vendor | ||
.dockerignore | ||
.gitignore | ||
.promu.yml | ||
.travis.yml | ||
bridge_test.go | ||
CHANGELOG.md | ||
circle.yml | ||
CONTRIBUTING.md | ||
Dockerfile | ||
exporter.go | ||
exporter_benchmark_test.go | ||
exporter_test.go | ||
LICENSE | ||
main.go | ||
MAINTAINERS.md | ||
Makefile | ||
mapper.go | ||
mapper_test.go | ||
NOTICE | ||
README.md | ||
telemetry.go | ||
timer.go | ||
VERSION |
statsd exporter
statsd_exporter
receives StatsD-style metrics and exports them as Prometheus metrics.
Overview
With StatsD
To pipe metrics from an existing StatsD environment into Prometheus, configure
StatsD's repeater backend to repeat all received metrics to a statsd_exporter
process. This exporter translates StatsD metrics to Prometheus metrics via
configured mapping rules.
+----------+ +-------------------+ +--------------+
| StatsD |---(UDP/TCP repeater)--->| statsd_exporter |<---(scrape /metrics)---| Prometheus |
+----------+ +-------------------+ +--------------+
Without StatsD
Since the StatsD exporter uses the same line protocol as StatsD itself, you can also configure your applications to send StatsD metrics directly to the exporter. In that case, you don't need to run a StatsD server anymore.
We recommend this only as an intermediate solution and recommend switching to native Prometheus instrumentation in the long term.
DogStatsD extensions
The exporter will convert DogStatsD-style tags to prometheus labels. See
Tags in the DogStatsD
documentation for the concept description and
Datagram Format
for specifics. It boils down to appending
|#tag:value,another_tag:another_value
to the normal StatsD format. Tags
without values (#some_tag
) are not supported.
Building and Running
$ go build
$ ./statsd_exporter --help
Usage of ./statsd_exporter:
-log.format value
If set use a syslog logger or JSON logging. Example: logger:syslog?appname=bob&local=7 or logger:stdout?json=true. Defaults to stderr.
-log.level value
Only log messages with the given severity or above. Valid levels: [debug, info, warn, error, fatal].
-statsd.add-suffix
Add the metric type (counter/gauge/timer) as suffix to the generated Prometheus metric (NOT recommended, but set by default for backward compatibility). (default true)
-statsd.listen-address string
The UDP address on which to receive statsd metric lines. DEPRECATED, use statsd.listen-udp instead.
-statsd.listen-tcp string
The TCP address on which to receive statsd metric lines. "" disables it. (default ":9125")
-statsd.listen-udp string
The UDP address on which to receive statsd metric lines. "" disables it. (default ":9125")
-statsd.mapping-config string
Metric mapping configuration file name.
-statsd.read-buffer int
Size (in bytes) of the operating system's transmit read buffer associated with the UDP connection. Please make sure the kernel parameters net.core.rmem_max is set to a value greater than the value specified.
-version
Print version information.
-web.listen-address string
The address on which to expose the web interface and generated Prometheus metrics. (default ":9102")
-web.telemetry-path string
Path under which to expose metrics. (default "/metrics")
Tests
$ go test
Metric Mapping and Configuration
The statsd_exporter
can be configured to translate specific dot-separated StatsD
metrics into labeled Prometheus metrics via a simple mapping language. A
mapping definition starts with a line matching the StatsD metric in question,
with *
s acting as wildcards for each dot-separated metric component. The
lines following the matching expression must contain one label="value"
pair
each, and at least define the metric name (label name name
). The Prometheus
metric is then constructed from these labels. $n
-style references in the
label value are replaced by the n-th wildcard match in the matching line,
starting at 1. Multiple matching definitions are separated by one or more empty
lines. The first mapping rule that matches a StatsD metric wins.
Metrics that don't match any mapping in the configuration file are translated
into Prometheus metrics without any labels and with certain characters escaped
(_
-> __
; -
-> __
; .
-> _
).
In general, the different metric types are translated as follows:
StatsD gauge -> Prometheus gauge
StatsD counter -> Prometheus counter
StatsD timer -> Prometheus summary <-- indicates timer quantiles
-> Prometheus counter (suffix `_total`) <-- indicates total time spent
-> Prometheus counter (suffix `_count`) <-- indicates total number of timer events
If -statsd.add-suffix
is set, the exporter appends the metric type (_gauge
,
_counter
, _timer
) to the resulting metrics. This is enabled by default for
backward compatibility but discouraged to use. Instead, it is better to
explicitly define the full metric name in your mapping and run the exporter
with -statsd.add-suffix=false
.
An example mapping configuration with -statsd.add-suffix=false
:
# comments are allowed
test.dispatcher.*.*.*
name="dispatcher_events_total"
processor="$1"
action="$2"
outcome="$3"
job="test_dispatcher"
*.signup.*.*
name="signup_events_total"
provider="$2"
outcome="$3"
job="${1}_server"
This would transform these example StatsD metrics into Prometheus metrics as follows:
test.dispatcher.FooProcessor.send.success
=> dispatcher_events_total{processor="FooProcessor", action="send", outcome="success", job="test_dispatcher"}
foo_product.signup.facebook.failure
=> signup_events_total{provider="facebook", outcome="failure", job="foo_product_server"}
test.web-server.foo.bar
=> test_web__server_foo_bar{}
YAML may also be used for the configuration if the passed filename ends in .yml
or
.yaml
. The above example mapping, in YAML, would be:
mappings:
- match: test.dispatcher.*.*.*
labels:
name: "dispatcher_events_total"
processor: "$1"
action: "$2"
outcome: "$3"
job: "test_dispatcher"
- match: *.signup.*.*
labels:
name: "signup_events_total"
provider: "$2"
outcome: "$3"
job: "${1}_server"
Using the YAML configuration, one may also set the timer type to "histogram". The default is "summary" as in the plain text configuration format. For example, to set the timer type for a single metric:
mappings:
- match: test.timing.*.*.*
timer_type: histogram
buckets: [ 0.01, 0.025, 0.05, 0.1 ]
labels:
name: "my_timer"
provider: "$2"
outcome: "$3"
job: "${1}_server"
Note, that one may also set the histogram buckets. If not set, then the default
Prometheus client values are used: [.005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10]
. +Inf
is added
automatically.
timer_type
is only used when the statsd metric type is a timer. buckets
is
only used when the statsd metric type is a timerand the timer_type
is set to
"histogram."
One may also set defaults for the timer type and the buckets. These will be used by all mappings that do not define these.
defaults:
timer_type: histogram
buckets: [.005, .01, .025, .05, .1, .25, .5, 1, 2.5 ]
mappings:
# This will be a histogram using the buckets set in `defaults`.
- match: test.timing.*.*.*
labels:
name: "my_timer"
provider: "$2"
outcome: "$3"
job: "${1}_server"
# This will be a summary timer.
- match: other.timing.*.*.*
timer_type: summary
labels:
name: "other_timer"
provider: "$2"
outcome: "$3"
job: "${1}_server_other"
Using Docker
You can deploy this exporter using the prom/statsd-exporter Docker image.
For example:
docker pull prom/statsd-exporter
docker run -d -p 9102:9102 -p 9125:9125 -p 9125:9125/udp \
-v $PWD/statsd_mapping.conf:/tmp/statsd_mapping.conf \
prom/statsd-exporter -statsd.mapping-config=/tmp/statsd_mapping.conf