Reader #

VictoriaMetrics Anomaly Detection (vmanomaly) primarily uses VmReader to ingest data. This reader focuses on fetching time-series data directly from VictoriaMetrics with the help of powerful MetricsQL expressions for aggregating, filtering and grouping your data, ensuring seamless integration and efficient data handling.

Future updates will introduce additional readers, expanding the range of data sources vmanomaly can work with.

VM reader #

Old format like

      # other config sections ...
reader:
  class: 'vm'
  datasource_url: 'http://localhost:8428'  # source victoriametrics/prometheus
  sampling_period: "10s"  # set it <= min(infer_every) in schedulers section
  queries:
    # old format {query_alias: query_expr}, prior to 1.13, will be converted to a new format automatically
    vmb: 'avg(vm_blocks)'
    

will be converted to a new one with a warning raised in logs:

      # other config sections ...
reader:
  class: 'vm'
  datasource_url: 'http://localhost:8428'  # source victoriametrics/prometheus
  sampling_period: '10s'
  queries:
    # old format {query_alias: query_expr}, prior to 1.13, will be converted to a new format automatically
    vmb:
      expr: 'avg(vm_blocks)'  # initial MetricsQL expression
      step: '10s'  # individual step for this query, will be filled with `sampling_period` from the root level
      data_range: ['-inf', 'inf']  # by default, no constraints applied on data range
      tz: 'UTC'  # by default, tz-free data is used throughout the model lifecycle
      # new query-level arguments will be added in backward-compatible way in future releases
    

Per-query parameters #

There is change Available from v1.13.0 of queries arg format. Now each query alias supports the next (sub)fields, which override reader-level parameters, if set:

• expr (string): MetricsQL/PromQL expression that defines an input for VmReader. As accepted by /query_range?query=%s. i.e. avg(vm_blocks)

• step (string): query-level frequency of the points returned, i.e. 30s. Will be converted to /query_range?step=%s param (in seconds). Useful to optimize total amount of data read from VictoriaMetrics, where different queries may have different frequencies for different machine learning models to run on.

  Note

  If not set explicitly (or if older config style prior to v1.13.0 ) is used, then it is set to reader-level sampling_period arg.

  Note

  Having different individual step args for queries (i.e. 30s for q1 and 2m for q2) is not yet supported for multivariate model if you want to run it on several queries simultaneously (i.e. setting queries arg of a model to [q1, q2]).

• data_range Available from v1.15.1 (list[float | string]): It allows defining valid data ranges for input per individual query in queries, resulting in:

  • High anomaly scores (>1) when the data falls outside the expected range, indicating a data range constraint violation (e.g. improperly configured metricsQL query, sensor malfunction, overflows in underlying metrics, etc.). Anomaly scores can be set to a specific value, like 5, to indicate a strong violation, using the anomaly_score_outside_data_range arg of a respective model this query is used in.
  • Lowest anomaly scores (=0) when the model’s predictions (yhat) fall outside the expected range, meaning uncertain predictions that does not really aligh with the data.

  Works together with anomaly_score_outside_data_range arg of a model to determine the anomaly score for such cases as well as with clip_predictions arg of a model to clip the predictions to the expected range.

  Note

  If not set explicitly (or if older config style prior to v1.13.0 ) is used, then it is set to reader-level data_range arg Available from v1.18.1

• max_points_per_query Available from v1.17.0 (int): Optional arg, overrides how search.maxPointsPerTimeseries flag Available from v1.14.1 impacts vmanomaly on splitting long fit_window queries into smaller sub-intervals. This helps users avoid hitting the search.maxQueryDuration limit for individual queries by distributing initial query across multiple subquery requests with minimal overhead. Set less than search.maxPointsPerTimeseries if hitting maxQueryDuration limits. If set on a query-level, it overrides the global max_points_per_query (reader-level).

• tz Available from v1.18.0 (string): this optional argument enables timezone specification per query, overriding the reader’s default tz. This setting helps to account for local timezone shifts, such as DST , in models that are sensitive to seasonal variations (e.g., ProphetModel or OnlineQuantileModel ).

• tenant_id Available from v1.19.0 (string): this optional argument enables tenant-level separation for queries (e.g. query1 to get the data from tenant “0:0”, query2 - from tenant “1:0”). It works as follows:

  • if not set, inherits reader-level tenant_id
  • if set, overrides reader-level tenant_id
  • raises config validation error, if reader-level is not set and query-level is found (mixing of VictoriaMetrics single-node and cluster is prohibited in a single config)
  • raises config validation warning, if writer.tenant_id is not explicitly set to multitenant when reader uses tenants, meaning VictoriaMetrics cluster will be used for data querying.
  • also raises config validation error if a set of reader.queries for multivariate models has different tenant_ids (meaning tenant data is mixed, and special labels like vm_project_id, vm_account_id will have ambiguous values )
  Note

  The recommended approach for using per-query tenant_ids is to set both reader.tenant_id and writer.tenant_id to multitenant. See this section for more details. Configurations where reader.tenant_id equals writer.tenant_id and is not multitenant are also considered safe, provided there is a single, DISTINCT tenant_id defined in the reader (either at the reader level or the query level, if set).

