# 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 differences 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 VictoriaMetrics playground or at your own VictoriaMetrics instance.

The list of MetricsQL features:

- 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 anywhere 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`

. - Numeric values can have
`K`

,`Ki`

,`M`

,`Mi`

,`G`

,`Gi`

,`T`

and`Ti`

suffixes. For example,`8K`

is equivalent to`8000`

, while`1.2Mi`

is equivalent to`1.2*1024*1024`

. - 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 label names may contain any unicode letter. For example
`температура{город="Киев"}`

is a value MetricsQL expression. - Metric names and labels names may contain escaped chars. For example,
`foo\-bar{baz\=aa="b"}`

is valid expression. It returns time series with name`foo-bar`

containing label`baz=aa`

with value`b`

. Additionally, the following escape sequences are supported:`\xXX`

, where`XX`

is hexadecimal representation of the escaped ascii char.`\uXXXX`

, where`XXXX`

is a hexadecimal representation of the escaped unicode 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`

.`WITH`

templates. This feature simplifies writing and managing complex queries. Go to WITH templates playground and try it.- String literals may be concatenated. This is useful with
`WITH`

templates:`WITH (commonPrefix="long_metric_prefix_") {__name__=commonPrefix+"suffix1"} / {__name__=commonPrefix+"suffix2"}`

. `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.

The list of supported rollup functions:

#### absent_over_time

`absent_over_time(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which calculates all the listed `rollup_func*`

for raw samples on the given lookbehind window `d`

. The calculations are performed 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])`

is a rollup function, which 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.

This function is useful for tracking height gains in GPS tracking. Metric names are stripped from the resulting rollups.

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])`

is a rollup function, which 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])`

is a rollup function, which 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])`

is a rollup function, which 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)`

is a rollup function, which 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.

See also count_over_time.

#### count_gt_over_time

`count_gt_over_time(series_selector[d], gt)`

is a rollup function, which 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.

See also count_over_time.

#### count_le_over_time

`count_le_over_time(series_selector[d], le)`

is a rollup function, which 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.

See also count_over_time.

#### count_ne_over_time

`count_ne_over_time(series_selector[d], ne)`

is a rollup function, which 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.

See also count_over_time.

#### count_over_time

`count_over_time(series_selector[d])`

is a rollup function, which calculates the number of 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 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])`

is a rollup function, which calculates the number of raw sample value decreases over the given lookbehind window `d`

per each time series returned from the given series_selector.

See also increases_over_time.

#### default_rollup

`default_rollup(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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 slightly different to the behaviour of `delta()`

function in Prometheus. See this article for details.

This function is supported by PromQL. See also increase and delta_prometheus.

#### delta_prometheus

`delta_prometheus(series_selector[d])`

is a rollup function, which 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.

See also delta.

#### deriv

`deriv(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also deriv_fast and ideriv.

#### deriv_fast

`deriv_fast(series_selector[d])`

is a rollup function, which 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.

#### descent_over_time

`descent_over_time(series_selector[d])`

is a rollup function, which 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.

This function is useful for tracking height loss in GPS tracking.

See also ascent_over_time.

#### distinct_over_time

`distinct_over_time(series_selector[d])`

is a rollup function, which returns the number of distinct raw sample values on the given lookbehind window `d`

per each time series returned from the given series_selector.

#### duration_over_time

`duration_over_time(series_selector[d], max_interval)`

is a rollup function, which 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.

#### first_over_time

`first_over_time(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which calculates geometric mean over raw samples on the given lookbehind window `d`

per each time series returned from the given series_selector.

#### histogram_over_time

`histogram_over_time(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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])`

is a rollup function, which 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)`

is a rollup function, which 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. See also range_linear_regression.

#### idelta

`idelta(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also delta.

#### ideriv

`ideriv(series_selector[d])`

is a rollup function, which 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.

See also deriv.

#### increase

`increase(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also increase_pure, increase_prometheus and delta.

#### increase_prometheus

`increase_prometheus(series_selector[d])`

is a rollup function, which 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.

See also increase_pure and increase.

#### increase_pure

`increase_pure(series_selector[d])`

iis a rollup function, which 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])`

is a rollup function, which calculates the number of raw sample value increases over the given lookbehind window `d`

per each time series returned from the given series_selector.

See also decreases_over_time.

#### integrate

`integrate(series_selector[d])`

is a rollup function, which calculates the integral over raw samples on the given lookbehind window `d`

per each time series returned from the given series_selector.

