CAUTION: This page documents an old version of Prometheus. Check out the latest stable version.

Query examples

Simple time series selection

Return all time series with the metric http_requests_total:

http_requests_total

Return all time series with the metric http_requests_total and the given job and handler labels:

http_requests_total{job="apiserver", handler="/api/comments"}

Return a whole range of time (in this case 5 minutes up to the query time) for the same vector, making it a range vector:

http_requests_total{job="apiserver", handler="/api/comments"}[5m]

Note that an expression resulting in a range vector cannot be graphed directly, but viewed in the tabular ("Console") view of the expression browser.

Using regular expressions, you could select time series only for jobs whose name match a certain pattern, in this case, all jobs that end with server:

http_requests_total{job=~".*server"}

All regular expressions in Prometheus use RE2 syntax.

To select all HTTP status codes except 4xx ones, you could run:

http_requests_total{status!~"4.."}

Subquery

Return the 5-minute rate of the http_requests_total metric for the past 30 minutes, with a resolution of 1 minute.

rate(http_requests_total[5m])[30m:1m]

This is an example of a nested subquery. The subquery for the deriv function uses the default resolution. Note that using subqueries unnecessarily is unwise.

max_over_time(deriv(rate(distance_covered_total[5s])[30s:5s])[10m:])

Using functions, operators, etc.

Return the per-second rate for all time series with the http_requests_total metric name, as measured over the last 5 minutes:

rate(http_requests_total[5m])

Assuming that the http_requests_total time series all have the labels job (fanout by job name) and instance (fanout by instance of the job), we might want to sum over the rate of all instances, so we get fewer output time series, but still preserve the job dimension:

sum by (job) (
  rate(http_requests_total[5m])
)

If we have two different metrics with the same dimensional labels, we can apply binary operators to them and elements on both sides with the same label set will get matched and propagated to the output. For example, this expression returns the unused memory in MiB for every instance (on a fictional cluster scheduler exposing these metrics about the instances it runs):

(instance_memory_limit_bytes - instance_memory_usage_bytes) / 1024 / 1024

The same expression, but summed by application, could be written like this:

sum by (app, proc) (
  instance_memory_limit_bytes - instance_memory_usage_bytes
) / 1024 / 1024

If the same fictional cluster scheduler exposed CPU usage metrics like the following for every instance:

instance_cpu_time_ns{app="lion", proc="web", rev="34d0f99", env="prod", job="cluster-manager"}
instance_cpu_time_ns{app="elephant", proc="worker", rev="34d0f99", env="prod", job="cluster-manager"}
instance_cpu_time_ns{app="turtle", proc="api", rev="4d3a513", env="prod", job="cluster-manager"}
instance_cpu_time_ns{app="fox", proc="widget", rev="4d3a513", env="prod", job="cluster-manager"}
...

...we could get the top 3 CPU users grouped by application (app) and process type (proc) like this:

topk(3, sum by (app, proc) (rate(instance_cpu_time_ns[5m])))

Assuming this metric contains one time series per running instance, you could count the number of running instances per application like this:

count by (app) (instance_cpu_time_ns)

This documentation is open-source. Please help improve it by filing issues or pull requests.