VictoriaMetrics/docs/MetricsQL.md
Dmytro Kozlov 86d85591a4
vmselect/promql: fix panic in histogram_quantiles function (#3029)
* vmselect/promql: fix panic in histogram_quantiles function

* Update docs/MetricsQL.md

Co-authored-by: Roman Khavronenko <roman@victoriametrics.com>

Co-authored-by: Aliaksandr Valialkin <valyala@victoriametrics.com>
Co-authored-by: Roman Khavronenko <roman@victoriametrics.com>
2022-08-30 11:16:13 +03:00

91 KiB

sort
14

MetricsQL

VictoriaMetrics implements MetricsQL - query language inspired by PromQL. MetricsQL is backwards-compatible with PromQL, so Grafana dashboards backed by Prometheus datasource should work the same after switching from Prometheus to VictoriaMetrics. However, there are some intentional differences between these two languages.

Standalone MetricsQL package can be used for parsing MetricsQL in external apps.

If you are unfamiliar with PromQL, then it is suggested reading this tutorial for beginners.

The following functionality is implemented differently in MetricsQL compared to PromQL. This improves user experience:

  • MetricsQL takes into account the previous point before the window in square brackets for range functions such as rate and increase. This allows returning the exact results users expect for increase(metric[$__interval]) queries instead of incomplete results Prometheus returns for such queries.
  • MetricsQL doesn't extrapolate range function results. This addresses this issue from Prometheus. See technical details about VictoriaMetrics and Prometheus calculations for rate and increase in this issue.
  • MetricsQL returns the expected non-empty responses for rate with step values smaller than scrape interval. This addresses this issue from Grafana. See also this blog post.
  • MetricsQL treats scalar type the same as instant vector without labels, since subtle differences between these types usually confuse users. See the corresponding Prometheus docs for details.
  • MetricsQL removes all the NaN values from the output, so some queries like (-1)^0.5 return empty results in VictoriaMetrics, while returning a series of NaN values in Prometheus. Note that Grafana doesn't draw any lines or dots for NaN values, so the end result looks the same for both VictoriaMetrics and Prometheus.
  • MetricsQL keeps metric names after applying functions, which don't change the meaning of the original time series. For example, min_over_time(foo) or round(foo) leaves foo metric name in the result. See this issue for details.

Read more about the diffferences between PromQL and MetricsQL in this article.

Other PromQL functionality should work the same in MetricsQL. File an issue if you notice discrepancies between PromQL and MetricsQL results other than mentioned above.

MetricsQL features

MetricsQL implements PromQL and provides additional functionality mentioned below, which is aimed towards solving practical cases. Feel free filing a feature request if you think MetricsQL misses certain useful functionality.

This functionality can be evaluated at an editable Grafana dashboard or at your own VictoriaMetrics instance.

  • Graphite-compatible filters can be passed via {__graphite__="foo.*.bar"} syntax. See these docs. VictoriaMetrics also can be used as Graphite datasource in Grafana. See these docs for details. See also label_graphite_group function, which can be used for extracting the given groups from Graphite metric name.
  • Lookbehind window in square brackets may be omitted. VictoriaMetrics automatically selects the lookbehind window depending on the current step used for building the graph (e.g. step query arg passed to /api/v1/query_range). For instance, the following query is valid in VictoriaMetrics: rate(node_network_receive_bytes_total). It is equivalent to rate(node_network_receive_bytes_total[$__interval]) when used in Grafana.
  • Aggregate functions accept arbitrary number of args. For example, avg(q1, q2, q3) would return the average values for every point across time series returned by q1, q2 and q3.
  • @ modifier can be put anywhere in the query. For example, sum(foo) @ end() calculates sum(foo) at the end timestamp of the selected time range [start ... end].
  • Arbitrary subexpression can be used as @ modifier. For example, foo @ (end() - 1h) calculates foo at the end - 1 hour timestamp on the selected time range [start ... end].
  • offset, lookbehind window in square brackets and step value for subquery may refer to the current step aka $__interval value from Grafana with [Ni] syntax. For instance, rate(metric[10i] offset 5i) would return per-second rate over a range covering 10 previous steps with the offset of 5 steps.
  • offset may be put anywere in the query. For instance, sum(foo) offset 24h.
  • Lookbehind window in square brackets and offset may be fractional. For instance, rate(node_network_receive_bytes_total[1.5m] offset 0.5d).
  • The duration suffix is optional. The duration is in seconds if the suffix is missing. For example, rate(m[300] offset 1800) is equivalent to rate(m[5m]) offset 30m.
  • The duration can be placed anywhere in the query. For example, sum_over_time(m[1h]) / 1h is equivalent to sum_over_time(m[1h]) / 3600.
  • Trailing commas on all the lists are allowed - label filters, function args and with expressions. For instance, the following queries are valid: m{foo="bar",}, f(a, b,), WITH (x=y,) x. This simplifies maintenance of multi-line queries.
  • Metric names and metric labels may contain escaped chars. For instance, foo\-bar{baz\=aa="b"} is valid expression. It returns time series with name foo-bar containing label baz=aa with value b. Additionally, \xXX escape sequence is supported, where XX is hexadecimal representation of escaped char.
  • Aggregate functions support optional limit N suffix in order to limit the number of output series. For example, sum(x) by (y) limit 3 limits the number of output time series after the aggregation to 3. All the other time series are dropped.
  • histogram_quantile accepts optional third arg - boundsLabel. In this case it returns lower and upper bounds for the estimated percentile. See this issue for details.
  • default binary operator. q1 default q2 fills gaps in q1 with the corresponding values from q2.
  • if binary operator. q1 if q2 removes values from q1 for missing values from q2.
  • ifnot binary operator. q1 ifnot q2 removes values from q1 for existing values from q2.
  • String literals may be concatenated. This is useful with WITH templates: WITH (commonPrefix="long_metric_prefix_") {__name__=commonPrefix+"suffix1"} / {__name__=commonPrefix+"suffix2"}.
  • WITH templates. This feature simplifies writing and managing complex queries. Go to WITH templates playground and try it.
  • keep_metric_names modifier can be applied to all the rollup functions and transform functions. This modifier prevents from dropping metric names in function results. See these docs.

keep_metric_names

By default metric names are dropped after applying functions, which change the meaning of the original time series. This may result in duplicate time series error when the function is applied to multiple time series with different names. This error can be fixed by applying keep_metric_names modifier to the function. For example, rate({__name__=~"foo|bar"}) keep_metric_names leaves foo and bar metric names in the returned time series.