#### irate

`irate(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also rate and rollup_rate.

#### lag

`lag(series_selector[d])`

is a rollup function, which 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.

See also lifetime and duration_over_time.

#### last_over_time

`last_over_time(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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.

See also duration_over_time and lag.

#### mad_over_time

`mad_over_time(series_selector[d])`

is a rollup function, which calculates median absolute deviation over raw samples on the given lookbehind window `d`

per each time series returned from the given series_selector.

#### max_over_time

`max_over_time(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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])`

is a rollup function, which 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])`

is a rollup function, which 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)`

is a rollup function, which 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. See also range_linear_regression.

#### present_over_time

`present_over_time(series_selector[d])`

is a rollup function, which returns 1 if there is at least a single raw sample on the given lookbehind window `d`

. Otherwise, an empty result is returned.

This function is supported by PromQL.

#### quantile_over_time

`quantile_over_time(phi, series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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])`

is a rollup function, which 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])`

.

#### rate

`rate(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also irate and rollup_rate.

#### rate_over_sum

`rate_over_sum(series_selector[d])`

is a rollup function, which calculates per-second rate over the sum of raw samples on the given lookbehind window `d`

. The calculations are performed individually per each time series returned from the given series_selector.

#### resets

`resets(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL.

#### rollup

`rollup(series_selector[d])`

is a rollup function, which calculates `min`

, `max`

and `avg`

values for raw samples on the given lookbehind window `d`

and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels. These values are calculated individually per each time series returned from the given series_selector.

Optional 2nd argument `"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

#### rollup_candlestick

`rollup_candlestick(series_selector[d])`

is a rollup function, which calculates `open`

, `high`

, `low`

and `close`

values (aka OHLC) over raw samples on the given lookbehind window `d`

and returns them in time series with `rollup="open"`

, `rollup="high"`

, `rollup="low"`

and `rollup="close"`

additional labels. The calculations are performed individually per each time series returned from the given series_selector. This function is useful for financial applications.

Optional 2nd argument `"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

#### rollup_delta

`rollup_delta(series_selector[d])`

is a rollup function, which calculates differences between adjacent raw samples on the given lookbehind window `d`

and returns `min`

, `max`

and `avg`

values for the calculated differences and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels. The calculations are performed individually per each time series returned from the given series_selector.

Optional 2nd argument `"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

See also rollup_increase.

#### rollup_deriv

`rollup_deriv(series_selector[d])`

is a rollup function, which calculates per-second derivatives for adjacent raw samples on the given lookbehind window `d`

and returns `min`

, `max`

and `avg`

values for the calculated per-second derivatives and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels. The calculations are performed individually per each time series returned from the given series_selector.

`"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

#### rollup_increase

`rollup_increase(series_selector[d])`

is a rollup function, which calculates increases for adjacent raw samples on the given lookbehind window `d`

and returns `min`

, `max`

and `avg`

values for the calculated increases and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels. The calculations are performed individually per each time series returned from the given series_selector.

`"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

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])`

is a rollup function, which calculates per-second change rates for adjacent raw samples on the given lookbehind window `d`

and returns `min`

, `max`

and `avg`

values for the calculated per-second change rates and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels.

See this article in order to understand better when to use `rollup_rate()`

.

`"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

The calculations are performed individually per each time series returned from the given series_selector.

#### rollup_scrape_interval

`rollup_scrape_interval(series_selector[d])`

is a rollup function, which calculates the interval in seconds between adjacent raw samples on the given lookbehind window `d`

and returns `min`

, `max`

and `avg`

values for the calculated interval and returns them in time series with `rollup="min"`

, `rollup="max"`

and `rollup="avg"`

additional labels. The calculations are performed individually per each time series returned from the given series_selector.

`"min"`

, `"max"`

or `"avg"`

can be passed to keep only one calculation result and without adding a label.

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])`

is a rollup function, which 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.

See also rollup_scrape_interval.

#### share_gt_over_time

`share_gt_over_time(series_selector[d], gt)`

is a rollup function, which 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.

This function is 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)`

is a rollup function, which 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.

This function is 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])`

is a rollup function, which calculates the number of staleness markers on the given lookbehind window `d`

per each time series matching the given series_selector.

#### stddev_over_time

`stddev_over_time(series_selector[d])`

is a rollup function, which calculates standard deviation 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 stdvar_over_time.

