backoff

Function decoration for backoff and retry

This module provides function decorators which can be used to wrap a function such that it will be retried until some condition is met. It is meant to be of use when accessing unreliable resources with the potential for intermittent failures i.e. network resources and external APIs. Somewhat more generally, it may also be of use for dynamically polling resources for externally generated content.

Decorators support both regular functions for synchronous code and asyncio's coroutines for asynchronous code.

Examples

Since Kenneth Reitz's requests module has become a defacto standard for synchronous HTTP clients in Python, networking examples below are written using it, but it is in no way required by the backoff module.

@backoff.on_exception

The on_exception decorator is used to retry when a specified exception is raised. Here's an example using exponential backoff when any requests exception is raised:

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException)
def get_url(url):
    return requests.get(url)

The decorator will also accept a tuple of exceptions for cases where the same backoff behavior is desired for more than one exception type:

@backoff.on_exception(backoff.expo,
                      (requests.exceptions.Timeout,
                       requests.exceptions.ConnectionError))
def get_url(url):
    return requests.get(url)

Give Up Conditions

Optional keyword arguments can specify conditions under which to give up.

The keyword argument max_time specifies the maximum amount of total time in seconds that can elapse before giving up.

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      max_time=60)
def get_url(url):
    return requests.get(url)

Keyword argument max_tries specifies the maximum number of calls to make to the target function before giving up.

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      max_tries=8,
                      jitter=None)
def get_url(url):
    return requests.get(url)

In some cases the raised exception instance itself may need to be inspected in order to determine if it is a retryable condition. The giveup keyword arg can be used to specify a function which accepts the exception and returns a truthy value if the exception should not be retried:

def fatal_code(e):
    return 400 <= e.response.status_code < 500

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      max_time=300,
                      giveup=fatal_code)
def get_url(url):
    return requests.get(url)

By default, when a give up event occurs, the exception in question is reraised and so code calling an on_exception-decorated function may still need to do exception handling. This behavior can optionally be disabled using the raise_on_giveup keyword argument.

In the code below, requests.exceptions.RequestException will not be raised when giveup occurs. Note that the decorated function will return None in this case, regardless of the logic in the on_exception handler.

def fatal_code(e):
    return 400 <= e.response.status_code < 500

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      max_time=300,
                      raise_on_giveup=False,
                      giveup=fatal_code)
def get_url(url):
    return requests.get(url)

This is useful for non-mission critical code where you still wish to retry the code inside of backoff.on_exception but wish to proceed with execution even if all retries fail.

@backoff.on_predicate

The on_predicate decorator is used to retry when a particular condition is true of the return value of the target function. This may be useful when polling a resource for externally generated content.

Here's an example which uses a fibonacci sequence backoff when the return value of the target function is the empty list:

@backoff.on_predicate(backoff.fibo, lambda x: x == [], max_value=13)
def poll_for_messages(queue):
    return queue.get()

Extra keyword arguments are passed when initializing the wait generator, so the max_value param above is passed as a keyword arg when initializing the fibo generator.

When not specified, the predicate param defaults to the falsey test, so the above can more concisely be written:

@backoff.on_predicate(backoff.fibo, max_value=13)
def poll_for_message(queue):
    return queue.get()

More simply, a function which continues polling every second until it gets a non-falsey result could be defined like like this:

@backoff.on_predicate(backoff.constant, jitter=None, interval=1)
def poll_for_message(queue):
    return queue.get()

The jitter is disabled in order to keep the polling frequency fixed.

Jitter

A jitter algorithm can be supplied with the jitter keyword arg to either of the backoff decorators. This argument should be a function accepting the original unadulterated backoff value and returning it's jittered counterpart.

As of version 1.2, the default jitter function backoff.full_jitter implements the 'Full Jitter' algorithm as defined in the AWS Architecture Blog's Exponential Backoff And Jitter post. Note that with this algorithm, the time yielded by the wait generator is actually the maximum amount of time to wait.

Previous versions of backoff defaulted to adding some random number of milliseconds (up to 1s) to the raw sleep value. If desired, this behavior is now available as backoff.random_jitter.