MetricsQL functions

If you are unfamiliar with PromQL, then please read this tutorial at first.

MetricsQL provides the following functions:

Rollup functions

Rollup functions (aka range functions or window functions) calculate rollups over raw samples on the given lookbehind window for the selected time series. For example, avg_over_time(temperature[24h]) calculates the average temperature over raw samples for the last 24 hours. Additional details:

  • If rollup functions are used for building graphs in Grafana, then the rollup is calculated independently per each point on the graph. For example, every point for avg_over_time(temperature[24h]) graph shows the average temperature for the last 24 hours ending at this point. The interval between points is set as step query arg passed by Grafana to /api/v1/query_range.
  • If the given series selector returns multiple time series, then rollups are calculated individually per each returned series.
  • If lookbehind window in square brackets is missing, then MetricsQL automatically sets the lookbehind window to the interval between points on the graph (aka step query arg at /api/v1/query_range, $__interval value from Grafana or 1i duration in MetricsQL). For example, rate(http_requests_total) is equivalent to rate(http_requests_total[$__interval]) in Grafana. It is also equivalent to rate(http_requests_total[1i]).
  • Every series selector in MetricsQL must be wrapped into a rollup function. Otherwise it is automatically wrapped into default_rollup. For example, foo{bar="baz"} is automatically converted to default_rollup(foo{bar="baz"}[1i]) before performing the calculations.
  • If something other than series selector is passed to rollup function, then the inner arg is automatically converted to a subquery.
  • All the rollup functions accept optional keep_metric_names modifier. If it is set, then the function keeps metric names in results. See these docs.

See also implicit query conversions.

absent_over_time

absent_over_time(series_selector[d]) returns 1 if the given lookbehind window d doesn't contain raw samples. Otherwise it returns an empty result. This function is supported by PromQL. See also present_over_time.

aggr_over_time

aggr_over_time(("rollup_func1", "rollup_func2", ...), series_selector[d]) calculates all the listed rollup_func* for raw samples on the given lookbehind window d. The calculations are perfomed individually per each time series returned from the given series_selector. rollup_func* can contain any rollup function. For instance, aggr_over_time(("min_over_time", "max_over_time", "rate"), m[d]) would calculate min_over_time, max_over_time and rate for m[d].

ascent_over_time

ascent_over_time(series_selector[d]) calculates ascent of raw sample values on the given lookbehind window d. The calculations are performed individually per each time series returned from the given series_selector. Useful for tracking height gains in GPS tracking. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also descent_over_time.

avg_over_time

avg_over_time(series_selector[d]) calculates the average value over raw samples on the given lookbehind window d per each time series returned from the given series_selector. This function is supported by PromQL. See also median_over_time.

changes

changes(series_selector[d]) calculates the number of times the raw samples changed on the given lookbehind window d per each time series returned from the given series_selector. Unlike changes() in Prometheus it takes into account the change from the last sample before the given lookbehind window d. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also changes_prometheus.

changes_prometheus

changes_prometheus(series_selector[d]) calculates the number of times the raw samples changed on the given lookbehind window d per each time series returned from the given series_selector. It doesn't take into account the change from the last sample before the given lookbehind window d in the same way as Prometheus does. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also changes.

count_eq_over_time

count_eq_over_time(series_selector[d], eq) calculates the number of raw samples on the given lookbehind window d, which are equal to eq. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also count_over_time.

count_gt_over_time

count_gt_over_time(series_selector[d], gt) calculates the number of raw samples on the given lookbehind window d, which are bigger than gt. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also count_over_time.

count_le_over_time

count_le_over_time(series_selector[d], le) calculates the number of raw samples on the given lookbehind window d, which don't exceed le. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also count_over_time.

count_ne_over_time

count_ne_over_time(series_selector[d], ne) calculates the number of raw samples on the given lookbehind window d, which aren't equal to ne. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also count_over_time.

count_over_time

count_over_time(series_selector[d]) calculates the number of raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also count_le_over_time, count_gt_over_time, count_eq_over_time and count_ne_over_time.

decreases_over_time

decreases_over_time(series_selector[d]) calculates the number of raw sample value decreases over the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also increases_over_time.

default_rollup

default_rollup(series_selector[d]) returns the last raw sample value on the given lookbehind window d per each time series returned from the given series_selector.

delta

delta(series_selector[d]) calculates the difference between the last sample before the given lookbehind window d and the last sample at the given lookbehind window d per each time series returned from the given series_selector. The behaviour of delta() function in MetricsQL is slighly different to the behaviour of delta() function in Prometheus. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also increase and delta_prometheus.

delta_prometheus

delta_prometheus(series_selector[d]) calculates the difference between the first and the last samples at the given lookbehind window d per each time series returned from the given series_selector. The behaviour of delta_prometheus() is close to the behaviour of delta() function in Prometheus. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also delta.

deriv

deriv(series_selector[d]) calculates per-second derivative over the given lookbehind window d per each time series returned from the given series_selector. The derivative is calculated using linear regression. Metric names are stripped from the resulting rollups. This function is supported by PromQL. Add keep_metric_names modifier in order to keep metric names. See also deriv_fast and ideriv.