#### stdvar_over_time

`stdvar_over_time(series_selector[d])`

is a rollup function, which calculates standard variance 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 stddev_over_time.

#### sum_over_time

`sum_over_time(series_selector[d])`

is a rollup function, which calculates the sum of raw sample values on the given lookbehind window `d`

per each time series returned from the given series_selector.

This function is supported by PromQL.

#### sum2_over_time

`sum2_over_time(series_selector[d])`

is a rollup function, which 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.

#### timestamp

`timestamp(series_selector[d])`

is a rollup function, which 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.

This function is supported by PromQL. See also timestamp_with_name.

#### timestamp_with_name

`timestamp_with_name(series_selector[d])`

is a rollup function, which 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])`

is a rollup function, which 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.

See also first_over_time.

#### tlast_change_over_time

`tlast_change_over_time(series_selector[d])`

is a rollup function, which 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`

.

See also last_over_time.

#### tlast_over_time

`tlast_over_time(series_selector[d])`

is a rollup function, which is an alias for timestamp.

See also tlast_change_over_time.

#### tmax_over_time

`tmax_over_time(series_selector[d])`

is a rollup function, which 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.

See also max_over_time.

#### tmin_over_time

`tmin_over_time(series_selector[d])`

is a rollup function, which 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.

See also min_over_time.

#### zscore_over_time

`zscore_over_time(series_selector[d])`

is a rollup function, which 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.

See also zscore and range_trim_zscore.

### 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.

The list of supported transform functions:

#### abs

`abs(q)`

is a transform function, which calculates the absolute value for every point of every time series returned by `q`

.

This function is supported by PromQL.

#### absent

`absent(q)`

is a transform function, which returns 1 if `q`

has no points. Otherwise, returns an empty result.

This function is supported by PromQL. See also absent_over_time.

#### acos

`acos(q)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which returns inverse hyperbolic sine for every point of every time series returned by `q`

.

This function is supported by PromQL. See also sinh.

#### atan

`atan(q)`

is a transform function, which returns inverse tangent for every point of every time series returned by `q`

.

This function is supported by PromQL. See also tan.

#### atanh

`atanh(q)`

is a transform function, which returns inverse hyperbolic tangent for every point of every time series returned by `q`

.

This function is supported by PromQL. See also tanh.

#### bitmap_and

`bitmap_and(q, mask)`

is a transform function, which calculates bitwise `v & mask`

for every `v`

point of every time series returned from `q`

.

#### bitmap_or

`bitmap_or(q, mask)`

is a transform function, which calculates bitwise `v | mask`

for every `v`

point of every time series returned from `q`

.

#### bitmap_xor

`bitmap_xor(q, mask)`

is a transform function, which calculates bitwise `v ^ mask`

for every `v`

point of every time series returned from `q`

.

#### buckets_limit

`buckets_limit(limit, buckets)`

is a transform function, which limits the number of histogram buckets to the given `limit`

.

See also prometheus_buckets and histogram_quantile.

#### ceil

`ceil(q)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which clamps every point 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)`

is a transform function, which returns `cos(v)`

for every `v`

point of every time series returned by `q`

.

This function is supported by PromQL. See also sin.

#### cosh

`cosh(q)`

is a transform function, which returns hyperbolic cosine for every point of every time series returned by `q`

.

This function is supported by PromQL. This function is supported by PromQL. See also acosh.

#### day_of_month

`day_of_month(q)`

is a transform function, which 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]`

.

This function is supported by PromQL.

#### day_of_week

`day_of_week(q)`

is a transform function, which 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.

This function is supported by PromQL.

#### days_in_month

`days_in_month(q)`

is a transform function, which 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]`

.

This function is supported by PromQL.

#### deg

`deg(q)`

is a transform function, which converts Radians to degrees for every point of every time series returned by `q`

.

This function is supported by PromQL. See also rad.

#### end

`end()`

is a transform function, which returns the unix timestamp in seconds for the last point. It is known as `end`

query arg passed to /api/v1/query_range.

#### exp

`exp(q)`

is a transform function, which calculates the `e^v`

for every point `v`

of every time series returned by `q`

.

This function is supported by PromQL. See also ln.

#### floor

`floor(q)`

is a transform function, which rounds every point for every time series returned by `q`

to the lower nearest integer.

This function is supported by PromQL. See also ceil and round.

#### histogram_avg

`histogram_avg(buckets)`

is a transform function, which 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 example, `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)`

is a transform function, which calculates `phi`

-percentile 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.

