tgs/dogslow

Dogslow is a Django watchdog middleware class that logs tracebacks of slow requests.

Dogslow -- Django Slow Request Watchdog

About this version

This is a fork of dogslow. The original package was by Erik van Zijst at Atlassian. Some changes were made by Misha Denisenko and Thomas Smith at NCBI. The canonical version of this fork is available at https://github.com/tgs/dogslow .

Overview

Dogslow is a Django watchdog middleware class that logs tracebacks of slow requests.

It started as an internal project inside Bitbucket to help trace operational problems.

Installation

Install dogslow:

$ pip install dogslow

Then add dogslow.WatchdogMiddleware to your list of middleware classes in your Django settings.py file:

MIDDLEWARE_CLASSES = (
    'dogslow.WatchdogMiddleware',
    ...
)

For best results, make it one of the first middlewares that is run.

Configuration

You can use the following configuration properties in your settings.py file to tune the watchdog:

# Watchdog is enabled by default, to temporarily disable, set to False:
DOGSLOW = True

# By default, Watchdog will create log files with the backtraces.
# You can also set the location of where it stores them:
DOGSLOW_LOG_TO_FILE = True
DOGSLOW_OUTPUT = '/tmp'

# Log requests taking longer than 25 seconds:
DOGSLOW_TIMER = 25

# When both specified, emails backtraces:
DOGSLOW_EMAIL_TO = 'errors@atlassian.com'
DOGSLOW_EMAIL_FROM = 'no-reply@atlassian.com'

# Also log to this logger (defaults to none):
DOGSLOW_LOGGER = 'syslog_logger'
DOGSLOW_LOG_LEVEL = 'WARNING'

# Tuple of url pattern names that should not be monitored:
# (defaults to none -- everything monitored)
# Note: this option is not compatible with Django < 1.3
DOGSLOW_IGNORE_URLS = ('some_view', 'other_view')

# Print (potentially huge!) local stack variables (off by default, use
# True for more detailed, but less manageable reports)
DOGSLOW_STACK_VARS = True

Usage

Every incoming HTTP request gets a 25 second timeout in the watchdog. If a request does not return within that time, the watchdog activates and takes a peek at the request thread's stack and writes the backtrace (including all local stack variables -- Django style) to a log file.

Each slow request is logged in a separate file that looks like this:

Undead request intercepted at: 16-05-2011 02:10:12 UTC

GET http://localhost:8000/?delay=2
Thread ID:  140539485042432
Process ID: 18010
Started:    16-05-2011 02:10:10 UTC

  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/management/commands/runserver.py", line 107, in inner_run
    run(self.addr, int(self.port), handler, ipv6=self.use_ipv6)
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/servers/basehttp.py", line 696, in run
    httpd.serve_forever()
  File "/usr/lib/python2.7/SocketServer.py", line 227, in serve_forever
    self._handle_request_noblock()
  File "/usr/lib/python2.7/SocketServer.py", line 284, in _handle_request_noblock
    self.process_request(request, client_address)
  File "/usr/lib/python2.7/SocketServer.py", line 310, in process_request
    self.finish_request(request, client_address)
  File "/usr/lib/python2.7/SocketServer.py", line 323, in finish_request
    self.RequestHandlerClass(request, client_address, self)
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/servers/basehttp.py", line 570, in __init__
    BaseHTTPRequestHandler.__init__(self, *args, **kwargs)
  File "/usr/lib/python2.7/SocketServer.py", line 639, in __init__
    self.handle()
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/servers/basehttp.py", line 615, in handle
    handler.run(self.server.get_app())
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/servers/basehttp.py", line 283, in run
    self.result = application(self.environ, self.start_response)
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/contrib/staticfiles/handlers.py", line 68, in __call__
    return self.application(environ, start_response)
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/handlers/wsgi.py", line 273, in __call__
    response = self.get_response(request)
  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/handlers/base.py", line 111, in get_response
    response = callback(request, *callback_args, **callback_kwargs)
  File "/home/erik/work/middleware/middleware/sleep/views.py", line 6, in sleep
    time.sleep(float(request.GET.get('delay', 1)))

Full backtrace with local variables:

  File "/home/erik/work/virtualenv/bit/lib/python2.7/site-packages/django/core/management/commands/runserver.py", line 107, in inner_run
    run(self.addr, int(self.port), handler, ipv6=self.use_ipv6)

  ...loads more...

The example above shows that the request thread was blocked in time.sleep() at the time dogslow took its snapshot.

Requests that return before dogslow's timeout expires do not get logged.

Note that dogslow only takes a peek at the thread's stack. It does not interrupt the request, or influence it in any other way. Using dogslow is therefore safe to use in production.

Sentry Integration

Dogslow natively integrates with Sentry. You can set it up by configuring Dogslow to use DOGSLOW_LOGGER and DOGSLOW_LOG_TO_SENTRY and by configuring Raven to collect Dogslow's reports.

DOGSLOW_LOGGER = 'dogslow' # can be anything, but must match `logger` below
DOGSLOW_LOG_TO_SENTRY = True

DOGSLOW_LOG_LEVEL = 'WARNING' # optional, defaults to 'WARNING'

# Add a new sentry handler to handle WARNINGs. It's not recommended to
# modify the existing sentry handler, as you'll probably start seeing
# other warnings unnecessarily sent to Sentry.
LOGGING = {
    ...
    'handlers': {
        ...
        'dogslow': {
            'level': 'WARNING',
            'class': 'raven.contrib.django.handlers.SentryHandler',
        }
        ...
    }
    'loggers': {
        ...
        'dogslow': {
            'level': 'WARNING',
            'handlers': ['dogslow'], # or whatever you named your handler
        }
        ...
    }
    ...
}

Caveats

Dogslow uses multithreading. It has a single background thread that handles the watchdog timeouts and takes the tracebacks, so that the original request threads are not interrupted. This has some consequences.

Multithreading and the GIL

In CPython, the GIL (Global Interpreter Lock) prevents multiple threads from executing Python code simultaneously. Only when a thread explicitly releases its lock on the GIL, can a second thread run.

Releasing the GIL is done automatically whenever a Python program makes blocking calls outside of the interpreter, for example when doing IO.

For dogslow this means that it can only reliably intercept requests that are slow because they are doing IO, calling sleep or busy waiting to acquire locks themselves.

In most cases this is fine. An important cause of slow Django requests is an expensive database query. Since this is IO, dogslow can intercept those fine. A scenario where CPython's GIL is problematic is when the request's thread hits an infinite loop in Python code (or legitimate Python that is extremely expensive and takes a long time to execute), never releasing the GIL. Even though dogslow's watchdog timer thread does become runnable, it cannot log the stack.

Co-routines and Greenlets

Dogslow is intended for use in a synchronous worker configuration. A webserver that uses dedicated threads (or single-threaded, dedicated worker processes) to serve requests. Django's built-in wsgi server does this, as does Gunicorn in its default sync-worker mode.

When running with a "co-routines framework" where multiple requests are served concurrently by one thread, backtraces might become nonsensical.