UQL reference

This document describes UQL itself. Check out our cheat sheet, as well as a topic for building UQL queries specifically for span queries.

The Cloud Observability Unified Query Language (UQL) allows you to:

• Retrieve and manipulate time series, which may be derived from metrics or spans (see time series queries). Metrics telemetry must be analyzed using time series queries, since metrics telemetry is time series data. Spans can be interpreted as time series, for example by counting their start time (e.g. spans count).
• Fetch whole spans (see Trace queries) and their metadata (for example, attributes and logs). Use span syntax when creating a dependency map.

UQL has the same structure whether it’s used to query for spans or time series. UQL queries are written as a pipeline made up of stages: stage1 | stage2 | stage3. Stages operate sequentially and independently; stage1 produces some data, then stage2 modifies that, and finally stage3 operates on the output of stage 2.

Time series: Query overview

Time series queries can be created from metrics and spans.

Syntax

The first stage of the pipeline produces a table, and subsequent stages operate on that table using table operations. Table operations may produce rows, combine rows, change the number of columns in a table, or even combine tables.

Tables

Each pipeline stage takes some number of tables as input and produces tables as output. A table has a column for each label on the data, a pair of timestamp columns, and a value column. For example, the following query produces a table with a column for the service-name label:

1
2
3
metric cpu.usage
| delta 1s
| group_by [service-name], mean

Table groups

Table Groups are sets of UQL tables that proceed through a pipeline together. Table Groups are what allow you to join tables and produce outputs like error ratios.

Here’s an example of a query that produces request error ratios per-service:

Query with a Table Group

1
2
3
4
5
6
7
(
  metric requests;
  metric errors
)
| delta 1m
| group_by [service], sum
| join right / (left + right)

The Table Group is created by the parallel Fetch stages: (metric requests; metric errors;) and consists of two tables.

Distributing Stages

Stages that can work with table groups may be written after the closing parenthesis of the group. They will distribute into the group. The term distribute is the same as in arithmetic: (a + b) * c == a * c + b * c You can also think of this as operating on each stage concurrently. These two pipelines are equivalent because the delta and group_by stages are distributed into the table group.

Distributing stages

1
2
3
4
5
6
7
8
9
(
  metric requests | delta 1m | group_by [hostname], sum;
  metric errors | delta 1m | group_by [hostname], sum
) | join right / left

(
  metric requests;
  metric errors
) | delta 1m | group_by [hostname], sum | join right / left

Data Types

The values in a table’s columns are always one of these types:

• string: UTF-8 string.
• integer: 64-bit integer.
• float: 64-bit IEEE-754 float
• boolean: either true or false
• distribution: a statistical distribution. For example, the latency of requests over a given time period may be represented as a distribution.
• timestamp: a Unix Epoch timestamp, in microseconds

Arguments to table operations may be any of the table column types, as well as:

• duration: an integer followed by a duration specifier. For example: 10m. Valid duration specifiers are:
  • s for seconds
  • m for minutes
  • h for hours
  • d for days
  • w for weeks
• table identifiers: in join’s the tables are called left and right.
• column identifiers: the label of a table’s column, for example: region or hostname.
• value expressions: are combinations of identifiers, operators, and literals that are evaluated to produce new values in a table. For example, in a join operation: join left / (1 + left + right).
• filter expressions: are expressions that can operate on column names and values to produce a boolean result. For example: region == "us-east-1" || region == "us-west-1"
• reducers: are mathematical functions that are applied to time series values. Examples include mean, max, and count.
• fetch type: specifies what span-derived time series should be fetched. Valid fetch types are:
  • count: produces a delta float time series counting the number of spans
  • latency: produces a delta distribution of the latency of spans
  • custom-numeric-attribute: produces a delta distribution of the values of custom-numeric-attribute

Data Literals

Numeric literals are written like: 110 or 110.33

Boolean literals are true and false

String literals can be written in three ways to help with escaping symbols. The string literal test \ “‘123’“ can be written:

• With double quotes: "test \\ \"'123'\"". If you want to write a \ or " literal in the string, you’ll need to escape it with a preceding \.
• With single quotes: 'test \\ "\'123\'"'. If you want to write a \ or ' literal in the string, you’ll need to escape it with a preceding \.
• With backticks (`): `test \ "'123'"`. All characters that appear in a raw string are interpreted literally, so there’s no need to escape anything. The one caveat is that there’s no way to put a ` character in these strings.

