gostatsd

An implementation of Etsy's statsd in Go, based on original code from @kisielk.

The project provides both a server called "gostatsd" which works much like Etsy's version, but also provides a library for developing customized servers.

Backends are pluggable and only need to support the backend interface.

Being written in Go, it is able to use all cores which makes it easy to scale up the server based on load.

Building the server

Gostatsd currently targets Go 1.15.6. If you are compiling from source, please ensure you are running this version.

From the gostatsd directory run make build. The binary will be built in build/bin/<arch>/gostatsd.

You will need to install the Golang build dependencies by running make setup in the gostatsd directory. This must be done before the first build, and again if the dependencies change. A protobuf installation is expected to be found in the tools/ directory. Managing this in a platform agnostic way is difficult, but PRs are welcome. Hopefully it will be sufficient to use the generated protobuf files in the majority of cases.

If you are unable to build gostatsd please check your Go version, and try running make setup again before reporting a bug.

Running the server

gostatsd --help gives a complete description of available options and their defaults. You can use make run to run the server with just the stdout backend to display info on screen.

You can also run through docker by running make run-docker which will use docker-compose to run gostatsd with a graphite backend and a grafana dashboard.

While not generally tested on Windows, it should work. Maximum throughput is likely to be better on a linux system, however.

Configuring the server mode

The server can currently run in two modes: standalone and forwarder. It is configured through the top level server-mode configuration setting. The default is standalone.

In standalone mode, raw metrics are processed and aggregated as normal, and aggregated data is submitted to configured backends (see below)

This configuration mode allows the following configuration options:

• expiry-interval: interval before metrics are expired, see Metric expiry and persistence section. Defaults to 5m. 0 to disable, -1 for immediate.
• expiry-interval-counter: interval before counters are expired, defaults to the value of expiry-interval.
• expiry-interval-gauge: interval before gauges are expired, defaults to the value of expiry-interval.
• expiry-interval-set: interval before sets are expired, defaults to the value of expiry-interval.
• expiry-interval-timer: interval before timers are expired, defaults to the value of expiry-interval.
• flush-aligned: whether or not the flush should be aligned. Setting this will flush at an exact time interval. With a 10 second flush-interval, if the service happens to be started at 12:47:13, then flushing will occur at 12:47:20, 12:47:30, etc, rather than 12:47:23, 12:47:33, etc. This removes query time ambiguity in a multi-server environment. Defaults to false.
• flush-interval: duration for how long to batch metrics before flushing. Should be an order of magnitude less than the upstream flush interval. Defaults to 1s.
• flush-offset: offset for flush interval when flush alignment is enabled. For example, with an offset of 7s and an interval of 10s, it will flush at 12:47:10+7 = 12:47:17, etc.
• ignore-host: indicates whether or not an explicit host field will be added to all incoming metrics and events. Defaults to false
• max-readers: the number of UDP receivers to run. Defaults to 8 or the number of logical cores, whichever is less.
• max-parsers: the number of workers available to parse metrics. Defaults to the number of logical cores.
• max-workers: the number of aggregators to process metrics. Defaults to the number of logical cores.
• max-queue-size: the size of the buffers between parsers and workers. Defaults to 10000, monitored via channel.* metric, with dispatch_aggregator_batch and dispatch_aggregator_map channels.
• max-concurrent-events: the maximum number of concurrent events to be dispatching. Defaults to 1024, monitored via channel.* metric, with backend_events_sem channel.
• estimated-tags: provides a hint to the system as to how many tags are expected to be seen on any particular metric, so that memory can be pre-allocated and reducing churn. Defaults to 4. Note: this is only a hint, and it is safe to send more.
• log-raw-metric: logs raw metrics received from the network. Defaults to false.
• metrics-addr: the address to listen to metrics on. Defaults to :8125.
• namespace: a namespace to prefix all metrics with. Defaults to ''.
• statser-type: configures where internal metrics are sent to. May be internal which sends them to the internal processing pipeline, logging which logs them, null which drops them. Defaults to internal, or null if the NewRelic backend is enabled.
• percent-threshold: configures the "percentiles" sent on timers. Space separated string. Defaults to 90.
• heartbeat-enabled: emits a metric named heartbeat every flush interval, tagged by version and commit. Defaults to false.
• receive-batch-size: the number of datagrams to attempt to read. It is more CPU efficient to read multiple, however it takes extra memory. See [Memory allocation for read buffers] section below for details. Defaults to 50.
• conn-per-reader: attempts to create a connection for every UDP receiver. Not supported by all OS versions. Defaults to false.
• bad-lines-per-minute: the number of metrics which fail to parse to log per minute. This is used to prevent a bad client spamming malformed statsd data, while still logging some information to enable troubleshooting. Defaults to 0.
• hostname: sets the hostname on internal metrics
• timer-histogram-limit: specifies the maximum number of buckets on histograms. See [Timer histograms] below.

