Mirror of StatsD to Prometheus metrics exporter
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
Matthias Rampke 7c792fba66
Merge pull request #344 from prometheus/repo_sync
2 days ago
.circleci Bump go to 1.15 1 month ago
pkg Use an interface for Registry so we may have multiple implementations 2 days ago
vendor go mod vendor 4 months ago
.dockerignore Build ARM container images (#242) 1 year ago
.gitattributes eol update 6 months ago
.gitignore Removed .exe 6 months ago
.golangci.yml Add .golangci.yml 1 year ago
.promu.yml Bump go to 1.15 1 month ago
CHANGELOG.md Version bump + changelog for #339 1 week ago
CODE_OF_CONDUCT.md Update common Prometheus files 4 months ago
CONTRIBUTING.md Replace AUTHORS.md by an updated MAINTAINERS.md 3 years ago
Dockerfile Build ARM container images (#242) 1 year ago
LICENSE License cleanup 5 years ago
MAINTAINERS.md Change maintainership to Matthias 3 years ago
Makefile Build ARM container images (#242) 1 year ago
Makefile.common Update common Prometheus files 2 days ago
NOTICE rename bridge -> exporter 5 years ago
README.md Minor documentation fixes 3 weeks ago
VERSION Version bump + changelog for #339 1 week ago
bridge_test.go move to line parser struct and option functions 2 months ago
exporter_benchmark_test.go move to line parser struct and option functions 2 months ago
go.mod go mod tidy 4 months ago
go.sum go mod tidy 4 months ago
line_benchmark_test.go move to line parser struct and option functions 2 months ago
main.go Merge pull request #339 from yaron-idan/add-health-endpoint 1 week ago

README.md

statsd exporter Build Status

CircleCI Docker Repository on Quay Docker Pulls

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.

Tagging Extensions

The exporter supports Librato, InfluxDB, DogStatsD, and SignalFX-style tags, which will be converted into Prometheus labels.

For Librato-style tags, they must be appended to the metric name with a delimiting #, as so:

metric.name#tagName=val,tag2Name=val2:0|c

See the statsd-librato-backend README for a more complete description.

For InfluxDB-style tags, they must be appended to the metric name with a delimiting comma, as so:

metric.name,tagName=val,tag2Name=val2:0|c

See this InfluxDB blog post for a larger overview.

For DogStatsD-style tags, they’re appended as a |# delimited section at the end of the metric, as so:

metric.name:0|c|#tagName:val,tag2Name:val2

See Tags in the DogStatsD documentation for the concept description and Datagram Format. If you encounter problems, note that this tagging style is incompatible with the original statsd implementation.

For SignalFX dimension, add the tags to the metric name in square brackets, as so:

metric.name[tagName=val,tag2Name=val2]:0|c

Be aware: If you mix tag styles (e.g., Librato/InfluxDB with DogStatsD), the exporter will consider this an error and the behavior is undefined. Also, tags without values (#some_tag) are not supported and will be ignored.

The exporter parses all tagging formats by default, but individual tagging formats can be disabled with command line flags:

--no-statsd.parse-dogstatsd-tags
--no-statsd.parse-influxdb-tags
--no-statsd.parse-librato-tags
--no-statsd.parse-signalfx-tags

Building and Running

NOTE: Version 0.7.0 switched to the kingpin flags library. With this change, flag behaviour is POSIX-ish:

  • long flags start with two dashes (--version)
  • boolean long flags are disabled by prefixing with no (--flag-name is true, --no-flag-name is false)
  • multiple short flags can be combined (but there currently is only one)
  • flag processing stops at the first --

    usage: statsd_exporter [<flags>]
    