deriv_fast

deriv_fast(series_selector[d]) calculates per-second derivative using the first and the last raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also deriv and ideriv.

descent_over_time

descent_over_time(series_selector[d]) calculates descent of raw sample values on the given lookbehind window d. The calculations are performed individually per each time series returned from the given series_selector. Useful for tracking height loss in GPS tracking. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also ascent_over_time.

distinct_over_time

distinct_over_time(series_selector[d]) returns the number of distinct raw sample values on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

duration_over_time

duration_over_time(series_selector[d], max_interval) returns the duration in seconds when time series returned from the given series_selector were present over the given lookbehind window d. It is expected that intervals between adjacent samples per each series don't exceed the max_interval. Otherwise such intervals are considered as gaps and aren't counted. See also lifetime and lag.

first_over_time

first_over_time(series_selector[d]) returns the first raw sample value on the given lookbehind window d per each time series returned from the given series_selector. See also last_over_time and tfirst_over_time.

geomean_over_time

geomean_over_time(series_selector[d]) calculates geometric mean over raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

histogram_over_time

histogram_over_time(series_selector[d]) calculates VictoriaMetrics histogram over raw samples on the given lookbehind window d. It is calculated individually per each time series returned from the given series_selector. The resulting histograms are useful to pass to histogram_quantile for calculating quantiles over multiple gauges. For example, the following query calculates median temperature by country over the last 24 hours: histogram_quantile(0.5, sum(histogram_over_time(temperature[24h])) by (vmrange,country)).

hoeffding_bound_lower

hoeffding_bound_lower(phi, series_selector[d]) calculates lower Hoeffding bound for the given phi in the range [0...1]. See also hoeffding_bound_upper.

hoeffding_bound_upper

hoeffding_bound_upper(phi, series_selector[d]) calculates upper Hoeffding bound for the given phi in the range [0...1]. See also hoeffding_bound_lower.

holt_winters

holt_winters(series_selector[d], sf, tf) calculates Holt-Winters value (aka double exponential smoothing) for raw samples over the given lookbehind window d using the given smoothing factor sf and the given trend factor tf. Both sf and tf must be in the range [0...1]. It is expected that the series_selector returns time series of gauge type. This function is supported by PromQL.

idelta

idelta(series_selector[d]) calculates the difference between the last two raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

ideriv

ideriv(series_selector[d]) calculates the per-second derivative based on the last two raw samples over the given lookbehind window d. The derivative is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also deriv.

increase

increase(series_selector[d]) calculates the increase over the given lookbehind window d per each time series returned from the given series_selector. It is expected that the series_selector returns time series of counter type. Unlike Prometheus it takes into account the last sample before the given lookbehind window d when calculating the result. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also increase_pure, increase_prometheus and delta.

increase_prometheus

increase_prometheus(series_selector[d]) calculates the increase over the given lookbehind window d per each time series returned from the given series_selector. It is expected that the series_selector returns time series of counter type. It doesn't take into account the last sample before the given lookbehind window d when calculating the result in the same way as Prometheus does. See this article for details. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also increase_pure and increase.

increase_pure

increase_pure(series_selector[d]) works the same as increase except of the following corner case - it assumes that counters always start from 0, while increase ignores the first value in a series if it is too big.

increases_over_time

increases_over_time(series_selector[d]) calculates the number of raw sample value increases over the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also decreases_over_time.

integrate

integrate(series_selector[d]) calculates the integral over raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

irate

irate(series_selector[d]) calculates the "instant" per-second increase rate over the last two raw samples on the given lookbehind window d per each time series returned from the given series_selector. It is expected that the series_selector returns time series of counter type. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also rate.

lag

lag(series_selector[d]) returns the duration in seconds between the last sample on the given lookbehind window d and the timestamp of the current point. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also lifetime and duration_over_time.

last_over_time

last_over_time(series_selector[d]) returns the last raw sample value on the given lookbehind window d per each time series returned from the given series_selector. This function is supported by PromQL. See also first_over_time and tlast_over_time.

lifetime

lifetime(series_selector[d]) returns the duration in seconds between the last and the first sample on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also duration_over_time and lag.

max_over_time

max_over_time(series_selector[d]) calculates the maximum value over raw samples on the given lookbehind window d per each time series returned from the given series_selector. This function is supported by PromQL. See also tmax_over_time.

median_over_time

median_over_time(series_selector[d]) calculates median value over raw samples on the given lookbehind window d per each time series returned from the given series_selector. See also avg_over_time.

min_over_time

min_over_time(series_selector[d]) calculates the minimum value over raw samples on the given lookbehind window d per each time series returned from the given series_selector. This function is supported by PromQL. See also tmin_over_time.

mode_over_time

mode_over_time(series_selector[d]) calculates mode for raw samples on the given lookbehind window d. It is calculated individually per each time series returned from the given series_selector. It is expected that raw sample values are discrete.

predict_linear

predict_linear(series_selector[d], t) calculates the value t seconds in the future using linear interpolation over raw samples on the given lookbehind window d. The predicted value is calculated individually per each time series returned from the given series_selector. This function is supported by PromQL.

present_over_time

present_over_time(series_selector[d]) returns 1 if there is at least a single raw sample on the given lookbehind window d. Otherwise an empty result is returned. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

quantile_over_time

quantile_over_time(phi, series_selector[d]) calculates phi-quantile over raw samples on the given lookbehind window d per each time series returned from the given series_selector. The phi value must be in the range [0...1]. This function is supported by PromQL. See also quantiles_over_time.

quantiles_over_time

quantiles_over_time("phiLabel", phi1, ..., phiN, series_selector[d]) calculates phi*-quantiles over raw samples on the given lookbehind window d per each time series returned from the given series_selector. The function returns individual series per each phi* with {phiLabel="phi*"} label. phi* values must be in the range [0...1]. See also quantile_over_time.

