AppMetrics

AppMetrics is a python library used to collect useful run-time application's metrics, based on Folsom from Boundary, which is in turn inspired by Metrics from Coda Hale.

The library's purpose is to help you collect real-time metrics from your Python applications, being them web apps, long-running batches or whatever. AppMetrics is not a persistent store, you must provide your own persistence layer, maybe by using well established monitoring tools.

AppMetrics works on python 2.7 and 3.3.

Getting started

Install AppMetrics into your python environment:

pip install appmetrics

or, if you don't use pip, download and unpack the package an then:

python setup.py install

Once you have installed AppMetrics you can access it by the metrics module:

>>> from appmetrics import metrics
>>> histogram = metrics.new_histogram("test")
>>> histogram.notify(1.0)
True
>>> histogram.notify(2.0)
True
>>> histogram.notify(3.0)
True
>>> histogram.get()
{'arithmetic_mean': 2.0, 'kind': 'histogram', 'skewness': 0.0, 'harmonic_mean': 1.6363636363636365, 'min': 1.0, 'standard_deviation': 1.0, 'median': 2.0, 'histogram': [(3.0, 3), (5.0, 0)], 'percentile': [(50, 2.0), (75, 2.0), (90, 3.0), (95, 3.0), (99, 3.0), (99.9, 3.0)], 'n': 3, 'max': 3.0, 'variance': 1.0, 'geometric_mean': 1.8171205928321397, 'kurtosis': -2.3333333333333335}

Basically you create a new metric by using one of the metrics.new_* functions. The metric will be stored into an internal registry, so you can access it in different places in your application:

>>> test_histogram = metrics.metric("test")
>>> test_histogram.notify(4.0)
True

The metrics registry is thread-safe, you can safely use it in multi-threaded web servers.

Now let's decorate some function:

>>> import time, random
>>> @metrics.with_histogram("test")
... def my_worker():
...     time.sleep(random.random())
...
>>> my_worker()
>>> my_worker()
>>> my_worker()

and let's see the results:

>>> metrics.get("test")
{'arithmetic_mean': 0.41326093673706055, 'kind': 'histogram', 'skewness': 0.2739718270714368, 'harmonic_mean': 0.14326954591313346, 'min': 0.0613858699798584, 'standard_deviation': 0.4319169569113129, 'median': 0.2831099033355713, 'histogram': [(1.0613858699798584, 3), (2.0613858699798584, 0)], 'percentile': [(50, 0.2831099033355713), (75, 0.2831099033355713), (90, 0.895287036895752), (95, 0.895287036895752), (99, 0.895287036895752), (99.9, 0.895287036895752)], 'n': 3, 'max': 0.895287036895752, 'variance': 0.18655225766752892, 'geometric_mean': 0.24964828731906127, 'kurtosis': -2.3333333333333335}

Let's print the metrics data on the screen every 5 seconds:

>>> from appmetrics import reporter
>>> def stdout_report(metrics):
...     print metrics
...
>>> reporter.register(stdout_report, reporter.fixed_interval_scheduler(5))
'5680173c-0279-46ec-bd88-b318f8058ef4'
>>> {'test': {'arithmetic_mean': 0.0, 'kind': 'histogram', 'skewness': 0.0, 'harmonic_mean': 0.0, 'min': 0, 'standard_deviation': 0.0, 'median': 0.0, 'histogram': [(0, 0)], 'percentile': [(50, 0.0), (75, 0.0), (90, 0.0), (95, 0.0), (99, 0.0), (99.9, 0.0)], 'n': 0, 'max': 0, 'variance': 0.0, 'geometric_mean': 0.0, 'kurtosis': 0.0}}
>>> my_worker()
>>> my_worker()
>>> {'test': {'arithmetic_mean': 0.5028266906738281, 'kind': 'histogram', 'skewness': 0.0, 'harmonic_mean': 0.2534044030939462, 'min': 0.14868521690368652, 'standard_deviation': 0.50083167520453, 'median': 0.5028266906738281, 'histogram': [(1.1486852169036865, 2), (2.1486852169036865, 0)], 'percentile': [(50, 0.14868521690368652), (75, 0.8569681644439697), (90, 0.8569681644439697), (95, 0.8569681644439697), (99, 0.8569681644439697), (99.9, 0.8569681644439697)], 'n': 2, 'max': 0.8569681644439697, 'variance': 0.2508323668881758, 'geometric_mean': 0.35695727672917066, 'kurtosis': -2.75}}
>>> reporter.remove('5680173c-0279-46ec-bd88-b318f8058ef4')
<Timer(Thread-1, started daemon 4555313152)>

