Introduction

The perfmetrics package provides a simple way to add software performance metrics to Python libraries and applications. Use perfmetrics to find the true bottlenecks in a production application.

The perfmetrics package is a client of the Statsd daemon by Etsy, which is in turn a client of Graphite (specifically, the Carbon daemon). Because the perfmetrics package sends UDP packets to Statsd, perfmetrics adds no I/O delays to applications and little CPU overhead. It can work equally well in threaded (synchronous) or event-driven (asynchronous) software.

Usage

Use the @metric and @metricmethod decorators to wrap functions and methods that should send timing and call statistics to Statsd. Add the decorators to any function or method that could be a bottleneck, including library functions.

Sample:

from perfmetrics import metric
from perfmetrics import metricmethod

@metric
def myfunction():
    """Do something that might be expensive"""

class MyClass(object):
    @metricmethod
    def mymethod(self):
        """Do some other possibly expensive thing"""

Next, tell perfmetrics how to connect to Statsd. (Until you do, the decorators have no effect.) Ideally, either your application should read the Statsd URI from a configuration file at startup time, or you should set the STATSD_URI environment variable. The example below uses a hard-coded URI:

from perfmetrics import set_statsd_client
set_statsd_client('statsd://localhost:8125')

for i in xrange(1000):
    myfunction()
    MyClass().mymethod()

If you run that code, it will fire 2000 UDP packets at port 8125. However, unless you have already installed Graphite and Statsd, all of those packets will be ignored and dropped. Dropping is a good thing: you don't want your production application to fail or slow down just because your performance monitoring system is stopped or not working.

Install Graphite and Statsd to receive and graph the metrics. One good way to install them is the graphite_buildout example at github, which installs Graphite and Statsd in a custom location without root access.

Pyramid and WSGI

If you have a Pyramid app, you can set the statsd_uri for each request by including perfmetrics in your configuration:

config = Configuration(...)
config.include('perfmetrics')

Also add a statsd_uri setting such as statsd://localhost:8125. Once configured, the perfmetrics tween will set up a Statsd client for the duration of each request. This is especially useful if you run multiple apps in one Python interpreter and you want a different statsd_uri for each app.

Similar functionality exists for WSGI apps. Add the app to your Paste Deploy pipeline:

[statsd]
use = egg:perfmetrics#statsd
statsd_uri = statsd://localhost:8125

[pipeline:main]
pipeline =
    statsd
    egg:myapp#myentrypoint

Threading

While most programs send metrics from any thread to a single global Statsd server, some programs need to use a different Statsd server for each thread. If you only need a global Statsd server, use the set_statsd_client function at application startup. If you need to use a different Statsd server for each thread, use the statsd_client_stack object in each thread. Use the push, pop, and clear methods.

Graphite Tips

Graphite stores each metric as a time series with multiple resolutions. The sample graphite_buildout stores 10 second resolution for 48 hours, 1 hour resolution for 31 days, and 1 day resolution for 5 years. To produce a coarse grained value from a fine grained value, Graphite computes the mean value (average) for each time span.

Because Graphite computes mean values implicitly, the most sensible way to treat counters in Graphite is as a "hits per second" value. That way, a graph can produce correct results no matter which resolution level it uses.

Treating counters as hits per second has unfortunate consequences, however. If some metric sees a 1000 hit spike in one second, then falls to zero for at least 9 seconds, the Graphite chart for that metric will show a spike of 100, not 1000, since Graphite receives metrics every 10 seconds and the spike looks to Graphite like 100 hits per second over a 10 second period.

If you want your graph to show 1000 hits rather than 100 hits per second, apply the Graphite hitcount() function, using a resolution of 10 seconds or more. The hitcount function converts per-second values to approximate raw hit counts. Be sure to provide a resolution value large enough to be represented by at least one pixel width on the resulting graph, otherwise Graphite will compute averages of hit counts and produce a confusing graph.