range_over_time

range_over_time(series_selector[d]) calculates value range over raw samples on the given lookbehind window d per each time series returned from the given series_selector. E.g. it calculates max_over_time(series_selector[d]) - min_over_time(series_selector[d]). Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

rate

rate(series_selector[d]) calculates the average per-second increase rate over the given lookbehind window d per each time series returned from the given series_selector. It is expected that the series_selector returns time series of counter type. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

rate_over_sum

rate_over_sum(series_selector[d]) calculates per-second rate over the sum of raw samples on the given lookbehind window d. The calculations are performed indiviually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

resets

resets(series_selector[d]) returns the number of counter resets over the given lookbehind window d per each time series returned from the given series_selector. It is expected that the series_selector returns time series of counter type. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

rollup

rollup(series_selector[d]) calculates min, max and avg values for raw samples on the given lookbehind window d. These values are calculated individually per each time series returned from the given series_selector.

rollup_candlestick

rollup_candlestick(series_selector[d]) calculates open, high, low and close values (aka OHLC) over raw samples on the given lookbehind window d. The calculations are perfomed individually per each time series returned from the given series_selector. This function is useful for financial applications.

rollup_delta

rollup_delta(series_selector[d]) calculates differences between adjancent raw samples on the given lookbehind window d and returns min, max and avg values for the calculated differences. The calculations are performed individually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also rollup_increase.

rollup_deriv

rollup_deriv(series_selector[d]) calculates per-second derivatives for adjancent raw samples on the given lookbehind window d and returns min, max and avg values for the calculated per-second derivatives. The calculations are performed individually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

rollup_increase

rollup_increase(series_selector[d]) calculates increases for adjancent raw samples on the given lookbehind window d and returns min, max and avg values for the calculated increases. The calculations are performed individually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also rollup_delta.

rollup_rate

rollup_rate(series_selector[d]) calculates per-second change rates for adjancent raw samples on the given lookbehind window d and returns min, max and avg values for the calculated per-second change rates. The calculations are perfomed individually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

rollup_scrape_interval

rollup_scrape_interval(series_selector[d]) calculates the interval in seconds between adjancent raw samples on the given lookbehind window d and returns min, max and avg values for the calculated interval. The calculations are perfomed individually per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also scrape_interval.

scrape_interval

scrape_interval(series_selector[d]) calculates the average interval in seconds between raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also rollup_scrape_interval.

share_gt_over_time

share_gt_over_time(series_selector[d], gt) returns share (in the range [0...1]) of raw samples on the given lookbehind window d, which are bigger than gt. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. Useful for calculating SLI and SLO. Example: share_gt_over_time(up[24h], 0) - returns service availability for the last 24 hours. See also share_le_over_time.

share_le_over_time

share_le_over_time(series_selector[d], le) returns share (in the range [0...1]) of raw samples on the given lookbehind window d, which are smaller or equal to le. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. Useful for calculating SLI and SLO. Example: share_le_over_time(memory_usage_bytes[24h], 100*1024*1024) returns the share of time series values for the last 24 hours when memory usage was below or equal to 100MB. See also share_gt_over_time.

stale_samples_over_time

stale_samples_over_time(series_selector[d]) calculates the number of staleness markers on the given lookbehind window d per each time series matching the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

stddev_over_time

stddev_over_time(series_selector[d]) calculates standard deviation over raw samples on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also stdvar_over_time.

stdvar_over_time

stdvar_over_time(series_selector[d]) calculates stadnard variance over raw samples on the given lookbheind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also stddev_over_time.

sum_over_time

sum_over_time(series_selector[d]) calculates the sum of raw sample values on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

sum2_over_time

sum2_over_time(series_selector[d]) calculates the sum of squares for raw sample values on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

timestamp

timestamp(series_selector[d]) returns the timestamp in seconds for the last raw sample on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also timestamp_with_name.

timestamp_with_name

timestamp_with_name(series_selector[d]) returns the timestamp in seconds for the last raw sample on the given lookbehind window d per each time series returned from the given series_selector. Metric names are preserved in the resulting rollups. See also timestamp.

tfirst_over_time

tfirst_over_time(series_selector[d]) returns the timestamp in seconds for the first raw sample on the given lookbehind window d per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also first_over_time.

tlast_change_over_time

tlast_change_over_time(series_selector[d]) returns the timestamp in seconds for the last change per each time series returned from the given series_selector on the given lookbehind window d. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also last_over_time.

tlast_over_time

tlast_over_time(series_selector[d]) is an alias for timestamp. See also tlast_change_over_time.

tmax_over_time

tmax_over_time(series_selector[d]) returns the timestamp in seconds for the raw sample with the maximum value on the given lookbehind window d. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also max_over_time.

tmin_over_time

tmin_over_time(series_selector[d]) returns the timestamp in seconds for the raw sample with the minimum value on the given lookbehind window d. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names. See also min_over_time.

zscore_over_time

zscore_over_time(series_selector[d]) calculates returns z-score for raw samples on the given lookbehind window d. It is calculated independently per each time series returned from the given series_selector. Metric names are stripped from the resulting rollups. Add keep_metric_names modifier in order to keep metric names.

Transform functions

Transform functions calculate transformations over rollup results. For example, abs(delta(temperature[24h])) calculates the absolute value for every point of every time series returned from the rollup delta(temperature[24h]). Additional details:

  • If transform function is applied directly to a series selector, then the default_rollup() function is automatically applied before calculating the transformations. For example, abs(temperature) is implicitly transformed to abs(default_rollup(temperature[1i])).
  • All the transform functions accept optional keep_metric_names modifier. If it is set, then the function doesn't drop metric names from the resulting time series. See these docs.