Decorators

The metrics module also provides a couple of decorators: with_histogram and with_meter which are an easy and fast way to use AppMetrics: just decorate your functions/methods and you will have metrics collected for them. You can decorate multiple functions with the same metric's name, as long as the decorator's type and parameters are the same, or a DuplicateMetricError will be raised. See the documentation for Histograms and Meters for more details.

API

AppMetrics exposes a simple and consistent API; all the metric objects have three methods:

• notify(value) - add a new value to the metric
• get() - get the computed metric's value (if any)
• raw_data() - get the raw data stored in the metrics

However, the notify input type and the get() and raw_data() data format depend on the kind of metric chosen. Please notice that get() returns a dictionary with the mandatory field kind which depends on the metric's type.

Metrics

Several metric types are available:

Counters

Counter metrics provide increment and decrement capabilities for a single integer value. The notify method accepts an integer: the counter will be incremented or decremented according to the value's sign. Notice that the function tries to cast the input value to integer, so a TypeError or a ValueError may be raised:

>>> counter = metrics.new_counter("test")
>>> counter.notify(10)
>>> counter.notify(-5)
>>> counter.get()
{'kind': 'counter', 'value': 5}
>>> counter.notify("wrong")
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "appmetrics/simple_metrics.py", line 40, in notify
    value = int(value)
ValueError: invalid literal for int() with base 10: 'wrong'

Gauges

Gauges are point-in-time single value metrics. The notify method accepts any data type:

>>> gauge = metrics.new_gauge("gauge_test")
>>> gauge.notify("version 1.0")
>>> gauge.get()
{'kind': 'gauge', 'value': 'version 1.0'}

The gauge metric is useful to expose almost-static values such as configuration parameters, constants and so on. Although you can use any python data type as the value, you won't be able to use the wsgi middleware unless you use a valid json type.

Histograms

Histograms are collections of values on which statistical analysis are performed automatically. They are useful to know how the application is performing. The notify method accepts a single floating-point value, while the get method computes and returns the following values:

• arithmetic mean
• geometric mean
• harmonic mean
• data distribution histogram with automatic bins
• kurtosis
• maximum value
• median
• minimum value
• number of values
• 50, 75, 90, 95, 99 and 99.9th percentiles of the data distribution
• skewness
• standard deviation
• variance

Notice that the notify method tries to cast the input value to a float, so a TypeError or a ValueError may be raised.

You can use the histogram metric also by the with_histogram decorator: the time spent in the decorated function will be collected by an histogram with the given name:

>>> @metrics.with_histogram("histogram_test")
... def fun(v):
...     return v*2
...
>>> fun(10)
20
>>> metrics.metric("histogram_test").raw_data()
[5.9604644775390625e-06]

The full signature is:

with_histogram(name, reservoir_type, *reservoir_args, **reservoir_kwargs)

where:

• name is the metric's name
• reservoir_type is a string which identifies a reservoir class, see reservoirs documentation
• reservoir_args and reservoir_kwargs are passed to the chosen reservoir's __init__

Sample types

To avoid unbound memory usage, the histogram metrics are generated from a reservoir of values.