Metric kinds

Metrics may be one of the following kinds. Kinds describe the relationship between successive points in the time series.

• Delta: A measurement of the change in the metric from the point to point. For example, the number of requests is usually a delta metric. Over a 5 second period, you might see a point every second like 1, 11, 2, 5, 6, which would indicate that 25 requests occurred in that timeframe.
• Gauge: An instantaneous measurement. There is no relationship between successive points. For example, the setting of a rate limit or other configuration value is often a gauge.

Certain reducers are only valid for certain kinds of metrics, as indicated below.

Time series: Fetch stages

Fetch stages take zero tables as input and produce a single table as output. They must come first in a pipeline.

metric

Syntax

1
metric <metric-name>

metric fetches the metric named metric-name.

Arguments

spans

Syntax

1
spans <fetch-type>

spans fetches a scalar time series of the count of spans or a distribution of either the spans latency or the values of a numeric attribute.

Arguments

constant

Syntax

1
constant <literal-value>

constant generates a gauge time series where all points have value literal-value. The generated points are aligned to the global output period, and the time series has no labels. This can be used to plot horizontal lines on charts.

Arguments

Time series: Aligners

Aligners take unaligned data and group it into a time series with points at regular intervals. They perform time aggregation, also called temporal aggregation. Aligners should generally follow immediately after a fetch operation. Aligners take two arguments: the input window and the output period. Checkout the alignment deep dive topic to learn more.

If there is more than one aligner stage in a query, all but the final aligner must provide both input window and output period arguments.

Cloud Observability supports these aligners:

• delta
• rate
• latest
• reduce

delta

Syntax

1
delta [input-window [, output-period]]

delta calculates the change in the metric across the input period, outputting a point every output period. The input-window functions as a smoothing argument. If the output-period is not provided, the Cloud Observability-calculated global output period is used.

Arguments

rate

Syntax

1
rate [input-window [, output-period]]

rate calculates the change per-second of the input tables, outputting a point every output period. The input-window functions as a smoothing argument: each output point is computed by taking the mean of the points within the input window. If the output-period is not provided, the Cloud Observability-calculated global output period is used.

Arguments

latest

Syntax

1
latest [input-window [, output-period]]

latest aligns gauge metrics to the output period by using the nearest point before an output point. The input-window tells latest how far to look back for input points. If the output-period is not provided, the Cloud Observability-calculated global output period is used.

Arguments

Example

Given this input data:

latest 1s, 1s will produce this output:

reduce

Syntax

1
reduce [input-window> [, output-period],] <reducer>

reduce performs temporal aggregation for a metric using the supplied reducer.

Arguments

Example

reduce 1h, max would produce the maximum value of the metric over the hour before the point.

Global output period

The global output period controls how many data points a UQL query produces per time series. The global output period isn’t something that’s set from within the language. It’s automatically set by the Cloud Observability UI or whatever interface is making the query. To show why the UI needs to have control over the time aggregation of a query, here’s an example.

In the Cloud Observability frontend, if you make a UQL query over the past 60 minutes the UI sets the global output period to 30 seconds. This means that there will be 120 data points per time series. Now let’s say you want to make that same query, but this time over the last day. When you change the query interval, the UI will adjust the global output period to 10 minutes. There will now be 144 data points in your chart per time series. This increase is necessary because if a query spanning the last day produced a data point every 30 seconds, madness would ensue! There would be a total of 7200 data points, which couldn’t be plotted easily.

The output of a UQL query must always be aligned to the global output period. In other words, neighboring points must always be global-output-period-apart. In order to meet this requirement, you must end a query with an aligner without an output period set, causing this aligner to default to the global output period.

For example, the query below isn’t allowed, because the trailing rate uses an output period of 1 minute.