The function accepts optional third arg - `boundsLabel`

. In this case it returns `lower`

and `upper`

bounds for the estimated percentile with the given `boundsLabel`

label. See this issue for details.

When the percentile is calculated over multiple histograms, then all the input histograms **must** have buckets with identical boundaries, e.g. they must have the same set of `le`

or `vmrange`

labels. Otherwise, the returned result may be invalid. See this issue for details.

This function is supported by PromQL (except of the `boundLabel`

arg). See also histogram_quantiles, histogram_share and quantile.

#### histogram_quantiles

`histogram_quantiles("phiLabel", phi1, ..., phiN, buckets)`

is a transform function, which 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)`

is a transform function, which calculates the share (in the range `[0...1]`

) for `buckets`

that fall below `le`

. This function is useful for calculating SLI and SLO. This is inverse to histogram_quantile.

The function accepts optional third arg - `boundsLabel`

. In this case it returns `lower`

and `upper`

bounds for the estimated share with the given `boundsLabel`

label.

#### histogram_stddev

`histogram_stddev(buckets)`

is a transform function, which calculates standard deviation for the given `buckets`

.

#### histogram_stdvar

`histogram_stdvar(buckets)`

is a transform function, which 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)`

is a transform function, which 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]`

.

This function is supported by PromQL.

#### interpolate

`interpolate(q)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which calculates `ln(v)`

for every point `v`

of every time series returned by `q`

.

This function is supported by PromQL. See also exp and log2.

#### log2

`log2(q)`

is a transform function, which calculates `log2(v)`

for every point `v`

of every time series returned by `q`

.

This function is supported by PromQL. See also log10 and ln.

#### log10

`log10(q)`

is a transform function, which calculates `log10(v)`

for every point `v`

of every time series returned by `q`

.

This function is supported by PromQL. See also log2 and ln.

#### minute

`minute(q)`

is a transform function, which 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]`

.

This function is supported by PromQL.

#### month

`month(q)`

is a transform function, which 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.

This function is supported by PromQL.

#### now

`now()`

is a transform function, which returns the current timestamp as a floating-point value in seconds.

See also time.

#### pi

`pi()`

is a transform function, which returns Pi number.

This function is supported by PromQL.

#### rad

`rad(q)`

is a transform function, which converts degrees to Radians for every point of every time series returned by `q`

.

This function is supported by PromQL. See also deg.

#### prometheus_buckets

`prometheus_buckets(buckets)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which returns pseudo-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)`

is a transform function, which calculates the avg value across points per each time series returned by `q`

.

#### range_first

`range_first(q)`

is a transform function, which returns the value for the first point per each time series returned by `q`

.

#### range_last

`range_last(q)`

is a transform function, which returns the value for the last point per each time series returned by `q`

.

#### range_linear_regression

`range_linear_regression(q)`

is a transform function, which calculates simple linear regression over the selected time range per each time series returned by `q`

. This function is useful for capacity planning and predictions.

#### range_mad

`range_mad(q)`

is a transform function, which calculates the median absolute deviation across points per each time series returned by `q`

.

See also mad and mad_over_time.

#### range_max

`range_max(q)`

is a transform function, which calculates the max value across points per each time series returned by `q`

.

#### range_median

`range_median(q)`

is a transform function, which calculates the median value across points per each time series returned by `q`

.

#### range_min

`range_min(q)`

is a transform function, which calculates the min value across points per each time series returned by `q`

.

#### range_normalize

`range_normalize(q1, ...)`

is a transform function, which normalizes values for time series returned by `q1, ...`

into `[0 ... 1]`

range. This function is useful for correlating time series with distinct value ranges.

See also share.

#### range_quantile

`range_quantile(phi, q)`

is a transform function, which returns `phi`

-quantile across points per each time series returned by `q`

. `phi`

must be in the range `[0...1]`

.

#### range_stddev

`range_stddev(q)`

is a transform function, which calculates standard deviation per each time series returned by `q`

on the selected time range.

#### range_stdvar

`range_stdvar(q)`

is a transform function, which calculates standard variance per each time series returned by `q`

on the selected time range.

#### range_sum

`range_sum(q)`

is a transform function, which calculates the sum of points per each time series returned by `q`

.

#### range_trim_outliers

`range_trim_outliers(k, q)`

is a transform function, which drops points located farther than `k*range_mad(q)`

from the `range_median(q)`

. E.g. it is equivalent to the following query: `q ifnot (abs(q - range_median(q)) > k*range_mad(q))`

.