Uniform reservoir

The default reservoir type is the uniform one, in which a fixed number of values (default 1028) is kept, and when the reservoir is full new values replace older ones randomly with an uniform probability distribution, ensuring that the sample is always statistically representative. This kind of reservoir must be used when you are interested in statistics over the whole stream of observations. Use "uniform" as reservoir_type in with_histogram.

Sliding window reservoir

This reservoir keeps a fixed number of observations (default 1028) and when a new value comes in the first one is discarded. The statistics are representative of the last N observations. Its reservoir_type is sliding_window.

Sliding time window reservoir

This reservoir keeps observation for a fixed amount of time (default 60 seconds), older values get discarded. The statistics are representative of the last N seconds, but if you have a lot of readings in N seconds this could eat a lot amount of memory. Its reservoir_type is sliding_time_window.

Exponentially-decaying reservoir

This reservoir keeps a fixed number of values (default 1028), with exponential decaying of older values in order to give greater significance to recent data. The bias towards newer values can be adjusted by specifying a proper alpha value to the reservoir's init (defaults to 0.015). Its reservoir_type is exp_decaying.

Meters

Meters are increment-only counters that measure the rate of events (such as "http requests") over time. This kind of metric is useful to collect throughput values (such as "requests per second"), both on average and on different time intervals:

>>> meter = metrics.new_meter("meter_test")
>>> meter.notify(1)
>>> meter.notify(1)
>>> meter.notify(3)
>>> meter.get()
{'count': 5, 'kind': 'meter', 'five': 0.0066114184713530035, 'mean': 0.27743058841197027, 'fifteen': 0.0022160607980413085, 'day': 2.3147478365093123e-05, 'one': 0.031982234148270686}

The return values of the get method are the following:

• count: number of operations collected so far
• mean: the average throughput since the metric creation
• one: one-minute exponentially-weighted moving average (EWMA)
• five: five-minutes EWMA
• fifteen: fifteen-minutes EWMA
• day: last day EWMA
• kind: "meter"

Notice that the notify method tries to cast the input value to an integer, so a TypeError or a ValueError may be raised.

You can use the meter metric also by the with_meter decorator: the number of calls to the decorated function will be collected by a meter with the given name.

Tagging

You can group several metrics together by "tagging" them:

>>> metrics.new_histogram("test1")
<appmetrics.histogram.Histogram object at 0x10ac2a950>
>>> metrics.new_gauge("test2")
<appmetrics.simple_metrics.Gauge object at 0x10ac2a990>
>>> metrics.new_meter("test3")
<appmetrics.meter.Meter object at 0x10ac2a9d0>
>>> metrics.tag("test1", "group1")
>>> metrics.tag("test3", "group1")
>>> metrics.tags()
{'group1': set(['test1', 'test3'])}
>>> metrics.metrics_by_tag("group1")
{'test1': {'arithmetic_mean': 0.0, 'skewness': 0.0, 'harmonic_mean': 0.0, 'min': 0, 'standard_deviation': 0.0, 'median': 0.0, 'histogram': [(0, 0)], 'percentile': [(50, 0.0), (75, 0.0), (90, 0.0), (95, 0.0), (99, 0.0), (99.9, 0.0)], 'n': 0, 'max': 0, 'variance': 0.0, 'geometric_mean': 0.0, 'kurtosis': 0.0}, 'test3': {'count': 0, 'five': 0.0, 'mean': 0.0, 'fifteen': 0.0, 'day': 0.0, 'one': 0.0}}

As you can see above, three functions are available:

• metrics.tag(metric_name, tag_name): tag the metric named <metric_name> with <tag_name>. Raise InvalidMetricError if <metric_name> does not exist.
• metrics.tags(): return the currently defined tags.
• metrics.metrics_by_tag(tag_name): return a dictionary with metric names as keys and metric values as returned by <metric_object>.get(). Return an empty dictionary if tag_name does not exist.