    Flags:
      -h, --help                    Show context-sensitive help (also try
                                    --help-long and --help-man).
          --web.listen-address=":9102"
                                    The address on which to expose the web interface
                                    and generated Prometheus metrics.
          --web.enable-lifecycle    Enable shutdown and reload via HTTP request.
          --web.telemetry-path="/metrics"
                                    Path under which to expose metrics.
          --statsd.listen-udp=":9125"
                                    The UDP address on which to receive statsd
                                    metric lines. "" disables it.
          --statsd.listen-tcp=":9125"
                                    The TCP address on which to receive statsd
                                    metric lines. "" disables it.
          --statsd.listen-unixgram=""
                                    The Unixgram socket path to receive statsd
                                    metric lines in datagram. "" disables it.
          --statsd.unixsocket-mode="755"
                                    The permission mode of the unix socket.
          --statsd.mapping-config=STATSD.MAPPING-CONFIG
                                    Metric mapping configuration file name.
          --statsd.read-buffer=STATSD.READ-BUFFER
                                    Size (in bytes) of the operating system's
                                    transmit read buffer associated with the UDP or
                                    Unixgram connection. Please make sure the kernel
                                    parameters net.core.rmem_max is set to a value
                                    greater than the value specified.
          --statsd.cache-size=1000  Maximum size of your metric mapping cache.
                                    Relies on least recently used replacement policy
                                    if max size is reached.
          --statsd.cache-type=lru   Metric mapping cache type. Valid options are
                                    "lru" and "random"
          --statsd.event-queue-size=10000
                                    Size of internal queue for processing events
          --statsd.event-flush-threshold=1000
                                    Number of events to hold in queue before
                                    flushing
          --statsd.event-flush-interval=200ms
                                    Maximum time between event queue flushes.
          --debug.dump-fsm=""       The path to dump internal FSM generated for
                                    glob matching as Dot file.
          --check-config            Check configuration and exit.
          --statsd.parse-dogstatsd-tags  
                                    Parse DogStatsd style tags. Enabled by default.
          --statsd.parse-influxdb-tags  
                                    Parse InfluxDB style tags. Enabled by default.
          --statsd.parse-librato-tags  
                                    Parse Librato style tags. Enabled by default.
          --statsd.parse-signalfx-tags  
                                    Parse SignalFX style tags. Enabled by default.
          --log.level=info          Only log messages with the given severity or
                                    above. One of: [debug, info, warn, error]
          --log.format=logfmt       Output format of log messages. One of: [logfmt,
                                    json]
          --version                 Show application version.
    

Lifecycle API

The statsd_exporter has an optional lifecycle API (disabled by default) that can be used to reload or quit the exporter by sending a PUT or POST request to the /-/reload or /-/quit endpoints.

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. The config file is reloaded on SIGHUP.

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 any non-alphanumeric characters, including periods, translated into underscores.

In general, the different metric types are translated as follows:

StatsD gauge   -> Prometheus gauge

StatsD counter -> Prometheus counter

StatsD timer, histogram, distribution   -> Prometheus summary or histogram

An example mapping configuration:

mappings:
- match: "test.dispatcher.*.*.*"
  name: "dispatcher_events_total"
  labels:
    processor: "$1"
    action: "$2"
    outcome: "$3"
    job: "test_dispatcher"
- match: "*.signup.*.*"
  name: "signup_events_total"
  labels:
    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{}

Each mapping in the configuration file must define a name for the metric. The metric’s name can contain $n-style references to be replaced by the n-th wildcard match in the matching line. That allows for dynamic rewrites, such as:

mappings:
- match: "test.*.*.counter"
  name: "${2}_total"
  labels:
    provider: "$1"

The metric name can also contain references to regex matches. The mapping above could be written as:

mappings:
- match: "test\\.(\\w+)\\.(\\w+)\\.counter"
  match_type: regex
  name: "${2}_total"
  labels:
    provider: "$1"

Be aware about yaml escape rules as a mapping like the following one will not work.

mappings:
- match: "test\.(\w+)\.(\w+)\.counter"
  match_type: regex
  name: "${2}_total"
  labels:
    provider: "$1"

Please note that metrics with the same name must also have the same set of label names.