See also range_trim_spikes and range_trim_zscore.

#### range_trim_spikes

`range_trim_spikes(phi, q)`

is a transform function, which drops `phi`

percent of biggest spikes from time series returned by `q`

. The `phi`

must be in the range `[0..1]`

, where `0`

means `0%`

and `1`

means `100%`

.

See also range_trim_outliers and range_trim_zscore.

#### range_trim_zscore

`range_trim_zscore(z, q)`

is a transform function, which drops points located farther than `z*range_stddev(q)`

from the `range_avg(q)`

. E.g. it is equivalent to the following query: `q ifnot (abs(q - range_avg(q)) > z*range_avg(q))`

.

See also range_trim_outliers and range_trim_spikes.

#### range_zscore

`range_zscore(q)`

is a transform function, which calculates z-score for points returned by `q`

, e.g. it is equivalent to the following query: `(q - range_avg(q)) / range_stddev(q)`

.

#### remove_resets

`remove_resets(q)`

is a transform function, which removes counter resets from time series returned by `q`

.

#### round

`round(q, nearest)`

is a transform function, which rounds 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)`

is a transform function, which 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)`

is a transform function, which calculates the running avg per each time series returned by `q`

.

#### running_max

`running_max(q)`

is a transform function, which calculates the running max per each time series returned by `q`

.

#### running_min

`running_min(q)`

is a transform function, which calculates the running min per each time series returned by `q`

.

#### running_sum

`running_sum(q)`

is a transform function, which calculates the running sum per each time series returned by `q`

.

#### scalar

`scalar(q)`

is a transform function, which returns `q`

if `q`

contains only a single time series. Otherwise, it returns nothing.

This function is supported by PromQL.

#### sgn

`sgn(q)`

is a transform function, which 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`

.

This function is supported by PromQL.

#### sin

`sin(q)`

is a transform function, which returns `sin(v)`

for every `v`

point of every time series returned by `q`

.

This function is supported by MetricsQL. See also cos.

#### sinh

`sinh(q)`

is a transform function, which returns hyperbolic sine for every point of every time series returned by `q`

.

This function is supported by MetricsQL. See also cosh.

#### tan

`tan(q)`

is a transform function, which returns `tan(v)`

for every `v`

point of every time series returned by `q`

.

This function is supported by MetricsQL. See also atan.

#### tanh

`tanh(q)`

is a transform function, which returns hyperbolic tangent for every point of every time series returned by `q`

.

This function is supported by MetricsQL. See also atanh.

#### smooth_exponential

`smooth_exponential(q, sf)`

is a transform function, which smooths points per each time series returned by `q`

using exponential moving average with the given smooth factor `sf`

.

#### sort

`sort(q)`

is a transform function, which 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 and sort_by_label.

#### sort_desc

`sort_desc(q)`

is a transform function, which 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 and sort_by_label.

#### sqrt

`sqrt(q)`

is a transform function, which calculates square root for every point of every time series returned by `q`

.

This function is supported by PromQL.

#### start

`start()`

is a transform function, which returns unix timestamp in seconds for the first point.

It is known as `start`

query arg passed to /api/v1/query_range.

#### step

`step()`

is a transform function, which 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()`

is a transform function, which returns unix timestamp for every returned point.

This function is supported by PromQL. See also now, start and end.

#### timezone_offset

`timezone_offset(tz)`

is a transform function, which 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)`

is a transform function, which 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)`

is a transform function, which returns a union of time series returned from `q1`

, …, `qN`

. The `union`

function name can be skipped - the following queries are equivalent: `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)`

is a transform function, which returns `q`

, e.g. it does nothing in MetricsQL.

This function is supported by PromQL.

#### year

`year(q)`

is a transform function, which returns the year for every point of every time series returned by `q`

. It is expected that `q`

returns unix timestamps.

This function is supported by PromQL.

### Label manipulation functions

**Label manipulation functions** perform manipulations with labels 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.

The list of supported label manipulation functions:

#### alias

`alias(q, "name")`

is label manipulation function, which 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)`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which deletes the given `label*`

labels from all the time series returned by `q`

.

#### label_graphite_group

`label_graphite_group(q, groupNum1, ... groupNumN)`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which maps `label`

values from `src_*`

to `dst*`

for all the time series returned by `q`

.