In forwarder mode, raw metrics are collected from a frontend, and instead of being aggregated they are sent via http to another gostatsd server after passing through the processing pipeline (cloud provider, static tags, filtering, etc).

A forwarder server is intended to run on-host and collect metrics, forwarding them on to a central aggregation service. At present the central aggregation service can only scale vertically, but horizontal scaling through clustering is planned.

Aligned flushing is deliberately not supported in forwarder mode, as it would impact the central aggregation server due to all for forwarder nodes transmitting at once, and the expectation that many forwarding flushes will occur per central flush anyway.

Configuring forwarder mode requires a configuration file, with a section named http-transport. The raw version spoken is not configurable per server (see HTTP.md for version guarantees). The configuration section allows the following configuration options:

• compress: boolean indicating if the payload should be compressed. Defaults to true
• api-endpoint: configures the endpoint to submit raw metrics to. This setting should be just a base URL, for example https://statsd-aggregator.private, with no path. Required, no default
• max-requests: maximum number of requests in flight. Defaults to 1000 (which is probably too high)
• max-request-elapsed-time: duration for the maximum amount of time to try submitting data before giving up. This includes retries. Defaults to 30s (which is probably too high). Setting this value to -1 will disable retries.
• consolidator-slots: number of slots in the metric consolidator. Memory usage is a function of this. Lower values may cause blocking in the pipeline (back pressure). A UDP only receiver will never use more than the number of configured parsers (--max-parsers option). Defaults to the value of --max-parsers, but may require tuning for HTTP based servers.
• transport: see TRANSPORT.md for how to configure the transport.
• custom-headers : a map of strings that are added to each request sent to allow for additional network routing / request inspection. Not required, default is empty. Example: --custom-headers='{"region" : "us-east-1", "service" : "event-producer"}'
• dynamic-headers : similar with custom-headers, but the header values are extracted from metric tags matching the provided list of string. Tag names are canonicalized by first replacing underscores with hyphens, then converting first letter and each letter after a hyphen to uppercase, the rest are converted to lower case. If a tag is specified in both custom-header and dynamic-header, the vaule set by custom-header takes precedence. Not required, default is empty. Example: --dynamic-headers='["region", "service"]'. This is an experimental feature and it may be removed or changed in future versions.

The following settings from the previous section are also supported:

• expiry-*
• ignore-host
• max-readers
• max-parsers
• estimated-tags
• log-raw-metric
• metrics-addr
• namespace
• statser-type
• heartbeat-enabled
• receive-batch-size
• conn-per-reader
• bad-lines-per-minute
• hostname
• log-raw-metric

Metric expiry and persistence

After a metric has been sent to the server, the server will continue to send the metric to the configured backend until it expires, even if no additional metrics are sent from the client. The value sent depends on the metric type:

• counter: sends 0 for both rate and count
• gauge: sends the last received value.
• set: sends 0
• timer: sends non-percentile values of 0. Percentile values are not sent at all (see issue #135)

Setting an expiry interval of 0 will persist metrics forever. If metrics are not carefully controlled in such an environment, the server may run out of memory or overload the backend receiving the metrics. Setting a negative expiry interval will result in metrics not being persisted at all.