If the default metric help text is insufficient for your needs you may use the YAML configuration to specify a custom help text for each mapping:

mappings:
- match: "http.request.*"
  help: "Total number of http requests"
  name: "http_requests_total"
  labels:
    code: "$1"

StatsD timers and distributions

By default, statsd timers and distributions (collectively “observers”) are represented as a Prometheus summary with quantiles. You may optionally configure the quantiles and acceptable error, as well as adjusting how the summary metric is aggregated:

mappings:
- match: "test.timing.*.*.*"
  observer_type: summary
  name: "my_timer"
  labels:
    provider: "$2"
    outcome: "$3"
    job: "${1}_server"
  summary_options:
    quantiles:
      - quantile: 0.99
        error: 0.001
      - quantile: 0.95
        error: 0.01
      - quantile: 0.9
        error: 0.05
      - quantile: 0.5
        error: 0.005
    max_summary_age: 30s
    summary_age_buckets: 3
    stream_buffer_size: 1000

The default quantiles are 0.99, 0.9, and 0.5.

The default summary age is 10 minutes, the default number of buckets is 5 and the default buffer size is 500. See also the golang_client docs. The max_summary_age corresponds to SummaryOptions.MaxAge, summary_age_buckets to SummaryOptions.AgeBuckets and stream_buffer_size to SummaryOptions.BufCap.

In the configuration, one may also set the observer type to “histogram”. For example, to set the observer type for a single timer metric:

mappings:
- match: "test.timing.*.*.*"
  observer_type: histogram
  histogram_options:
    buckets: [ 0.01, 0.025, 0.05, 0.1 ]
  name: "my_timer"
  labels:
    provider: "$2"
    outcome: "$3"
    job: "${1}_server"

If not set, then the default Prometheus client values are used for the histogram buckets: [.005, .01, .025, .05, .1, .25, .5, 1, 2.5, 5, 10]. +Inf is added automatically.

observer_type is only used when the statsd metric type is a timer, histogram, or distribution. buckets is only used when the statsd metric type is one of these, and the observer_type is set to histogram.

Timers will be accepted with the ms statsd type. Statsd timer data is transmitted in milliseconds, while Prometheus expects the unit to be seconds. The exporter converts all timer observations to seconds.

Histogram and distribution events (h and d metric type) are not subject to unit conversion.

DogStatsD Client Behavior

timed() decorator

The DogStatsD client’s timed decorator emits the metric in seconds but uses the ms type. Set use_ms=True to send the correct units.

Regular expression matching

Another capability when using YAML configuration is the ability to define matches using raw regular expressions as opposed to the default globbing style of match. This may allow for pulling structured data from otherwise poorly named statsd metrics AND allow for more precise targetting of match rules. When no match_type parameter is specified the default value of glob will be assumed:

mappings:
- match: "(.*)\.(.*)--(.*)\.status\.(.*)\.count"
  match_type: regex
  name: "request_total"
  labels:
    hostname: "$1"
    exec: "$2"
    protocol: "$3"
    code: "$4"

Global defaults

One may also set defaults for the observer type, buckets or quantiles, and match type. These will be used by all mappings that do not define them.

An option that can only be configured in defaults is glob_disable_ordering, which is false if omitted. By setting this to true, glob match type will not honor the occurance of rules in the mapping rules file and always treat * as lower priority than a concrete string.

defaults:
  observer_type: histogram
  buckets: [.005, .01, .025, .05, .1, .25, .5, 1, 2.5 ]
  match_type: glob
  glob_disable_ordering: false
  ttl: 0 # metrics do not expire
mappings:
# This will be a histogram using the buckets set in `defaults`.
- match: "test.timing.*.*.*"
  name: "my_timer"
  labels:
    provider: "$2"
    outcome: "$3"
    job: "${1}_server"
# This will be a summary timer.
- match: "other.distribution.*.*.*"
  observer_type: summary
  name: "other_distribution"
  labels:
    provider: "$2"
    outcome: "$3"
    job: "${1}_server_other"