See also implicit query conversions.

abs

abs(q) calculates the absolute value for every point of every time series returned by q. This function is supported by PromQL.

absent

absent(q) returns 1 if q has no points. Otherwise returns an empty result. This function is supported by PromQL.

acos

acos(q) returns inverse cosine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also asin and cos.

acosh

acosh(q) returns inverse hyperbolic cosine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also sinh.

asin

asin(q) returns inverse sine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also acos and sin.

asinh

asinh(q) returns inverse hyperbolic sine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also sinh.

atan

atan(q) returns inverse tangent for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also tan.

atanh

atanh(q) returns inverse hyperbolic tangent for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also tanh.

bitmap_and

bitmap_and(q, mask) - calculates bitwise v & mask for every v point of every time series returned from q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names.

bitmap_or

bitmap_or(q, mask) calculates bitwise v | mask for every v point of every time series returned from q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names.

bitmap_xor

bitmap_xor(q, mask) calculates bitwise v ^ mask for every v point of every time series returned from q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names.

buckets_limit

buckets_limit(limit, buckets) limits the number of histogram buckets to the given limit. See also prometheus_buckets and histogram_quantile.

ceil

ceil(q) rounds every point for every time series returned by q to the upper nearest integer. This function is supported by PromQL. See also floor and round.

clamp

clamp(q, min, max) clamps every point for every time series returned by q with the given min and max values. This function is supported by PromQL. See also clamp_min and clamp_max.

clamp_max

clamp_max(q, max) clamps every point for every time series returned by q with the given max value. This function is supported by PromQL. See also clamp and clamp_min.

clamp_min

clamp_min(q, min) clamps every pount for every time series returned by q with the given min value. This function is supported by PromQL. See also clamp and clamp_max.

cos

cos(q) returns cos(v) for every v point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also sin.

cosh

cosh(q) returns hyperbolic cosine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. This function is supported by PromQL. See also acosh.

day_of_month

day_of_month(q) returns the day of month for every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [1...31]. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

day_of_week

day_of_week(q) returns the day of week for every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [0...6], where 0 means Sunday and 6 means Saturday. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

days_in_month

days_in_month(q) returns the number of days in the month identified by every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [28...31]. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

deg

deg(q) converts Radians to degrees for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also rad.

end

end() returns the unix timestamp in seconds for the last point. See also start. It is known as end query arg passed to /api/v1/query_range.

exp

exp(q) calculates the e^v for every point v of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. See also ln. This function is supported by PromQL.

floor

floor(q) rounds every point for every time series returned by q to the lower nearest integer. See also ceil and round. This function is supported by PromQL.

histogram_avg

histogram_avg(buckets) calculates the average value for the given buckets. It can be used for calculating the average over the given time range across multiple time series. For exmple, histogram_avg(sum(histogram_over_time(response_time_duration_seconds[5m])) by (vmrange,job)) would return the average response time per each job over the last 5 minutes.

histogram_quantile

histogram_quantile(phi, buckets) calculates phi-quantile over the given histogram buckets. phi must be in the range [0...1]. For example, histogram_quantile(0.5, sum(rate(http_request_duration_seconds_bucket[5m]) by (le)) would return median request duration for all the requests during the last 5 minutes. It accepts optional third arg - boundsLabel. In this case it returns lower and upper bounds for the estimated percentile. See this issue for details. This function is supported by PromQL (except of the boundLabel arg). See also histogram_quantiles and histogram_share.

histogram_quantiles

histogram_quantiles("phiLabel", phi1, ..., phiN, buckets) calculates the given phi*-quantiles over the given histogram buckets. Argument phi* must be in the range [0...1]. For example, histogram_quantiles('le', 0.3, 0.5, sum(rate(http_request_duration_seconds_bucket[5m]) by (le)). Each calculated quantile is returned in a separate time series with the corresponding {phiLabel="phi*"} label. See also histogram_quantile.

histogram_share

histogram_share(le, buckets) calculates the share (in the range [0...1]) for buckets that fall below le. Useful for calculating SLI and SLO. This is inverse to histogram_quantile.

histogram_stddev

histogram_stddev(buckets) calculates standard deviation for the given buckets.

histogram_stdvar

histogram_stdvar(buckets) calculates standard variance for the given buckets. It can be used for calculating standard deviation over the given time range across multiple time series. For example, histogram_stdvar(sum(histogram_over_time(temperature[24])) by (vmrange,country)) would return standard deviation for the temperature per each country over the last 24 hours.

hour

hour(q) returns the hour for every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [0...23]. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

interpolate

interpolate(q) fills gaps with linearly interpolated values calculated from the last and the next non-empty points per each time series returned by q. See also keep_last_value and keep_next_value.

keep_last_value

keep_last_value(q) fills gaps with the value of the last non-empty point in every time series returned by q. See also keep_next_value and interpolate.

keep_next_value

keep_next_value(q) fills gaps with the value of the next non-empty point in every time series returned by q. See also keep_last_value and interpolate.

limit_offset

limit_offset(limit, offset, q) skips offset time series from series returned by q and then returns up to limit of the remaining time series per each group. This allows implementing simple paging for q time series. See also limitk.

ln

ln(q) calculates ln(v) for every point v of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also exp and log2.

log2

log2(q) calculates log2(v) for every point v of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also log10 and ln.

log10

log10(q) calculates log10(v) for every point v of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also log2 and ln.

minute

minute(q) returns the minute for every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [0...59]. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

month

month(q) returns the month for every point of every time series returned by q. It is expected that q returns unix timestamps. The returned values are in the range [1...12], where 1 means January and 12 means December. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

now

now() returns the current timestamp as a floating-point value in seconds. See also time.

pi

pi() returns Pi number. This function is supported by PromQL.

rad

rad(q) converts degrees to Radians for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL. See also deg.

prometheus_buckets

prometheus_buckets(buckets) converts VictoriaMetrics histogram buckets with vmrange labels to Prometheus histogram buckets with le labels. This may be useful for building heatmaps in Grafana. See also histogram_quantile and buckets_limit.

