pushgateway

Prometheus Pushgateway

The Prometheus Pushgateway exists to allow ephemeral and batch jobs to expose their metrics to Prometheus. Since these kinds of jobs may not exist long enough to be scraped, they can instead push their metrics to a Pushgateway. The Pushgateway then exposes these metrics to Prometheus.

The Pushgateway is explicitly not an aggregator, but rather a metrics cache. It does not have a statsd-like semantics. The metrics pushed are exactly the same as you would present for scraping in a permanently running program.

For machine-level metrics, the textfile collector of the Node exporter is usually more appropriate. The Pushgateway is best used for service-level metrics.

Run it

Compile the binary using the provided Makefile (type make). The binary will be put into the bin directory.

For the most basic setup, just start the binary. To change the address to listen on, use the -addr flag. The -persistence.file flag allows you to specify a file in which the pushed metrics will be persisted (so that they survive restarts of the Pushgateway).

Use it

Libraries

Prometheus client libraries should have a feature to push the registered metrics to a Pushgateway. Usually, a Prometheus client passively presents metric for scraping by a Prometheus server. A client library that supports pushing has a push function, which needs to be called by the client code. It will then actively push the metrics to a Pushgateway, using the API described below.

Command line

Using the Prometheus text protocol, pushing metrics is so easy that no separate CLI is provided. Simply use a command-line HTTP tool like curl. Your favorite scripting language has most likely some built-in HTTP capabilities you can leverage here as well.

Caveat: Note that in the text protocol, each line has to end with a line-feed character (aka 'LF' or '\n'). Ending a line in other ways, e.g. with 'CR' aka '\r', 'CRLF' aka '\r\n', or just the end of the packet, will result in a protocol error.

Examples:

• Push a single sample:

    echo "some_metric 3.14" | curl --data-binary @- http://pushgateway.example.org:8080/metrics/jobs/some_job
  

  Since no type information has been provided, some_metric will be of type untyped.

• Push something more complex:

    cat <<EOF | curl --data-binary @- http://pushgateway.example.org:8080/metrics/jobs/some_job/instances/some_instance
    # TYPE some_metric counter
    some_metric{label="val1"} 42
    # This one even has a timestamp (but beware, see below).
    some_metric{label="val2"} 34 1398355504000
    # TYPE another_metric gauge
    # HELP another_metric Just an example.
    another_metric 2398.283
    EOF
  

  Note how type information and help strings are provided. Those lines are optional, but strongly encouraged for anything more complex.

• Delete all metrics of an instance:

    curl -X DELETE http://pushgateway.example.org:8080/metrics/jobs/some_job/instances/some_instance
  

• Delete all metrics of a job:

    curl -X DELETE http://pushgateway.example.org:8080/metrics/jobs/some_job
  

About timestamps

If you push metrics at time t₁, you might be tempted to believe that Prometheus will scrape them with that same timestamp t₁. Instead, what Prometheus attaches as a timestamp is the time when it scrapes the push gateway. Why so?

In the world view of Prometheus, a metric can be scraped at any time. A metric that cannot be scraped has basically ceased to exist. Prometheus is somewhat tolerant, but if it cannot get any samples for a metric in 5min, it will behave as if that metric does not exist anymore. Preventing that is actually one of the reasons to use a push gateway. The push gateway will make the metrics of your ephemeral job scrapable at any time. Attaching the time of pushing as a timestamp would defeat that purpose because 5min after the last push, your metric will look as stale to Prometheus as if it could not be scraped at all anymore. (Prometheus knows only one timestamp per sample, there is no way to distinguish a 'time of pushing' and a 'time of scraping'.)

You can still force Prometheus to attach a different timestamp by using the optional timestamp field in the exchange format. However, there are very few use cases where that would make sense. (Essentially, if you push more often than every 5min, you could attach the time of pushing as a timestamp.)

API

All pushes are done via HTTP. The interface is REST-like.

URL

The default port the push gateway is listening to is 8080. The path looks like

/metrics/jobs/<JOBNAME>[/instances/<INSTANCENAME>]

<JOBNAME> is used as the value of the job label, and <INSTANCE> as the value of the instance label. The instance part of the URL is optional. If it is missing, the IP number of the pushing host is used as the value for the 'instance' label instead.

If those labels are already set in the body of the request (as regular labels, e.g. name{job="foo",instance="bar"} 42), the values of those labels will be overwritten with values determined as described above! (This behavior might be changed in the future if a valid use-case can be shown.)

POST method

POST is used to add metrics to previously pushed metrics. Note that only previously pushed metrics with a different metric name, job label or instance label are preserved. Each POST will completely replace existing metrics with the same metric name, job label, and instance label as the metrics pushed, even if the existing metrics had a different label set otherwise.

The response code upon success is always 202 (even if that same metric has never been pushed before, i.e. there is no feedback to the client if the push has replaced a metric or created a new one).

The body of the request contains the metrics to push either as delimited binary protocol buffers or in the simple flat text format (both in version 0.0.4, see the data exposition format specification. Discrimination between the two variants is done via content-type header. (In case of an unknown content-type, the text format is tried as a fall-back.)

If using the protobuf format, do not send duplicate MetricFamily proto messages (i.e. more than one with the same name) in one push, as they will overwrite each other.

A successfully finished request means that the pushed metrics are queued for an update of the storage. Scraping the push gateway may still yield the old sample value for that metric (or nothing at all if this is the first time that metric is pushed) until the queued update is processed. Neither is there a guarantee that the metric is persisted to disk. (A server crash may cause data loss. Or the push gateway is configured to not persist to disk at all.)

PUT method

PUT works exactly as POST with the important distinction that all metrics with the same job label and instance label are deleted before pushing any of the newly submitted metrics.

DELETE method

DELETE is used to delete metrics from the push gateway. The request must not contain any content. If both a job and an instance are specified in the URL, all metrics matching that job and instance are deleted. If only a job is specified, all metrics matching that job are deleted. The response code upon success is always 202. The delete request is merely queued at the moment. There is no guarantee that the request will actually be executed or that the result will make it to the persistence layer (e.g. in case of a server crash). However, the order of PUT/POST and DELETE request is guaranteed, i.e. if you have successfully sent a DELETE request and then send a PUT, it is guaranteed that the DELETE will be processed first (and vice versa).

Deleting non-existing metrics is a no-op and will not result in an error.

Development

The normal binary embeds the files in resources. For development purposes, it is handy to have a running binary use those files directly (so that you can see the effect of changes immediately). To switch to direct usage, type make bindata-debug just before compiling the binary. Switch back to "normal" mode by typing make bindata-embed. (Just make after a resource has changed will result in the same.)

Contributing

Relevant style guidelines are the Go Code Review Comments and the Formatting and style section of Peter Bourgon's Go: Best Practices for Production Environments.