_alias="s1", run="test-format"}