Illegal Query

1
metric requests | rate 1m, 1m

It’s important to note that even if the global output period were set to 1 minute, the above query would still be considered illegal. Although its output period matches the global output period, if the global output period were to change to 2 minutes the query would produce incorrectly aligned data. Legal queries must produce data aligned to the global output period regardless of what it is set to.

The query below is legal, because the trailing reduce doesn’t specify an output period, which causes the global output period to be used instead:

Legal Query

1
metric requests | reduce 1m, sum

Time series: Modifiers

These stages change tables by dropping or combining rows, or by joining tables together. UQL has the following modifier stages:

• filter
• group_by
• join
• top
• bottom
• point
• point_filter
• time_shift

filter

Syntax

1
filter <predicate>`

filter drops rows from the table that do not match the predicate. The predicate is generally based on the label columns so that a complete series is either kept or filtered out.

Arguments

Examples

All requests time series where service is checkout and zone is us-east-1:

1
metric requests | delta | filter service == "checkout" && zone == "us-east-1"

All requests time series where the host name isn’t i-12345:

1
metric requests | delta | filter host != "i-12345"

All requests time series where the host name doesn’t start with i-123:

1
metric requests | delta | filter host =~ "i-123.*"

All responses time series where the integer-valued response_code label is in the 400s:

1
metric responses | delta | filter response_code >= 400 && response_code < 500

Count the number of spans where the host label is defined:

1
spans count | filter defined(host) | delta | group_by [], sum

group_by

Syntax

1
group_by [<label-list>], <reducer>

group_by combines rows with the same timestamps and the same values in the label columns within the [] using the provided reducer.

Arguments

Examples

Given this input data, produced by metric rate-limit | latest:

metric rate-limit | latest | group_by [region, zone], max produces:

An empty list of group_by fields means group all rows together, stripping off all attributes and producing a single time series. metric rate-limit | latest | group_by [], max produces:

join

Syntax

1
join <join-expression> [left-default [, right-default]]

join merges two tables into a single table using the join_expression. A join expression has two identifiers, left and right, that represent the values of rows in the input tables.

Join uses columns in the left and right tables with identical names for joining, and join matches all columns with identical names. There is a row for the cross-product of all the combinations of values in the join columns. This can result in a very large table. For this reason, it’s usually useful to reduce the number of columns in tables using group_by before a join. In general, using one or two columns for joining is a good rule of thumb.

Group-By before Join

1
(metric cpu.usage; metric cpu.limit) | delta | group_by [service], mean | join left / right * 100

The output tables of the group_by stage will have only one label column: service.

If there is a combination in one of the tables that is not present in the other, the default value argument is used.

Default values for Join

1
(metric cpu.usage; metric cpu.limit) | latest | group_by [service], mean | join left / right * 100, 0, 1

Arguments

Example

Given these two input tables:

join right / left produces:

Joining more than two tables

Using the with keyword and naming your join inputs allows joining 2 or more tables into a single table.

Percentage of cpu limits currently being used, broken down by container and pod

1
2
3
4
with
  usage = metric cpu.usage | latest | group_by [container, pod], mean;
  limit = metric cpu.limit | latest | group_by [container, pod], mean;
join usage / limit * 100, usage=0, limit=1

Mean error response percentage by service

1
2
3
4
5
with
  client_errors = metric responses | delta | filter response_class = 4xx | group_by [service], mean;
  server_errors = metric responses | delta | filter response_class = 5xx | group_by [service], mean;
  total = metric responses | delta | group_by [service], mean;
join ((client_errors + server_errors) / total) * 100

top

Syntax

1
top [[label-list],] <k>, <reducer>, [window]

top returns the k time series with the highest value as computed by the supplied reducer and time window. All other time series are filtered out. If labels are specified, top returns k time series for each combination of label values. As an example:

Grouped Top-K

1
metric cpu.usage | top [service, region], 10, mean, 10m

Will return ten time series for each combination of service and region in the table. For each of those combinations, up to ten time series will be returned, and the ranking will be based on the mean over the last ten minutes of the query.

The inverse of top is bottom.