rand

rand(seed) returns pseudo-random numbers on the range [0...1] with even distribution. Optional seed can be used as a seed for pseudo-random number generator. See also rand_normal and rand_exponential.

rand_exponential

rand_exponential(seed) returns pseudo-random numbers with exponential distribution. Optional seed can be used as a seed for pseudo-random number generator. See also rand and rand_normal.

rand_normal

rand_normal(seed) returns pesudo-random numbers with normal distribution. Optional seed can be used as a seed for pseudo-random number generator. See also rand and rand_exponential.

range_avg

range_avg(q) calculates the avg value across points per each time series returned by q.

range_first

range_first(q) returns the value for the first point per each time series returned by q.

range_last

range_last(q) returns the value for the last point per each time series returned by q.

range_max

range_max(q) calculates the max value across points per each time series returned by q.

range_median

range_median(q) calculates the median value across points per each time series returned by q.

range_min

range_min(q) calculates the min value across points per each time series returned by q.

range_quantile

range_quantile(phi, q) returns phi-quantile across points per each time series returned by q. phi must be in the range [0...1].

range_sum

range_sum(q) calculates the sum of points per each time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names.

remove_resets

remove_resets(q) removes counter resets from time series returned by q.

round

round(q, nearest) round every point of every time series returned by q to the nearest multiple. If nearest is missing then the rounding is performed to the nearest integer. This function is supported by PromQL. See also floor and ceil.

ru

ru(free, max) calculates resource utilization in the range [0%...100%] for the given free and max resources. For instance, ru(node_memory_MemFree_bytes, node_memory_MemTotal_bytes) returns memory utilization over node_exporter metrics.

running_avg

running_avg(q) calculates the running avg per each time series returned by q.

running_max

running_max(q) calculates the running max per each time series returned by q.

running_min

running_min(q) calculates the running min per each time series returned by q.

running_sum

running_sum(q) calculates the running sum per each time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names.

scalar

scalar(q) returns q if q contains only a single time series. Otherwise it returns nothing. This function is supported by PromQL.

sgn

sgn(q) returns 1 if v>0, -1 if v<0 and 0 if v==0 for every point v of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

sin

sin(q) returns sin(v) for every v point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by MetricsQL. See also cos.

sinh

sinh(q) returns hyperbolic sine for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by MetricsQL. See also cosh.

tan

tan(q) returns tan(v) for every v point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by MetricsQL. See also atan.

tanh

tanh(q) returns hyperbolic tangent for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by MetricsQL. See also atanh.

smooth_exponential

smooth_exponential(q, sf) smooths points per each time series returned by q using exponential moving average with the given smooth factor sf.

sort

sort(q) sorts series in ascending order by the last point in every time series returned by q. This function is supported by PromQL. See also sort_desc.

sort_by_label

sort_by_label(q, label1, ... labelN) sorts series in ascending order by the given set of labels. For example, sort_by_label(foo, "bar") would sort foo series by values of the label bar in these series. See also sort_by_label_desc.

sort_by_label_desc

sort_by_label_desc(q, label1, ... labelN) sorts series in descending order by the given set of labels. For example, sort_by_label(foo, "bar") would sort foo series by values of the label bar in these series. See also sort_by_label.

sort_desc

sort_desc(q) sorts series in descending order by the last point in every time series returned by q. This function is supported by PromQL. See also sort.

sqrt

sqrt(q) calculates square root for every point of every time series returned by q. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

start

start() returns unix timestamp in seconds for the first point. See also end. It is known as start query arg passed to /api/v1/query_range.

step

step() returns the step in seconds (aka interval) between the returned points. It is known as step query arg passed to /api/v1/query_range.

time

time() returns unix timestamp for every returned point. This function is supported by PromQL.

timezone_offset

timezone_offset(tz) returns offset in seconds for the given timezone tz relative to UTC. This can be useful when combining with datetime-related functions. For example, day_of_week(time()+timezone_offset("America/Los_Angeles")) would return weekdays for America/Los_Angeles time zone. Special Local time zone can be used for returning an offset for the time zone set on the host where VictoriaMetrics runs. See the list of supported timezones.

ttf

ttf(free) estimates the time in seconds needed to exhaust free resources. For instance, ttf(node_filesystem_avail_byte) returns the time to storage space exhaustion. This function may be useful for capacity planning.

union

union(q1, ..., qN) returns a union of time series returned from q1, ..., qN. The union function name can be skipped - the following queries are quivalent: union(q1, q2) and (q1, q2). It is expected that each q* query returns time series with unique sets of labels. Otherwise only the first time series out of series with identical set of labels is returned. Use alias and label_set functions for giving unique labelsets per each q* query:

vector

vector(q) returns q, e.g. it does nothing in MetricsQL. This function is supported by PromQL.

year

year(q) returns the year for every point of every time series returned by q. It is expected that q returns unix timestamps. Metric names are stripped from the resulting series. Add keep_metric_names modifier in order to keep metric names. This function is supported by PromQL.

Label manipulation functions

Label manipulation functions perform manipulations with lables on the selected rollup results. Additional details:

  • If label manipulation function is applied directly to a series_selector, then the default_rollup() function is automatically applied before performing the label transformation. For example, alias(temperature, "foo") is implicitly transformed to alias(default_rollup(temperature[1i]), "foo").