Using multiple decorators

The backoff decorators may also be combined to specify different backoff behavior for different cases:

@backoff.on_predicate(backoff.fibo, max_value=13)
@backoff.on_exception(backoff.expo,
                      requests.exceptions.HTTPError,
                      max_time=60)
@backoff.on_exception(backoff.expo,
                      requests.exceptions.Timeout,
                      max_time=300)
def poll_for_message(queue):
    return queue.get()

Runtime Configuration

The decorator functions on_exception and on_predicate are generally evaluated at import time. This is fine when the keyword args are passed as constant values, but suppose we want to consult a dictionary with configuration options that only become available at runtime. The relevant values are not available at import time. Instead, decorator functions can be passed callables which are evaluated at runtime to obtain the value:

def lookup_max_time():
    # pretend we have a global reference to 'app' here
    # and that it has a dictionary-like 'config' property
    return app.config["BACKOFF_MAX_TIME"]

@backoff.on_exception(backoff.expo,
                      ValueError,
                      max_time=lookup_max_time)

Event handlers

Both backoff decorators optionally accept event handler functions using the keyword arguments on_success, on_backoff, and on_giveup. This may be useful in reporting statistics or performing other custom logging.

Handlers must be callables with a unary signature accepting a dict argument. This dict contains the details of the invocation. Valid keys include:

• target: reference to the function or method being invoked
• args: positional arguments to func
• kwargs: keyword arguments to func
• tries: number of invocation tries so far
• elapsed: elapsed time in seconds so far
• wait: seconds to wait (on_backoff handler only)
• value: value triggering backoff (on_predicate decorator only)

A handler which prints the details of the backoff event could be implemented like so:

def backoff_hdlr(details):
    print ("Backing off {wait:0.1f} seconds after {tries} tries "
           "calling function {target} with args {args} and kwargs "
           "{kwargs}".format(**details))

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      on_backoff=backoff_hdlr)
def get_url(url):
    return requests.get(url)

Multiple handlers per event type

In all cases, iterables of handler functions are also accepted, which are called in turn. For example, you might provide a simple list of handler functions as the value of the on_backoff keyword arg:

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      on_backoff=[backoff_hdlr1, backoff_hdlr2])
def get_url(url):
    return requests.get(url)

Getting exception info

In the case of the on_exception decorator, all on_backoff and on_giveup handlers are called from within the except block for the exception being handled. Therefore exception info is available to the handler functions via the python standard library, specifically sys.exc_info() or the traceback module.

Asynchronous code

Backoff supports asynchronous execution in Python 3.5 and above.

To use backoff in asynchronous code based on asyncio you simply need to apply backoff.on_exception or backoff.on_predicate to coroutines. You can also use coroutines for the on_success, on_backoff, and on_giveup event handlers, with the interface otherwise being identical.

The following examples use aiohttp asynchronous HTTP client/server library.

@backoff.on_exception(backoff.expo, aiohttp.ClientError, max_time=60)
async def get_url(url):
    async with aiohttp.ClientSession(raise_for_status=True) as session:
        async with session.get(url) as response:
            return await response.text()

Logging configuration

By default, backoff and retry attempts are logged to the 'backoff' logger. By default, this logger is configured with a NullHandler, so there will be nothing output unless you configure a handler. Programmatically, this might be accomplished with something as simple as:

logging.getLogger('backoff').addHandler(logging.StreamHandler())

The default logging level is INFO, which corresponds to logging anytime a retry event occurs. If you would instead like to log only when a giveup event occurs, set the logger level to ERROR.

logging.getLogger('backoff').setLevel(logging.ERROR)

It is also possible to specify an alternate logger with the logger keyword argument. If a string value is specified the logger will be looked up by name.

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      logger='my_logger')
# ...

It is also supported to specify a Logger (or LoggerAdapter) object directly.

my_logger = logging.getLogger('my_logger')
my_handler = logging.StreamHandler()
my_logger.addHandler(my_handler)
my_logger.setLevel(logging.ERROR)

@backoff.on_exception(backoff.expo,
                      requests.exceptions.RequestException,
                      logger=my_logger)
# ...

Default logging can be disabled all together by specifying logger=None. In this case, if desired alternative logging behavior could be defined by using custom event handlers.