Arguments

Examples

Top five time series where the CPU usage was the highest on average over the duration of the query:

1
metric kubernetes.cpu.usage.total | latest | top 5, mean

For each region and environment, the five time series where the CPU usage was the highest on average over the last hour of the query:

1
metric kubernetes.cpu.usage.total | latest | top ["region", "environment"], 5, mean, 1h

bottom

Syntax

1
bottom [[label-list],] <k>, <reducer>, [window]

Bottom is the exact inverse of top. It returns the k time series with the lowest values according to the provided reducer. For details about ranking by label combinations and time windows, see top

Arguments

Example

Bottom five time series where the CPU usage was lowest on average over the duration of the query.

1
metric kubernetes.cpu.usage.total | latest | bottom 5, mean

point

Syntax

1
point <value-expression>

Point evaluates the value-expression for each point in the input table and stores the result in a row in the output table. The value expression can make use of one identifier: value.

Arguments

Examples

Time series of the 95th percentile of spans latency for the warehouse service:

1
spans latency | delta | group_by [], sum | point percentile(value, 95)

Boolean-valued time series where a point is true if the value of my.metric is greater than 5.

1
metric my.metric | latest | point value > 5

Squared value of each point in my.metric:

1
metric my.metric | latest | point pow(value, 2)

Determine how long until SSL certificates expire:

1
metric ssl_expiration_unix_secs | point value - timestamp(value) | latest

point_filter

Syntax

1
point_filter <predicate>

point_filter evaluates the provided predicate against the value of each point. If the predicate returns false, the point is removed. Otherwise, the point remains unchanged.

Arguments

Example

Remove all negative gauge points:

1
metric cpu.utilization | point_filter value >= 0 | latest

Keep only distributions where the p99 latency is above 10 seconds:

1
2
3
4
spans latency
| delta
| point_filter percentile(value, 99) > 10000
| group_by [], sum

time_shift

Syntax

1
time_shift <shift-duration>

time_shift moves each point in the time series forwards by the amount of time specified in the duration.

This is equivalent to adding duration to the timestamp of each point.

time_shift does not have to immediately follow a fetch stage, but it must come before an aligner or group_by in the UQL query.

Arguments

Example

Rate of requests to the warehouse service, where each point in the time series is a point from the prior day, shifted forward.

spans count | time_shift 1d | rate | filter service = warehouse | group_by [], sum

Difference in request rate from one week prior:

1
2
3
4
5
(
  metric requests | rate | group_by [customer, method], sum;
  metric requests | time_shift 7d | rate | group_by [customer, method], sum
)
| join left - right

Trace queries

This section describes queries that fetch and manipulate whole traces. Trace queries fetch spans with all of their labels and logs. This is not to be confused with queries like spans count or spans latency, which build time series from spans. For help with those queries, see this topic.

spans_sample

Syntax

1
spans_sample [filter]

spans_sample must be the first stage in every trace query. The spans sampled at this stage will be the basis of assembling a trace, also known as seed spans in later stages. If a predicate is specified, spans_sample fetches only spans that match the predicate. If not, it may fetch any span. The stage will try to fetch a large sample of spans, but may not fetch that many if only a few spans match the predicate.

Arguments

assemble

Syntax

1
assemble

assemble takes a set of sampled spans as input, assembles a trace for each of those spans, and returns a set of traces. It takes no arguments.

trace_filter

Syntax

1
trace_filter <predicate>

trace_filter keeps or drops traces based on the labels on the spans that make up each trace. If predicate returns true for the set of span labels, the trace will be kept. Otherwise, it will be discarded.

For example, trace_filter operation == foo && service == bar will keep only traces which have operation == foo on some span and service == bar on some span. The labels may be on different spans or the same span.

Arguments

summarize_by

Syntax

1
summarize_by [<group-attribute-list>], <reducer>([<to-reduce-attribute>])...

summarize_by takes as input a set of traces and produces a table where each row summarizes a sub-group of those traces. summarize_by is almost identical to a SQL GROUP BY, with the notable difference that it also keeps the input traces that were used to build each summary row. [<group-attribute-list>] determines how traces will be grouped. The remaining arguments determine how each group will be translated into a row in the output table.