Choosing between glob or regex match type

Despite from the missing flexibility of using regular expression in mapping and formatting labels, glob matching is optimized to have better performance than regex in certain use cases. In short, glob will have best performance if the rules amount is not so less and captures (using of *) is not to much in a single rule. Whether disabling ordering in glob or not won’t have a noticable effect on performance in general use cases. In edge cases like the below however, disabling ordering will be beneficial:

a.*.*.*.*
a.b.*.*.*
a.b.c.*.*
a.b.c.d.*

The reason is that the list assignment of captures (using of *) is the most expensive operation in glob. Honoring ordering will result in up to 10 list assignments, while without ordering it will need only 4 at most.

For details, see pkg/mapper/fsm/README.md. Running go test -bench . in pkg/mapper directory will produce a detailed comparison between the two match type.

drop action

You may also drop metrics by specifying a “drop” action on a match. For example:

mappings:
# This metric would match as normal.
- match: "test.timing.*.*.*"
  name: "my_timer"
  labels:
    provider: "$2"
    outcome: "$3"
    job: "${1}_server"
# Any metric not matched will be dropped because "." matches all metrics.
- match: "."
  match_type: regex
  action: drop
  name: "dropped"

You can drop any metric using the normal match syntax. The default action is “map” which does the normal metrics mapping.

Explicit metric type mapping

StatsD allows emitting of different metric types under the same metric name, but the Prometheus client library can’t merge those. For this use-case the mapping definition allows you to specify which metric type to match:

mappings:
- match: "test.foo.*"
  name: "test_foo"
  match_metric_type: counter
  labels:
    provider: "$1"

Possible values for match_metric_type are gauge, counter and observer.

Mapping cache size and cache replacement policy

There is a cache used to improve the performance of the metric mapping, that can greatly improvement performance. The cache has a default maximum of 1000 unique statsd metric names -> prometheus metrics mappings that it can store. This maximum can be adjust using the statsd.cache-size flag.

If the maximum is reached, entries are by default rotated using the least recently used replacement policy. This strategy is optimal when memory is constrained as only the most recent entries are retained.

Alternatively, you can choose a random-replacement cache strategy. This is less optimal if the cache is smaller than the cacheable set, but requires less locking. Use this for very high throughput, but make sure to allow for a cache that holds all metrics.

The optimal cache size is determined by the cardinality of the incoming metrics.

Time series expiration

The ttl parameter can be used to define the expiration time for stale metrics. The value is a time duration with valid time units: “ns”, “us” (or “µs”), “ms”, “s”, “m”, “h”. For example, ttl: 1m20s. 0 value is used to indicate metrics that do not expire.

TTL configuration is stored for each mapped metric name/labels combination whenever new samples are received. This means that you cannot immediately expire a metric only by changing the mapping configuration. At least one sample must be received for updated mappings to take effect.

### Event flushing configuration

Internally statsd_exporter runs a goroutine for each network listener (UDP, TCP & Unix Socket). These each receive and parse metrics received into an event. For performance purposes, these events are queued internally and flushed to the main exporter goroutine periodically in batches. The size of this queue and the flush criteria can be tuned with the --statsd.event-queue-size, --statsd.event-flush-threshold and --statsd.event-flush-interval. However, the defaults should perform well even for very high traffic environments.

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.yml:/tmp/statsd_mapping.yml \
        prom/statsd-exporter --statsd.mapping-config=/tmp/statsd_mapping.yml

Library packages

Parts of the implementation of this exporter are available as separate packages. See the documentation for details.

For the time being, there are no stability guarantees for library interfaces. We will try to call out any significant changes in the changelog. Semantic versioning of the exporter is based on the impact on users of the exporter, not users of the library.

We encourage re-use of these packages and welcome issues related to their usability as a library.