#### label_match

`label_match(q, "label", "regexp")`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which 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")`

is label manipulation function, which sets `{label1="value1", ..., labelN="valueN"}`

labels to all the time series returned by `q`

.

#### label_transform

`label_transform(q, "label", "regexp", "replacement")`

is label manipulation function, which substitutes all the `regexp`

occurrences by the given `replacement`

in the given `label`

.

#### label_uppercase

`label_uppercase(q, "label1", ..., "labelN")`

is label manipulation function, which uppercases values for the given `label*`

labels in all the time series returned by `q`

.

See also label_lowercase.

#### label_value

`label_value(q, "label")`

is label manipulation function, which returns numeric 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. Function will return no data for non-numeric label values.

#### sort_by_label

`sort_by_label(q, label1, ... labelN)`

is label manipulation function, which 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 and sort_by_label_numeric.

#### sort_by_label_desc

`sort_by_label_desc(q, label1, ... labelN)`

is label manipulation function, which 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 and sort_by_label_numeric_desc.

#### sort_by_label_numeric

`sort_by_label_numeric(q, label1, ... labelN)`

is label manipulation function, which sorts series in ascending order by the given set of labels using numeric sort. For example, if `foo`

series have `bar`

label with values `1`

, `101`

, `15`

and `2`

, then `sort_by_label_numeric(foo, "bar")`

would return series in the following order of `bar`

label values: `1`

, `2`

, `15`

and `101`

.

See also sort_by_label_numeric_desc and sort_by_label.

#### sort_by_label_numeric_desc

`sort_by_label_numeric_desc(q, label1, ... labelN)`

is label manipulation function, which sorts series in descending order by the given set of labels using numeric sort. For example, if `foo`

series have `bar`

label with values `1`

, `101`

, `15`

and `2`

, then `sort_by_label_numeric(foo, "bar")`

would return series in the following order of `bar`

label values: `101`

, `15`

, `2`

and `1`

.

See also sort_by_label_numeric and sort_by_label_desc.

### 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 calculating 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.

The list of supported aggregate functions:

#### any

`any(q) by (group_labels)`

is aggregate function, which returns a single series per `group_labels`

out of time series returned by `q`

.

See also group.

#### avg

`avg(q) by (group_labels)`

is aggregate function, which 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)`

is aggregate function, which 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")`

is aggregate function, which 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")`

is aggregate function, which 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")`

is aggregate function, which 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")`

is aggregate function, which returns up to `k`

time series from `q`

with the smallest 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, `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")`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which counts the number of points with the same value and stores the counts in a time series with an additional `label`

, which 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)`

is aggregate function, which calculates the number of unique values per each group of points with the same timestamp.

#### geomean

`geomean(q)`

is aggregate function, which calculates geometric mean per each group of points with the same timestamp.

#### group

`group(q) by (group_labels)`

is aggregate function, which returns `1`

per each `group_labels`

for time series returned by `q`

.

This function is supported by PromQL. See also any.

#### histogram

`histogram(q)`

is aggregate function, which 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.

See also histogram_over_time and histogram_quantile.

#### limitk

`limitk(k, q) by (group_labels)`

is aggregate function, which 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)`

is aggregate function, which 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 range_mad, mad_over_time, outliers_mad and stddev.

#### max

`max(q) by (group_labels)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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)`

.

#### outliersk

`outliersk(k, q)`

is aggregate function, which 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)`

is aggregate function, which 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 and histogram_quantile.

#### quantiles

`quantiles("phiLabel", phi1, ..., phiN, q)`

is aggregate function, which 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.

#### share

`share(q) by (group_labels)`

is aggregate function, which returns shares in the range `[0..1]`

for every non-negative points returned by `q`

per each timestamp, so the sum of shares per each `group_labels`

equals 1.

This function is useful for normalizing histogram bucket shares into `[0..1]`

range:

```
share(
sum(
rate(http_request_duration_seconds_bucket[5m])
) by (le, vmrange)
)
```

See also range_normalize.

#### stddev

`stddev(q) by (group_labels)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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)`

is aggregate function, which 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")`

is aggregate function, which 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")`

is aggregate function, which 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 maximums 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")`

is aggregate function, which 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 maximums 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")`

is aggregate function, which 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")`

is aggregate function, which 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)`

is aggregate function, which 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. This function is useful for detecting anomalies in the group of related time series.

See also zscore_over_time and range_trim_zscore.

## 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])`

.