Arguments

[<group-attribute-list>] is a comma-separated list of whole-trace attributes enclosed by brackets. summarize_by will group traces based on the values of the specified attributes. Two types of attributes can be used:

Note that a “seed span” is the span that the trace was assembled from.

<reducer>([to-reduce-attribute])... is a comma-separated list of reducers operating on whole-trace attributes. For each group of traces, summarize_by will produce a row. The values of the columns in that row will be computed by the reducers in the summarize_by query. The allowed reducers are mean, count, and fraction_true. For the time being, only a few patterns are supported:

Applying reducers in other ways is not yet supported. If you try to write something else, such as mean(byte_size), you’ll see an error.

Example

Now, let’s walk through an example. Consider the query:

1
summarize_by [host, seed_span.response_code], mean(seed_span.latency), count(), fraction_true(seed_span.error)

First, the [host, seed_span.response_code] portion of the query indicates that input traces should be separated into groups based on:

• The set of values of the host attribute on all spans in each trace
• The value of the response_code attribute on the seed span of each trace

Next, a row will be produced for each group. The first cell of that row will contain an average of the latency of the seed spans in the group. The second cell will contain the number of traces in the group. And the third cell will contain the fraction of seed spans in the group that contain errors.

Pipeline arguments

This section explains all of the types that are references in the Argument Stages

Identifiers

Identifiers are unquoted strings that are used in several pipeline stages. In join the identifiers are left and right. In point the identifier is value. group_by and filter use identifiers as the first part of a label. For example, group_by[service] uses the identifiers service.

Value expressions

Value expressions are supported in the join and point table operations. They are made up of identifiers, literals, and arithmetic operators. Value expressions produce a single output value.

Examples

• left + right
• 2 * value - 3
• percentile(value, 99)
• dist_sum(value)

Filter expressions

Filter expressions are used in the filter table operation. Filter expressions take as input a set of labels and return true or false depending on whether those labels satisfy the predicate. A filter expression can be composed of many different operators:

• regular expressions (=~, !~)
• comparison operators (==, !=, <, <=, >, >=)
• presence operators:
  • defined (defined(some_key))
  • undefined (undefined(some_key))
• logical operators (&&, ||)

Examples

• region == "us-east-1"
• region =~ "us-.\*"
• response_code != 200
• response_code >= 400 && response_code < 500
• undefined(response_code) || response_code != 200

Arithmetic operators

Arithmetic expressions are listed in order from high to low precedence. They take and return numeric arguments on the value column of a table

Boolean operators

Boolean operators have the lowest precedence and are applied left-to-right.

Regular expressions

Regular expressions use the RE2 regular expression syntax.

By default, regular expressions:

• are case-sensitive
• are greedy
• allow . to match new lines
• are not multi-line aware (i.e. ^ will only match the start of text, not the start of a line)

You can modify this behavior with flags in the regular expression string. For example, if you want ^ / $ to match the start / end of lines, you can use (?m) to enable multi-line mode.

To match spans whose db.statement attribute has a WHERE clause at the start of a line, you could write:

Example multi-line db.statement attribute

1
2
3
SELECT *
FROM table
WHERE id = 2;

Query with flags

1
2
3
4
spans count
| delta
| filter db.statement =~ "(?mi)^where"
| group_by [service, operation], sum

The m flag lets ^ match the start of each line, not just the start of the string. The i flag allows a case-insensitive search, so the regex will match all variations of “where”.

Reducers

Reducers are required in the reduce, group_by, bottom, and top stages. Reducers perform calculations on groups of points and produce a single output.

Comments

All text on each line after the # character is treated as a comment. (Of course, the # inside a string won’t start a comment). Here is an example query with comments:

1
2
3
4
# This query plots the rate of requests to each service.
metric requests
| rate # this must be before the group_by
| group_by [service], sum

See also

UQL cheat sheet

Use UQL to create advanced queries

Getting started with spans queries in UQL

Updated Jun 16, 2023