Overview¶
docs | |
---|---|
tests | |
package |
dogstatsd-collector
is a library to make it easy to collect DataDog-style
StatsD counters
and histograms
with tags and control when they are flushed. It
gives you a drop-in wrapper for the DogStatsD library for counters and
histograms and allows you to defer flushing the metrics until you choose to. This
capability enables you to collect StatsD metrics at arbitrary granularity, for
example on a per-web request or per-job basis (instead of the per-flush
interval basis).
Counters and histograms are tracked separately for each metric series (unique
set of tag key-value pairs) and a single metric is emitted for each series when
the collector is flushed. You don’t have to think about tracking your metric
series separately; you just use the DogstatsdCollector
object as you would the
normal DogStatsD
object, and flush when you’re ready; the library will take
care of emitting all the series for you.
- Free software: BSD 3-Clause License
Installation¶
pip install dogstatsd-collector
Example Usage¶
Imagine you want to track a distribution of the number of queries issued by requests to your webapp, and tag them by which database is queried and which verb is used. You collect the following metrics as you issue your queries:
collector = DogstatsdCollector(dogstatsd)
...
collector.histogram('query', tags=['database:master','verb:insert'])
collector.histogram('query', tags=['database:master','verb:update'])
collector.histogram('query', tags=['database:master','verb:update'])
collector.histogram('query', tags=['database:replica','verb:select'])
collector.histogram('query', tags=['database:replica','verb:select'])
Then, at the end of your web request, when you flush the collector, the
following metrics will be pushed to DogStatsD
(shown in DogStatsD datagram
format):
collector.flush()
# query:1|h|#database:master,verb:insert
# query:2|h|#database:master,verb:update
# query:2|h|#database:replica,verb:select
Base Tags¶
The collector object also supports specifying a set of base tags, which will be included on every metric that gets emitted.
base_tags = ['mytag:myvalue']
collector = DogstatsdCollector(dogstatsd, base_tags=base_tags)
collector.histogram('query', tags=['database:master','verb:insert'])
collector.histogram('query', tags=['database:master','verb:update'])
collector.flush()
# query:1|h|#database:master,verb:insert,mytag:myvalue
# query:1|h|#database:master,verb:update,mytag:myvalue
Motivation¶
The StatsD model is to run an agent on each server/container in your infrastructure and periodically flush aggregations at a regular interval to a centralized location. This model scales very well because the volume of metrics sent to the centralized location grows very slowly even as you scale your application; each StatsD agent calculates aggregations to flush to the backend instead of every datapoint, so the storage volume is quite low even for a large application with lots of volume.
A drawback to this model is that you don’t have much control of the granularity that your metrics represent. When your aggregations reach the centralized location (DataDog in this case), you only know the counts or distributions within the flush interval. You can’t represent any other execution granularity beyond “across X seconds” (where X is the flush interval). This limitation precludes you from easily representings metrics on a “per-request” basis, for example.
The purpose of this library is to make it simple to control when your StatsD metrics are emitted so that you can defer emission of the metrics until a point you determine. This allows you to represent a finer granularity than “across X seconds” such as “across a web request” or “across a cron job.” It also preserves metric tags by emitting each series independently when the collector is flushed, which ensures you don’t lose any of the benefit of tagging your metrics (such as aggregating/slicing in DataDog).
Patterns¶
The DogstatsdCollector
object is a singleton that provides a similar
interface as the DogStatsD
increment
and histogram
methods. As you invoke these methods, you collect counters and histograms for
each series (determined by any tags you include). After calling flush()
,
each series is separately emitted as a StatsD metric.
Simple Request Metrics¶
You can collect various metrics over a request and emit them at the end of the request to get per-request granularity.
In Django:
from datadog.dogstatsd.base import DogStatsd
from dogstatsd_collector import DogstatsdCollector
# Middleware
class MetricsMiddleware:
def __init__(self, get_response):
self.get_response = get_response
self.dogstatsd = DogStatsd()
def __call__(self, request):
request.metrics = DogstatsdCollector(self.dogstatsd)
response = self.get_response(request)
request.metrics.flush()
return response
# Inside a view
def my_view(request):
# Do some stuff...
request.metrics.increment('my.count')
request.metrics.histogram('my.time', 0.5)
return HttpResponse('ok')
In Flask:
from datadog.dogstatsd.base import DogStatsd
from dogstatsd_collector import DogstatsdCollector
from flask import Flask
from flask import request
app = Flask(__name__)
dogstatsd = DogStatsd()
@app.before_request
def init_metrics():
request.metrics = DogstatsdCollector(dogstatsd)
@app.after_request
def flush_metrics():
request.metrics.flush()
@app.route('/')
def my_view():
# Do some stuff...
request.metrics.increment('my.count')
request.metrics.histogram('my.time', 0.5)
return 'ok'
Celery Task Metrics¶
Same as above, but over a Celery task.
from datadog.dogstatsd.base import DogStatsd
from dogstatsd_collector import DogstatsdCollector
from celery import Celery
from celery import current_task
from celery.signals import task_prerun
from celery.signals import task_postrun
app = Celery('tasks', broker='pyamqp://guest@localhost//')
dogstatsd = DogStatsd()
@task_prerun.connect
def init_metrics(task_id, task, *args, **kwargs):
task.request.metrics = DogstatsdCollector(dogstatsd)
@task_postrun.connect
def flush_metrics(task_id, task, *args, **kwargs):
task.request.metrics.flush()
@app.task
def my_task():
# Do some stuff...
current_task.request.metrics.increment('my.count')
current_task.request.metrics.histogram('my.time', 0.5)
Metrics Within a Function¶
Emit a set of metrics for a particular function you execute.
from datadog.dogstatsd.base import DogStatsd
from dogstatsd_collector import DogstatsdCollector
dogstatsd = DogStatsd()
def do_stuff(metrics):
# Do some stuff...
metrics.increment('my.count')
metrics.histogram('my.time', 0.5)
metrics = DogstatsdCollector(dogstatsd)
do_stuff(metrics)
metrics.flush()
Thread Safety¶
The DogstatsdCollector
singleton is not threadsafe. Do not share a
single DogstatsdCollector
object among multiple threads.
More Documentation¶
Full documentation can be found on ReadTheDocs: