Skip to main content

Python library for reporting metrics to Atlas.

Project description

Build Status


Python port of the Spectator library for Java.

See the Spectator documentation for an overview of core concepts and details on usage.

Supports Python >= 3.5, which is the oldest system Python 3 available on our commonly used OSes.

Note that there is a risk of deadlock if you are running Python 3.6 or lower and using os.fork() or using a library that will fork the process (see this section for workarounds), so we recommend using Python >= 3.7

Local Development

  • Install pyenv, possibly with Homebrew.
  • Install Python versions: 3.5, 3.6, 3.7, and 3.8. Enable all versions globally.
  • Make changes and add tests.
  • tox



The netflix-spectator-py package alone is not sufficient to report data to an Atlas backend - a configuration package must also be installed.

At Netflix, the internal configuration package is named netflix-spectator-pyconf and it declares a dependency on this client package.

pip3 install netflix-spectator-pyconf


Standard Usage

At Netflix, your initialization script should load the environment, to ensure that the standard variables are available to the Python application.

source /etc/nflx/environment

Importing the GlobalRegistry configures common tags based on environment variables, sets the Atlas Aggregator URL, and starts a background thread which reports metrics data every five seconds.

from spectator import GlobalRegistry

Once the GlobalRegistry is imported, it is used to create and manage Meters.

Concurrent Usage Under Older Python

:warning: Use Python 3.7+ if possible: But if you can't, here's a workaround to prevent deadlocks

There is a known issue in Python where forking a process after a thread is started can lead to deadlocks. This is commonly seen when using the multiprocessing module with default settings. The root cause of the deadlocks is that fork() copies everything in memory, including globals that have been set in imported modules, but it does not copy threads - any threads started in the parent process will not exist in the child process. The possibility of deadlocks occurs when global state sets a lock and it then depends upon a thread to remove the lock.

Under the standard usage model for spectator-py, it starts a background publishing thread when the module is imported, which is responsible for manipulating a lock. Below, there are a couple of options described for working around this issue.


If you are using spectator-py while running under Gunicorn, then do not use the --preload flag, which loads application code before worker processes are forked. Preloading triggers the conditions that allow deadlocks to occur in the background publish thread.

At Netflix, you can set the following flag in /etc/default/ezconfig to achieve this configuration for Gunicorn:


Task Worker Forking

:warning: Use Python 3.7+ if possible: But if you can't, here's a workaround to prevent deadlocks

For other pre-fork worker processing frameworks, such as huey, you need to be careful about how and when you start the GlobalRegistry to avoid deadlocks in the background publish thread. You should set the SPECTATOR_PY_DISABLE_AUTO_START_GLOBAL environment variable to disable automatic startup of the Registry, so that you can plan to start it manually after all of the workers have been forked.

You can set the variable as a part of your initialization script:


You can set the variable in Python code, as long as it is at the top of the module where you plan to use spectator-py:

import os


After your workers have started, you can then start the GlobalRegistry as follows:

from spectator import GlobalRegistry


It is often best to have the import and start() within an initialization function for the workers, to help ensure that it is not started when the module is loaded.

Generic Multiprocessing

:warning: Use Python 3.7+ if possible: But if you can't, here's a workaround to prevent deadlocks

In Python 3, you can configure the start method to spawn for multiprocessing. This will cause the module to do a fork() followed by an execve() to start a brand new Python process.

To configure this option globally:

from multiprocessing import set_start_method

To configure this option within a context:

from multiprocessing import get_context

def your_func():
    with get_context("spawn").Pool() as pool:


This package provides three loggers:

  • spectator.init
  • spectator.HttpClient
  • spectator.Registry

When troubleshooting metrics collection and reporting, you should set Registry logging to the DEBUG level. For example:

import logging

# record the human-readable time, name of the logger, logging level, thread id and message
    format='%(asctime)s - %(name)s - %(levelname)s - %(thread)d - %(message)s'

# silence the HttpClient logger output, to minimize confusion while reading logs

# set the Registry logger to INFO or ERROR when done troubleshooting

Detecting Deadlocks

If you need to detect whether or not your application is affected by deadlocks, then you can use sys._current_frames to collect stack frames periodically and check them. A common pattern is to run this on a background thread every 10 seconds.

Working with IDs

The IDs used for looking up a meter in the GlobalRegistry consist of a name and a set of tags. IDs will be consumed by users many times after the data has been reported, so they should be chosen thoughtfully, while considering how they will be used. See the naming conventions page for general guidelines.

IDs are immutable, so they can be freely passed around and used in a concurrent context. Tags can be added to an ID when it is created, to track the dimensionality of the metric. All tag keys and values must be strings. For example, if you want to keep track of the number of successful requests, you must cast integers to strings.

requests_id = GlobalRegistry.counter('server.numRequests', {'statusCode': str(200)})

Meter Types


A Counter is used to measure the rate at which an event is occurring. Considering an API endpoint, a Counter could be used to measure the rate at which it is being accessed.

Counters are reported to the backend as a rate-per-second. In Atlas, the :per-step operator can be used to convert them back into a value-per-step on a graph.

Call increment() when an event occurs:


You can also pass a value to increment(). This is useful when a collection of events happens together:


Distribution Summaries

A Distribution Summary is used to track the distribution of events. It is similar to a Timer, but more general, in that the size does not have to be a period of time. For example, a Distribution Summary could be used to measure the payload sizes of requests hitting a server.

Always use base units when recording data, to ensure that the tick labels presented on Atlas graphs are readable. If you are measuring payload size, then use bytes, not kilobytes (or some other unit). This means that a 4K tick label will represent 4 kilobytes, rather than 4 kilo-kilobytes.

Call record() with a value:



A gauge is a value that is sampled at some point in time. Typical examples for gauges would be the size of a queue or number of threads in a running state. Since gauges are not updated inline when a state change occurs, there is no information about what might have occurred between samples.

Consider monitoring the behavior of a queue of tasks. If the data is being collected once a minute, then a gauge for the size will show the size when it was sampled. The size may have been much higher or lower at some point during interval, but that is not known.

Call set() with a value:


Gauges are designed to report the last set value for 15 minutes. This done so that updates to the values do not need to be collected on a tight 1-minute schedule to ensure that Atlas shows unbroken lines in graphs.

If you wish to no longer report a Gauge value, then set it to float('nan'). This is a separate and distinct value from 'nan' or 'NaN', which are strings.


A Timer is used to measure how long (in seconds) some event is taking.

Call record() with a value:


Timers will keep track of the following statistics as they are used:

  • count
  • totalTime
  • totalOfSquares
  • max

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for netflix-spectator-py, version 0.1.16
Filename, size File type Python version Upload date Hashes
Filename, size netflix_spectator_py-0.1.16-py2.py3-none-any.whl (21.6 kB) File type Wheel Python version py2.py3 Upload date Hashes View
Filename, size netflix-spectator-py-0.1.16.tar.gz (19.2 kB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring DigiCert DigiCert EV certificate Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page