External access

You can access the metrics provided by AppMetrics externally by the WSGI middleware found in appmetrics.wsgi.AppMetricsMiddleware. It is a standard WSGI middleware without external dependencies and it can be plugged in any framework supporting the WSGI standard, for example in a Flask application:

from flask import Flask
from appmetrics import metrics

metrics.new_histogram("test-histogram")
metrics.new_gauge("test-counter")
metrics.metric("test-counter").notify(10)

app = Flask(__name__)

@app.route('/hello')
def hello_world():
    return 'Hello World!'

if __name__ == '__main__':
    from appmetrics.wsgi import AppMetricsMiddleware
    app.wsgi_app = AppMetricsMiddleware(app.wsgi_app)
    app.run()

If you launch the above application you can ask for metrics:

$ curl http://localhost:5000/hello
Hello World!

$ curl http://localhost:5000/_app-metrics
["test-counter", "test-histogram"]

$ curl http://localhost:5000/_app-metrics/test-counter
10

In this way you can easily expose your application's metrics to an external monitoring service. Moreover, since the AppMetricsMiddleware exposes a full RESTful API, you can create metrics from anywhere and also populate them with foreign application's data.

Usage

As usual, instantiate the middleware with the wrapped WSGI application; it looks for request paths starting with "/_app-metrics": if not found, the wrapped application is called. The following resources are defined:

/_app-metrics

• GET: return the list of the registered metrics

/_app-metrics/<name>

• GET: return the value of the given metric or 404.
• PUT: create a new metric with the given name. The body must be a JSON object with a mandatory attribute named "type" which must be one of the metrics types allowed, by the "metrics.METRIC_TYPES" dictionary, while the other attributes are passed to the new_<type> function as keyword arguments. Request's content-type must be "application/json".
• POST: add a new value to the metric. The body must be a JSON object with a mandatory attribute named "value": the notify method will be called with the given value. Other attributes are ignored. Request's content-type must be "application/json".

The root doesn't have to be "/_app-metrics", you can customize it by providing your own to the middleware constructor.

A standalone AppMetrics webapp can be started by using werkzeug's development server:

$ python -m werkzeug.serving appmetrics.wsgi.standalone_app
* Running on http://127.0.0.1:5000/

The standalone app mounts on the root (no _app-metrics prefix). DON'T use it for production purposes!!!

Reporting

AppMetrics provides another easy way to get your application's metrics: the reporter module. It allows to register any number of callbacks that will be called at scheduled times with the metrics, allowing you to "export" your application's metrics into your favourite storage system. The main entry point for the reporter feature is reporter.register:

reporter.register(callback, schedule, tag=None)

where:

• callback must be a callback function that will be called with a dictionary of {metric name: metric values}
• schedule must be an iterable object yielding a future timestamp (in time.time() format) at each iteration
• tag must be a tag to narrow the involved metrics to the ones with that tag, if None all the available metrics will be used.

When a callback is registered, a new thread will be started, waiting for the next scheduled call. Please notice that the callback will be executed in a thread. register returns an opaque id identifying the registration.

A callback registration can be removed by calling reporter.remove with the id returned by register.

reporter provides a simple scheduler object, fixed_interval_scheduler:

>>> sched = reporter.fixed_interval_scheduler(10)
>>> next(sched)
1397297405.672592
>>> next(sched)
1397297415.672592
>>> next(sched)
1397297425.672592

CSV reporter

A simple reporter callback is exposed by reporter.CSVReporter. As the name suggests, it will create csv reports with metric values, a file for each metric, a row for each call. See examples/csv_reporter.py

Testing

AppMetrics has an exhaustive, fully covering test suite, made up by both doctests and unit tests. To run the whole test suite (including the coverage test), just issue:

$ nosetests --with-coverage --cover-package=appmetrics --cover-erase

You will need to install a couple of packages in your python environment, the list is in the "requirements.txt" file.