scales - Metrics for Python
Tracks server state and statistics, allowing you to see what your server is doing. It can also send metrics to Graphite for graphing or to a file for crash forensics.
scales is inspired by the fantastic metrics library, though it is by no means a port.
This is a brand new release - issue reports and pull requests are very much appreciated!
Installation
You can get a release from PyPI:
pip install scales
Or you can get it from GitHub:
git clone https://github.com/Cue/scales
cd scales
python setup.py install
The HTTP statistics viewer in scales requires one of the following web frameworks:
If you aren't sure, go with Flask; it's compatible with most every other event
loop. You can get it with pip install flask.
Scales is tested with Python 2.7 and 3.3. For some reason it does not work with PyPy; pull requests for this are welcome, if you can figure out what's up.
How to use it
Getting started and adding stats only takes a few lines of code:
from greplin import scales
STATS = scales.collection('/web',
scales.IntStat('errors'),
scales.IntStat('success'))
# In a request handler
STATS.success += 1This code will collect two integer stats, which is nice, but what you really want to do is look at those stats, to get insight into what your server is doing. There are two main ways of doing this: the HTTP server and Graphite logging.
The HTTP server is the simplest way to get stats out of a running server. The easiest way, if you have Flask installed, is to do this:
import greplin.scales.flaskhandler as statserver
statserver.serveInBackground(8765, serverName='something-server-42')This will spawn a background thread that will listen on port 8765, and serve up a very convenient view of all your stats. To see it, go to
You can also get the stats in JSON by appending ?format=json to the
URL. ?format=prettyjson is the same thing, but pretty-printed.
The HTTP server is good for doing spot checks on the internals of running servers, but what about continuous monitoring? How do you generate graphs of stats over time? This is where Graphite comes in. Graphite is a server for collecting stats and graphing them, and scales has easy support for using it. Again, this is handled in a background thread:
graphitePeriodicPusher = graphite.GraphitePeriodicPusher('graphite-collector-hostname', 2003, 'my.server.prefix.')
graphitePeriodicPusher.allow("*") # Logs everything to graphite
graphitePeriodicPusher.start()That's it! Numeric stats will now be pushed to Graphite every minute.
Note that, by default, if you don't use allow, nothing is logged to graphite.
You can also exclude stats from graphite logging with the forbid(prefix) method
of the GraphitePeriodicPusher class.
Timing sections of code
To better understand the performance of certain critical sections of your code, scales lets you collect timing information:
from greplin import scales
STATS = scales.collection('/web',
scales.IntStat('errors'),
scales.IntStat('success'),
scales.PmfStat('latency'))
# In a request handler
with STATS.latency.time():
do_something_expensive()This will collect statistics on the running times of that section of code: mean time, median, standard deviation, and several percentiles to help you locate outlier times. This happens in pretty small constant memory, so don't worry about the cost; time anything you like.
You can gather this same kind of sample statistics about any quantity. Just make
a PmfStat and assign new values to it:
for person in people:
person.perturb(42)
STATS.wistfulness = person.getFeelings('wistfulness')Metering Rates
Scales can track 1/5/15 minute averages with MeterStat:
from greplin.scales.meter import MeterStat
STATS = scales.collection('/web', MeterStat('hits'))
def handleRequest(..):
STATS.hits.mark() # or .mark(NUMBER), or STATS.hits = NUMBERClass Stats
While global stats are easy to use, sometimes making stats class-based makes
more sense. This is supported; just make sure to give each instance of the class
a unique identifier with scales.init.
class Handler(object):
requests = scales.IntStat('requests')
latency = scales.PmfStat('latency')
byPath = scales.IntDictStat('byPath')
def __init__(self):
scales.init(self, '/handler')
def handleRequest(self, request):
with self.latency.time():
doSomething()
self.requests += 1
self.byPath[request.path] += 1Gauges
Simple lambdas can be used to generate stat values.
STATS = scales.collection(scales.Stat('currentTime', lambda: time.time())Of course this works with arbitrary function objects, so the example above could also be written:
STATS = scales.collection(scales.Stat('currentTime', time.time)Hierarchical Stats + Aggregation
Stats can inherit their path from the object that creates them, and (non-gauge) stats can be aggregated up to ancestors.
class Processor(object):
"""Example processing management object."""
threadStates = scales.HistogramAggregationStat('state')
finished = scales.SumAggregationStat('finished')
def __init__(self):
scales.init(self, '/processor')
self.threads = 0
def createThread(self):
threadId = self.threads
self.threads += 1
SomeThread(threadId).start()
class SomeThread(object):
"""Stub of a processing thread object."""
state = scales.Stat('state')
finished = scales.IntStat('finished')
def __init__(self, threadId):
scales.initChild(self, 'thread-%d' % threadId)
def processingLoop(self):
while True:
self.state = 'waitingForTask'
getTask()
self.state = 'performingTask'
doTask()
self.finished += 1This will result in a stat at the path /processor/finished which counts the
total of the finished stats in each SomeThread object, as well as per-object
stats with paths like /processor/thread-0/finished. There will also be stats
like /processor/state/waitingForTask which aggregates the number of threads in
the waitingForTask state.
Authors
License
Copyright 2011 The scales Authors.
Published under The Apache License, see LICENSE