python-statsd

python-statsd is a client for statsd, the metrics aggregation daemon that started at Etsy and now lives in its own organisation, sitting in front of Graphite. It supports Python 3.10 and newer, and it has no dependencies.

pip install python-statsd

import statsd

counter = statsd.Counter('app')
counter += 1

with statsd.Timer('app').time('render'):
    pass  # the work you are measuring

Two metrics, two UDP packets, nothing blocking. A statsd server that is down costs you graphs rather than requests.

What goes on the wire

Every line under udp :8125 < in this recording is a packet the client sent, caught by a real listener on the other end:

Note the last one. At sample_rate=0.5 four increments produced a single packet, and it carries |@0.5 so the server knows to multiply back up.

The metric types

Type	A burst of values becomes	Reach for it when
Counter	their sum	you are counting events
Gauge	the last one	you are reporting a level
Timer	mean, median, percentiles	you are measuring duration
Average	their mean	the server should average samples
Raw	stored as sent	you already did the summarising

import statsd

counter = statsd.Counter('app')
counter.increment('requests')  # app.requests:1|c

gauge = statsd.Gauge('app')
gauge.send('queue_depth', 42)  # app.queue_depth:42|g

average = statsd.Average('app')
average.send('batch', 123)  # app.batch:123|a

raw = statsd.Raw('app')
raw.send('summary', 42, timestamp=1234567890)  # app.summary:42|r|1234567890

Timers come in three forms, and the context manager is the one to reach for, because it reports the block that raised as well as the block that did not:

import statsd

timer = statsd.Timer('app')

with timer.time('render'):
    pass  # the work you are measuring


@timer.decorate
def render_page():  # sends app.render_page
    pass

Names build themselves when you nest clients, which keeps the string formatting out of your call sites:

import statsd

app = statsd.Client('app')
queries = app.get_client('database').get_client('queries', statsd.Counter)

queries.increment()  # app.database.queries:1|c

See it working

The repository ships a compose file with the real statsd, Graphite and Grafana, so you can watch a metric arrive instead of taking anyone's word for it:

docker compose up -d --wait
uv run python examples/send_metrics.py --seconds 120

--wait returns once every service reports healthy. No local Python? docker compose --profile demo up -d --wait generates the traffic from a container instead. Then open http://localhost:3000: Grafana has the datasource configured and this dashboard loaded, no login in the way:

That screenshot is the stack in this repository, fed by examples/send_metrics.py through this client. statsd runs from examples/statsd/config.js, so the aggregation you see is configured in a file you can read and change.

Already running Grafana? Import the dashboard straight into it from examples/grafana/import/python-statsd.json, which asks which datasource to use instead of assuming one. The local stack guide covers how statsd renames your metrics on the way through, and what to check when nothing shows up.

Configuration

Set the defaults once at startup and every client built afterwards follows:

import statsd

statsd.Connection.set_defaults(host='localhost', port=8125, sample_rate=1)

Or build connections yourself when one destination is not enough:

import statsd

connection = statsd.Connection(host='statsd-1', port=8125, sample_rate=0.1)
statsd.Counter('app.requests', connection).increment()

One trap worth knowing before it costs you an afternoon: a falsy argument means "use the default", so Connection(sample_rate=0) sends everything. Pass disabled=True to send nothing.

Documentation

Full documentation is at python-statsd.readthedocs.io.

• Quickstart: nothing to metrics on a graph, in two tracks
• Metrics: the five types, how to choose, and the exact bytes each one writes
• Connections: destinations, sampling, disabling, failure behaviour, threads and forks
• Patterns: naming, cardinality, client trees, WSGI and Celery integration
• Local stack: docker compose, and how to debug a metric that never arrives

For Django, use django-statsd, the sister project built on this client. It times views and reports the queries per request without you writing any of it.

Note on the package name

This project is published on PyPI as python-statsd and installs a module called statsd. A different project, jsocol's client, is published as statsd and installs a module called statsd as well. Installing both in one environment leaves you with whichever was written last, so pick one.

Contributing

Bug reports and patches are welcome, and CONTRIBUTING.md covers the development setup: uv sync --all-extras, uv run pytest, and uv run tox -p auto to run everything CI runs. Every code sample in this README and in the documentation is executed by the test suite, so a change in behaviour tends to tell you which paragraph it just made wrong.

Links

• Source: https://github.com/WoLpH/python-statsd
• Issues: https://github.com/WoLpH/python-statsd/issues
• Statsd: https://github.com/statsd/statsd
• Graphite: https://graphiteapp.org/

Support

python-statsd is maintained by Rick van Hattem in his own time.

If it saved you an afternoon, a tip covers an hour of issue triage: Ko-fi or GitHub Sponsors.

If your company funds its dependencies, this package is on thanks.dev.