See also implicit query conversions.

alias

alias(q, "name") sets the given name to all the time series returned by q. For example, alias(up, "foobar") would rename up series to foobar series.

drop_common_labels

drop_common_labels(q1, ...., qN) drops common label="value" pairs among time series returned from q1, ..., qN.

label_copy

label_copy(q, "src_label1", "dst_label1", ..., "src_labelN", "dst_labelN") copies label values from src_label* to dst_label* for all the time series returned by q. If src_label is empty, then the corresponding dst_label is left untouched.

label_del

label_del(q, "label1", ..., "labelN") deletes the given label* labels from all the time series returned by q.

label_graphite_group

label_graphite_group(q, groupNum1, ... groupNumN) replaces metric names returned from q with the given Graphite group values concatenated via . char. For example, label_graphite_group({__graphite__="foo*.bar.*"}, 0, 2) would substitute foo<any_value>.bar.<other_value> metric names with foo<any_value>.<other_value>. This function is useful for aggregating Graphite metrics with aggregate functions. For example, the following query would return per-app memory usage:

sum by (__name__) (
    label_graphite_group({__graphite__="app*.host*.memory_usage"}, 0)
)

label_join

label_join(q, "dst_label", "separator", "src_label1", ..., "src_labelN") joins src_label* values with the given separator and stores the result in dst_label. This is performed individually per each time series returned by q. For example, label_join(up{instance="xxx",job="yyy"}, "foo", "-", "instance", "job") would store xxx-yyy label value into foo label. This function is supported by PromQL.

label_keep

label_keep(q, "label1", ..., "labelN") deletes all the labels except of the listed label* labels in all the time series returned by q.

label_lowercase

label_lowercase(q, "label1", ..., "labelN") lowercases values for the given label* labels in all the time series returned by q.

label_map

label_map(q, "label", "src_value1", "dst_value1", ..., "src_valueN", "dst_valueN") maps label values from src_* to dst* for all the time seires returned by q.

label_match

label_match(q, "label", "regexp") drops time series from q with label not matching the given regexp. This function can be useful after rollup-like functions, which may return multiple time series for every input series. See also label_mismatch.

label_mismatch

label_mismatch(q, "label", "regexp") drops time series from q with label matching the given regexp. This function can be useful after rollup-like functions, which may return multiple time series for every input series. See also label_match.

label_move

label_move(q, "src_label1", "dst_label1", ..., "src_labelN", "dst_labelN") moves label values from src_label* to dst_label* for all the time series returned by q. If src_label is empty, then the corresponding dst_label is left untouched.

label_replace

label_replace(q, "dst_label", "replacement", "src_label", "regex") applies the given regex to src_label and stores the replacement in dst_label if the given regex matches src_label. The replacement may contain references to regex captures such as $1, $2, etc. These references are substituted by the corresponding regex captures. For example, label_replace(up{job="node-exporter"}, "foo", "bar-$1", "job", "node-(.+)") would store bar-exporter label value into foo label. This function is supported by PromQL.

label_set

label_set(q, "label1", "value1", ..., "labelN", "valueN") sets {label1="value1", ..., labelN="valueN"} labels to all the time series returned by q.

label_transform

label_transform(q, "label", "regexp", "replacement") substitutes all the regexp occurences by the given replacement in the given label.

label_uppercase

label_uppercase(q, "label1", ..., "labelN") uppercases values for the given label* labels in all the time series returned by q.

label_value

label_value(q, "label") returns number values for the given label for every time series returned by q. For example, if label_value(foo, "bar") is applied to foo{bar="1.234"}, then it will return a time series foo{bar="1.234"} with 1.234 value.

Aggregate functions

Aggregate functions calculate aggregates over groups of rollup results. Additional details:

  • By default a single group is used for aggregation. Multiple independent groups can be set up by specifying grouping labels in by and without modifiers. For example, count(up) by (job) would group rollup results by job label value and calculate the count aggregate function independently per each group, while count(up) without (instance) would group rollup results by all the labels except instance before calculating count aggregate function independently per each group. Multiple labels can be put in by and without modifiers.
  • If the aggregate function is applied directly to a series_selector, then the default_rollup() function is automatically applied before cacluating the aggregate. For example, count(up) is implicitly transformed to count(default_rollup(up[1i])).
  • Aggregate functions accept arbitrary number of args. For example, avg(q1, q2, q3) would return the average values for every point across time series returned by q1, q2 and q3.
  • Aggregate functions support optional limit N suffix, which can be used for limiting the number of output groups. For example, sum(x) by (y) limit 3 limits the number of groups for the aggregation to 3. All the other groups are ignored.

See also implicit query conversions.

any

any(q) by (group_labels) returns a single series per group_labels out of time series returned by q. See also group.

avg

avg(q) by (group_labels) returns the average value per group_labels for time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

bottomk

bottomk(k, q) returns up to k points with the smallest values across all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL. See also topk.

bottomk_avg

bottomk_avg(k, q, "other_label=other_value") returns up to k time series from q with the smallest averages. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, bottomk_avg(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the smallest averages plus a time series with {job="other"} label with the sum of the remaining series if any. See also topk_avg.

bottomk_last

bottomk_last(k, q, "other_label=other_value") returns up to k time series from q with the smallest last values. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, bottomk_max(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the smallest maximums plus a time series with {job="other"} label with the sum of the remaining series if any. See also topk_last.

bottomk_max

bottomk_max(k, q, "other_label=other_value") returns up to k time series from q with the smallest maximums. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, bottomk_max(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the smallest maximums plus a time series with {job="other"} label with the sum of the remaining series if any. See also topk_max.

bottomk_median

bottomk_median(k, q, "other_label=other_value") returns up to k time series from q with the smallest medians. If an optionalother_label=other_valuearg is set, then the sum of the remaining time series is returned with the given label. For example,bottomk_median(3, sum(process_resident_memory_bytes) by (job), "job=other")would return up to 3 time series with the smallest medians plus a time series with{job="other"}` label with the sum of the remaining series if any. See also topk_median.

bottomk_min

bottomk_min(k, q, "other_label=other_value") returns up to k time series from q with the smallest minimums. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, bottomk_min(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the smallest minimums plus a time series with {job="other"} label with the sum of the remaining series if any. See also topk_min.

count

count(q) by (group_labels) returns the number of non-empty points per group_labels for time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

count_values

count_values("label", q) counts the number of points with the same value and stores the counts in a time series with an additional label, wich contains each initial value. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

distinct

distinct(q) calculates the number of unique values per each group of points with the same timestamp.