It usually makes sense to treat null values in Graphite as zero, though that is not the default; by default, Graphite draws nothing for null values. You can turn on that option for each graph.

Reference Documentation

Decorators

@metric

Notifies Statsd using UDP every time the function is called. Sends both call counts and timing information. The name of the metric sent to Statsd is <module>.<function name>.

@metricmethod

Like @metric, but the name of the Statsd metric is <class module>.<class name>.<method name>.

Metric(stat=None, rate=1, method=False, count=True, timing=True)

A decorator or context manager with options.

stat is the name of the metric to send; set it to None to use the name of the function or method. rate lets you reduce the number of packets sent to Statsd by selecting a random sample; for example, set it to 0.1 to send one tenth of the packets. If the method parameter is true, the default metric name is based on the method's class name rather than the module name. Setting count to False disables the counter statistics sent to Statsd. Setting timing to False disables the timing statistics sent to Statsd.

Sample use as a decorator:

@Metric('frequent_func', rate=0.1, timing=False)
def frequent_func():
    """Do something fast and frequently"""

Sample use as a context manager:

def do_something():
    with Metric('doing_something'):
        pass

If perfmetrics sends packets too frequently, UDP packets may be lost and the application performance may be affected. You can reduce the number of packets and the CPU overhead using the Metric decorator with options instead of metric or metricmethod. The decorator example above uses a sample rate and a static metric name. It also disables the collection of timing information.

When using Metric as a context manager, you must provide the stat parameter or nothing will be recorded.

Functions

statsd_client()

Return the currently configured StatsdClient. Returns the thread-local client if there is one, or the global client if there is one, or None.

set_statsd_client(client_or_uri)

Set the global StatsdClient. The client_or_uri can be a StatsdClient, a statsd:// URI, or None. Note that when the perfmetrics module is imported, it looks for the STATSD_URI environment variable and calls set_statsd_client() if that variable is found.

statsd_client_from_uri(uri)

Create a StatsdClient from a URI, but do not install it as the global StatsdClient. A typical URI is statsd://localhost:8125. Supported optional query parameters are prefix and gauge_suffix. The default prefix is empty and the default gauge_suffix is .<host_name>. See the StatsdClient documentation for more information about gauge_suffix.

StatsdClient Methods

Python code can send custom metrics by first getting the current StatsdClient using the statsd_client() function. Note that statsd_client() returns None if no client has been configured.

Most of the methods below have optional rate, rate_applied, and buf parameters. The rate parameter, when set to a value less than 1, causes StatsdClient to send a random sample of packets rather than every packet. The rate_applied parameter, if true, informs StatsdClient that the sample rate has already been applied and the packet should be sent regardless of the specified sample rate.

If the buf parameter is a list, StatsdClient appends the packet contents to the buf list rather than send the packet, making it possible to send multiple updates in a single packet. Keep in mind that the size of UDP packets is limited (the limit varies by the network, but 1000 bytes is usually a good guess) and any extra bytes will be ignored silently.

timing(stat, value, rate=1, buf=None, rate_applied=False)

Record timing information. stat is the name of the metric to record and value is the timing measurement in milliseconds. Note that Statsd maintains several data points for each timing metric, so timing metrics can take more disk space than counters or gauges.

gauge(stat, value, suffix=None, rate=1, buf=None, rate_applied=False)

Update a gauge value. stat is the name of the metric to record and value is the new gauge value. A gauge represents a persistent value such as a pool size. Because gauges from different machines often conflict, a suffix is usually applied to gauge names. If the suffix parameter is a string (including an empty string), it overrides the default gauge suffix.

incr(stat, count=1, rate=1, buf=None, rate_applied=False)

Increment a counter by count. Note that Statsd clears all counter values every time it sends the metrics to Graphite, which usually happens every 10 seconds. If you need a persistent value, it may be more appropriate to use a gauge instead of a counter.

decr(stat, count=1, rate=1, buf=None, rate_applied=False)

Decrement a counter by count.

sendbuf(buf)

Send the contents of the buf list to Statsd.