• offset Available from v1.25.3 (string): this optional argument allows specifying a time offset for the query, which can be useful for adjusting the query time range to account for data collection delays or other timing issues. The offset is specified as a string (e.g., “15s”, “-20s”) and will be applied to the query time range. Valid resolutions are ms, s, m, h, d (miliseconds, seconds, minutes, hours, days). If not set, defaults to 0s (0). See FAQ for more details.

Per-query config example #

      reader:
  class: 'vm'
  sampling_period: '1m'
  datasource_url: 'https://play.victoriametrics.com/'  # source victoriametrics/prometheus
  max_points_per_query: 10000
  data_range: [0, 'inf']
  tenant_id: 'multitenant'
  offset: '0s'  # optional, defaults to 0s if not set
  # other reader params ...
  queries:
    ingestion_rate_t1:
      expr: 'sum(rate(vm_rows_inserted_total[5m])) by (type) > 0'
      step: '2m'  # overrides global `sampling_period` of 1m
      data_range: [10, 'inf']  # meaning only positive values > 10 are expected, i.e. a value `y` < 10 will trigger anomaly score > 1
      max_points_per_query: 5000 # overrides reader-level value of 10000 for `ingestion_rate` query
      tz: 'America/New_York'  # to override reader-wise `tz`
      tenant_id: '1:0'  # overriding tenant_id to isolate data
    ingestion_rate_t2:
      expr: 'sum(rate(vm_rows_inserted_total[5m])) by (type) > 0'
      step: '2m'  # overrides global `sampling_period` of 1m
      data_range: [10, 'inf']  # meaning only positive values > 10 are expected, i.e. a value `y` < 10 will trigger anomaly score > 1
      max_points_per_query: 5000 # overrides reader-level value of 10000 for `ingestion_rate` query
      tz: 'America/New_York'  # to override reader-wise `tz`
      tenant_id: '2:0'  # overriding tenant_id to isolate data
      offset: '-15s'  # to override reader-wise `offset` and query data 15 seconds earlier to account for data collection delays
    

Config parameters #

      reader:
  class: "vm"  # or "reader.vm.VmReader" until v1.13.0
  datasource_url: "https://play.victoriametrics.com/"
  tenant_id: '0:0'
  tz: 'America/New_York'
  data_range: [1, 'inf']  # reader-level
  offset: '0s'  # reader-level
  queries:
    ingestion_rate:
      expr: 'sum(rate(vm_rows_inserted_total[5m])) by (type) > 0'
      step: '1m' # can override reader-level `sampling_period` on per-query level
      data_range: [0, 'inf']  # if set, overrides reader-level data_range
      tz: 'Australia/Sydney'  # if set, overrides reader-level tz
      # tenant_id: '1:0'  # if set, overrides reader-level tenant_id
      # offset: '-15s'  # if set, overrides reader-level offset
  sampling_period: '1m'
  query_from_last_seen_timestamp: True  # false by default
  latency_offset: '1ms'
    

mTLS protection #

vmanomaly supports mutual TLS (mTLS) Available from v1.16.3 for secure communication across its components, including VmReader , VmWriter , and Monitoring/Push . This allows for mutual authentication between the client and server when querying or writing data to VictoriaMetrics Enterprise, configured for mTLS .

mTLS ensures that both the client and server verify each other’s identity using certificates, which enhances security by preventing unauthorized access.

To configure mTLS, the following parameters can be set in the config :

• verify_tls: If set to a string, it functions like the -mtlsCAFile command-line argument of VictoriaMetrics, specifying the CA bundle to use. Set to True to use the system’s default certificate store.
• tls_cert_file: Specifies the path to the client certificate, analogous to the -tlsCertFile argument of VictoriaMetrics.
• tls_key_file: Specifies the path to the client certificate key, similar to the -tlsKeyFile argument of VictoriaMetrics.

These options allow you to securely interact with mTLS-enabled VictoriaMetrics endpoints.