geomean

geomean(q) calculates geometric mean per each group of points with the same timestamp.

group

group(q) by (group_labels) returns 1 per each group_labels for time series returned by q. This function is supported by PromQL. See also any.

histogram

histogram(q) calculates VictoriaMetrics histogram per each group of points with the same timestamp. Useful for visualizing big number of time series via a heatmap. See this article for more details.

limitk

limitk(k, q) by (group_labels) returns up to k time series per each group_labels out of time series returned by q. The returned set of time series remain the same across calls. See also limit_offset.

mad

mad(q) by (group_labels) returns the Median absolute deviation per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. See also outliers_mad and stddev.

max

max(q) by (group_labels) returns the maximum value per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

median

median(q) by (group_labels) returns the median value per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp.

min

min(q) by (group_labels) returns the minimum value per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

mode

mode(q) by (group_labels) returns mode per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp.

outliers_mad

outliers_mad(tolerance, q) returns time series from q with at least a single point outside Median absolute deviation (aka MAD) multiplied by tolerance. E.g. it returns time series with at least a single point below median(q) - mad(q) or a single point above median(q) + mad(q). See also outliersk and mad.

outliersk

outliersk(k, q) returns up to k time series with the biggest standard deviation (aka outliers) out of time series returned by q. See also outliers_mad.

quantile

quantile(phi, q) by (group_labels) calculates phi-quantile per each group_labels for all the time series returned by q. phi must be in the range [0...1]. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL. See also quantiles.

quantiles

quantiles("phiLabel", phi1, ..., phiN, q) calculates phi*-quantiles for all the time series returned by q and return them in time series with {phiLabel="phi*"} label. phi* must be in the range [0...1]. The aggregate is calculated individually per each group of points with the same timestamp. See also quantile.

stddev

stddev(q) by (group_labels) calculates standard deviation per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

stdvar

stdvar(q) by (group_labels) calculates standard variance per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

sum

sum(q) by (group_labels) returns the sum per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL.

sum2

sum2(q) by (group_labels) calculates the sum of squares per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp.

topk

topk(k, q) returns up to k points with the biggest values across all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. This function is supported by PromQL. See also bottomk.

topk_avg

topk_avg(k, q, "other_label=other_value") returns up to k time series from q with the biggest averages. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, topk_avg(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the biggest averages plus a time series with {job="other"} label with the sum of the remaining series if any. See also bottomk_avg.

topk_last

topk_last(k, q, "other_label=other_value") returns up to k time series from q with the biggest last values. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, topk_max(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the biggest amaximums plus a time series with {job="other"} label with the sum of the remaining series if any. See also bottomk_last.

topk_max

topk_max(k, q, "other_label=other_value") returns up to k time series from q with the biggest maximums. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, topk_max(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the biggest amaximums plus a time series with {job="other"} label with the sum of the remaining series if any. See also bottomk_max.

topk_median

topk_median(k, q, "other_label=other_value") returns up to k time series from q with the biggest medians. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, topk_median(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the biggest medians plus a time series with {job="other"} label with the sum of the remaining series if any. See also bottomk_median.

topk_min

topk_min(k, q, "other_label=other_value") returns up to k time series from q with the biggest minimums. If an optional other_label=other_value arg is set, then the sum of the remaining time series is returned with the given label. For example, topk_min(3, sum(process_resident_memory_bytes) by (job), "job=other") would return up to 3 time series with the biggest minimums plus a time series with {job="other"} label with the sum of the remaining series if any. See also bottomk_min.

zscore

zscore(q) by (group_labels) returns z-score values per each group_labels for all the time series returned by q. The aggregate is calculated individually per each group of points with the same timestamp. Useful for detecting anomalies in the group of related time series.

Subqueries

MetricsQL supports and extends PromQL subqueries. See this article for details. Any rollup function for something other than series selector form a subquery. Nested rollup functions can be implicit thanks to the implicit query conversions. For example, delta(sum(m)) is implicitly converted to delta(sum(default_rollup(m[1i]))[1i:1i]), so it becomes a subquery, since it contains default_rollup nested into delta.

VictoriaMetrics performs subqueries in the following way:

  • It calculates the inner rollup function using the step value from the outer rollup function. For example, for expression max_over_time(rate(http_requests_total[5m])[1h:30s]) the inner function rate(http_requests_total[5m]) is calculated with step=30s. The resulting data points are aligned by the step.
  • It calculates the outer rollup function over the results of the inner rollup function using the step value passed by Grafana to /api/v1/query_range.

Implicit query conversions

VictoriaMetrics performs the following implicit conversions for incoming queries before starting the calculations:

  • If lookbehind window in square brackets is missing inside rollup function, then [1i] is automatically added there. The [1i] means one step value, which is passed to /api/v1/query_range. It is also known as $__interval in Grafana. For example, rate(http_requests_count) is automatically transformed to rate(http_requests_count[1i]).
  • All the series selectors, which aren't wrapped into rollup functions, are automatically wrapped into default_rollup function. Examples:
    • foo is transformed to default_rollup(foo[1i])
    • foo + bar is transformed to default_rollup(foo[1i]) + default_rollup(bar[1i])
    • count(up) is transformed to count(default_rollup(up[1i])), because count isn't a rollup function - it is aggregate function
    • abs(temperature) is transformed to abs(default_rollup(temperature[1i])), because abs isn't a rollup function - it is transform function
  • If step in square brackets is missing inside subquery, then 1i step is automatically added there. For example, avg_over_time(rate(http_requests_total[5m])[1h]) is automatically converted to avg_over_time(rate(http_requests_total[5m])[1h:1i]).
  • If something other than series selector is passed to rollup function, then a subquery with 1i lookbehind window and 1i step is automatically formed. For example, rate(sum(up)) is automatically converted to rate((sum(default_rollup(up[1i])))[1i:1i]).