Each metric type has its own interval, which is configured using the following precedence (from highest to lowest): expiry-interval-<type> > expiry-interval > default (5 minutes).

Configuring HTTP servers

The service supports multiple HTTP servers, with different configurations for different requirements. All http servers are named in the top level http-servers setting. It should be a space separated list of names. Each server is then configured by creating a section in the configuration file named http.<servername>. An http server section has the following configuration options:

• address: the address to bind to
• enable-prof: boolean indicating if profiler endpoints should be enabled. Default false
• enable-expvar: boolean indicating if expvar endpoints should be enabled. Default false
• enable-ingestion: boolean indicating if ingestion should be enabled. Default false
• enable-healthcheck: boolean indicating if healthchecks should be enabled. Default true

For example, to configure a server with a localhost only diagnostics endpoint, and a regular ingestion endpoint that can sit behind an ELB, the following configuration could be used:

backends='stdout'
http-servers='receiver profiler'

[http.receiver]
address='0.0.0.0:8080'
enable-ingestion=true

[http.profiler]
address='127.0.0.1:6060'
enable-expvar=true
enable-prof=true

There is no capability to run an https server at this point in time, and no auth (which is why you might want different addresses). You could also put a reverse proxy in front of the service. Documentation for the endpoints can be found under HTTP.md

Configuring backends

Refer to backends for configuration options for the backends.

Cloud providers

Cloud providers are a way to automatically enrich metrics with metadata from a cloud vendor.

Refer to cloud providers for configuration options for the cloud providers.

Configuring timer sub-metrics

By default, timer metrics will result in aggregated metrics of the form (exact name varies by backend):

<base>.Count
<base>.CountPerSecond
<base>.Mean
<base>.Median
<base>.Lower
<base>.Upper
<base>.StdDev
<base>.Sum
<base>.SumSquares

In addition, the following aggregated metrics will be emitted for each configured percentile:

<base>.Count_XX
<base>.Mean_XX
<base>.Sum_XX
<base>.SumSquares_XX
<base>.Upper_XX - for positive only
<base>.Lower_-XX - for negative only

These can be controlled through the disabled-sub-metrics configuration section:

[disabled-sub-metrics]
# Regular metrics
count=false
count-per-second=false
mean=false
median=false
lower=false
upper=false
stddev=false
sum=false
sum-squares=false

# Percentile metrics
count-pct=false
mean-pct=false
sum-pct=false
sum-squares-pct=false
lower-pct=false
upper-pct=false

By default (for compatibility), they are all false and the metrics will be emitted.

Timer histograms (experimental feature)

Timer histograms inspired by Prometheus implementation can be enabled on a per time series basis using gsd_histogram meta tag with value containing histogram bucketing definition (joined with _) e.g. gsd_histogram:-10_0_2.5_5_10_25_50.

It will:

• output additional counter time series with name <base>.histogram and le tags specifying histogram buckets.
• disable default sub-aggregations for timers e.g. <base>.Count, <base>.Mean, <base>.Upper, <base>.Upper_XX, etc.

For timer with gsd_histogram:-10_0_2.5_5_10_25_50 meta tag, following time series will be generated

• <base>.histogram with tag le:-10
• <base>.histogram with tag le:0
• <base>.histogram with tag le:2.5
• <base>.histogram with tag le:5
• <base>.histogram with tag le:10
• <base>.histogram with tag le:25
• <base>.histogram with tag le:50
• <base>.histogram with tag le:+Inf

Each time series will contain a total number of timer data points that had a value less or equal le value, e.g. counter <base>.histogram with the tag le:5 will contain the number of all observations that had a value not bigger than 5. Counter <base>.histogram with tag le:+Inf is equivalent to <base>.count and contains the total number.

All original timer tags are preserved and added to all the time series.