Example configuration to enable mTLS with custom certificates:

      reader:
  class: "vm"
  datasource_url: "https://your-victoriametrics-instance-with-mtls"
  # tenant_id: "0:0" uncomment and set for cluster version
  queries:
    vm_blocks_example:
      expr: 'avg(rate(vm_blocks[5m]))'
      step: 30s
  sampling_period: 30s
  verify_tls: "path/to/ca.crt"  # path to CA bundle for TLS verification
  tls_cert_file: "path/to/client.crt"  # path to the client certificate
  tls_key_file:  "path/to/client.key"  # path to the client certificate key
  # additional reader parameters ...

# other config sections, like models, schedulers, writer, ...
    

Healthcheck metrics #

VmReader exposes several healthchecks metrics .

VictoriaLogs reader #

Available from v1.26.0 vmanomaly adds support for reading data from VictoriaLogs stats queries endpoint with VLogsReader. This reader allows quering and analyzing log data stored in VictoriaLogs, enabling anomaly detection on metrics generated from logs.

Its queries should be expressed in a subset of LogsQL , which is similar to MetricsQL/PromQL but adapted for log data.

Valid stats functions #

VLogsReader relies on stats pipe functions that return numeric values, which can be used for anomaly detection on timeseries (metrics). The future addition of similar stats functions in VictoriaLogs will be supported automatically, as long as they return numeric values.

The supported stats functions currently include:

• avg - returns the average value over the given numeric log fields .
• count - returns the number of log entries.
• count_empty - returns the number logs with empty log fields .
• count_uniq - returns the number of unique non-empty values for the given log fields .
• count_uniq_hash - returns the number of unique hashes for non-empty values at the given log fields .
• max - returns the maximum value over the given log fields .
• median - returns the median value over the given log fields .
• min - returns the minimum value over the given log fields .
• quantile - returns the given quantile for the given log fields .
• rate - returns the average per-second rate of matching logs on the selected time range.
• rate_sum - returns the average per-second rate of sum for the given log fields .
• sum - returns the sum for the given numeric log fields .
• sum_len - returns the sum of lengths for the given log fields .

Query Examples #

Here are examples of simple valid LogsQL queries with stats pipe functions that can be used with VLogsReader.

The following query returns the average value for the duration field over logs matching the filter for error word :

      error | stats avg(duration) as avg_error_duration
    

It is possible to calculate the average over fields with common prefix via avg(prefix*) syntax. For example, the following query calculates the number of logs with foo prefix having error word :

      error | stats count(foo*) as foo_error_count
    

Config parameters #

Per-query parameters #

The names, types and the logic of the per-query parameters subset used in VLogsReader are exactly the same as those of VmReader , please see per-query parameters section above for the details. The only difference is that expr parameter should contain a valid LogsQL expression with stats pipe , as described in query examples section above.

Per-query config example #

      reader:
  class: 'vlogs'  # or 'reader.vlogs.VLogsReader'
  # don't include /select/stats_query_range part in the URL, it is added automatically
  datasource_url: 'https://play-vmlogs.victoriametrics.com/'  # source victorialogs
  # tenant_id: '0:0'  # for cluster version only
  sampling_period: '1m'
  max_points_per_query: 10000
  data_range: [0, 'inf']  # reader-level
  offset: '0s'  # reader-level
  timeout: '30s'
  queries:
    # one query returning 1 result fields (avg_duration), it will have __name__ label (series name) as `duration_30m__avg`
    duration_avg_30m:
      expr: "* | stats avg(duration) as avg"  # initial LogsQL expression
      step: '2m'  # overrides global `sampling_period` of 1m
      data_range: [0, 'inf']  # meaning only positive values > 0 are expected, i.e. a value `y` < 0 will trigger anomaly score > 1
      tz: 'America/New_York'  # to override reader-wise `tz`
      # tenant_id: '1:0'  # overriding tenant_id to isolate data
      # offset: '-15s'  # to override reader-wise `offset` and query data 15 seconds earlier to account for data collection delays
      # max_points_per_query: 5000 # overrides reader-level value of 10000 for `avg_duration` query

    # one query returning 3 result fields (p50, p90, p99), they will have __name__ label (series name) as
    # `duration_quantiles_30m__p50`, `duration_quantiles_30m__p90`, `duration_quantiles_30m__p99`, respectively
    duration_quantiles_30m:
      expr: |
            * | stats
                quantile(0.5, request_duration_seconds) p50,
                quantile(0.9, request_duration_seconds) p90,
                quantile(0.99, request_duration_seconds) p99
      step: '2m'  # overrides global `sampling_period` of 1m
      # other per-query parameters as needed
  # other reader-level parameters as needed
    
# other config sections, like models, schedulers, writer, ...
    

mTLS protection #

Please refer to the mTLS protection section above for details on how to configure mTLS for VLogsReader. It uses the same config parameters as VmReader for mTLS setup.

Healthcheck metrics #

Similarly to VmReader, VLogsReader also exposes several healthchecks metrics .