To limit cardinality, timer-histogram-limit option can be specified to limit the number of buckets that will be created (default is math.MaxUint32). Value of 0 won't disable the feature, 0 buckets will be emitted which effectively drops metrics with gsd_hostogram tags.

Incorrect meta tag values will be handled in best effort manner, i.e.

• gsd_histogram:10__20_50 & gsd_histogram:10_incorrect_20_50 will generate le:10, le:20, le:50 and le:+Inf buckets
• gsd_histogram:incorrect will result in only le:+Inf bucket

This is an experimental feature and it may be removed or changed in future versions.

Load testing

There is a tool under cmd/loader with support for a number of options which can be used to generate synthetic statsd load. There is also another load generation tool under cmd/tester which is deprecated and will be removed in a future release.

Help for the loader tool can be found through --help.

Sending metrics

The server listens for UDP packets on the address given by the --metrics-addr flag, aggregates them, then sends them to the backend servers given by the --backends flag (space separated list of backend names).

Currently supported backends are:

• cloudwatch
• datadog
• graphite
• influxdb
• newrelic
• statsdaemon
• stdout

The format of each metric is:

<bucket name>:<value>|<type>\n

• <bucket name> is a string like abc.def.g, just like a graphite bucket name
• <value> is a string representation of a floating point number
• <type> is one of c, g, or ms for "counter", "gauge", and "timer" respectively.

A single packet can contain multiple metrics, each ending with a newline.

Optionally, gostatsd supports sample rates (for simple counters, and for timer counters) and tags:

• <bucket name>:<value>|c|@<sample rate>\n where sample rate is a float between 0 and 1
• <bucket name>:<value>|c|@<sample rate>|#<tags>\n where tags is a comma separated list of tags
• <bucket name>:<value>|<type>|#<tags>\n where tags is a comma separated list of tags

Tags format is: simple or key:value.

A simple way to test your installation or send metrics from a script is to use echo and the netcat utility nc:

echo 'abc.def.g:10|c' | nc -w1 -u localhost 8125

Monitoring

Many metrics for the internal processes are emitted. See METRICS.md for details. Go expvar is also exposed if the --profile flag is used.

Memory allocation for read buffers

By default gostatsd will batch read multiple packets to optimise read performance. The amount of memory allocated for these read buffers is determined by the config options:

max-readers * receive-batch-size * 64KB (max packet size)

The metric avg_packets_in_batch can be used to track the average number of datagrams received per batch, and the --receive-batch-size flag used to tune it. There may be some benefit to tuning the --max-readers flag as well.

Using the library

In your source code:

import "github.com/atlassian/gostatsd/pkg/statsd"

Note that this project uses Go modules for dependency management.

Documentation can be found via go doc github.com/atlassian/gostatsd/pkg/statsd or at https://godoc.org/github.com/atlassian/gostatsd/pkg/statsd

Versioning

Gostatsd uses semver versioning for both API and configuration settings, however it does not use it for packages.

This is due to gostatsd being an application first and a library second. Breaking API changes occur regularly, and the overhead of managing this is too burdensome.

Contributors

Pull requests, issues and comments welcome. For pull requests:

• Add tests for new features and bug fixes
• Follow the existing style
• Separate unrelated changes into multiple pull requests

See the existing issues for things to start contributing.

For bigger changes, make sure you start a discussion first by creating an issue and explaining the intended change.

Atlassian requires contributors to sign a Contributor License Agreement, known as a CLA. This serves as a record stating that the contributor is entitled to contribute the code/documentation/translation to the project and is willing to have it used in distributions and derivative works (or is willing to transfer ownership).

Prior to accepting your contributions we ask that you please follow the appropriate link below to digitally sign the CLA. The Corporate CLA is for those who are contributing as a member of an organization and the individual CLA is for those contributing as an individual.

• CLA for corporate contributors
• CLA for individuals

License

Copyright (c) 2012 Kamil Kisiel. Copyright @ 2016-2020 Atlassian Pty Ltd and others.

Licensed under the MIT license. See LICENSE file.