Does Python need yet another retry / poll library? It needs at least one that isn't coupled to decorators and functions. Decorators prevent the caller from customizing delay options, and organizing the code around functions hinders any custom handling of failures.

Waiter is built around iteration instead, because the foundation of retrying / polling is a slowly executing loop. The resulting interface is both easier to use and more flexible, decoupling the delay algorithms from the application logic.

Usage

creation

Supply a number of seconds to repeat endlessly, or any iterable of seconds.

from waiter import wait

wait(1)                 # 1, 1, 1, 1, ...
wait([1] * 3)           # 1, 1, 1
wait([0.5, 0.5, 60])    # circuit breaker

Iterable delays can express any waiting strategy, and constructors for common algorithms are also provided.

wait.count(1)           # incremental backoff 1, 2, 3, 4, 5, ...
wait(1) + 1             # alternate syntax 1, 2, 3, 4, 5, ...
wait.fibonacci(1)       # 1, 1, 2, 3, 5, ...
wait.polynomial(2)      # 0, 1, 4, 9, 16, ...

wait.exponential(2)     # exponential backoff 1, 2, 4, 8, ...
backoff = wait(1) * 2   # alternate syntax 1, 2, 4, 8, ...
backoff[:3]             # limit attempt count 1, 2, 4
backoff <= 5            # set maximum delay   1, 2, 4, 5, 5, 5, ...
backoff.random(-1, 1)   # add random jitter

iteration

Then simply use the wait object like any iterable, yielding the amount of elapsed time. Timeouts also supported of course.

from waiter import wait, suppress, first

for elapsed in wait(delays):            # first iteration is immediate
    with suppress(exception):           # then each subsequent iteration sleeps as necessary
        ...
        break

for _ in wait(delays, timeout):         # standard convention for ignoring a loop variable
    ...                                 # won't sleep past the timeout
    if ...:
        break

results = (... for _ in wait(delays))   # expressions are even easier
first(predicate, results[, default])    # filter for first true item
assert any(results)                     # perfect for tests too

functions

Yes, functional versions are provided, as well as being trivial to implement.

wait(...).throttle(iterable)                      # generate items from iterable
wait(...).repeat(func, *args, **kwargs)           # generate successive results
wait(...).retry(exception, func, *args, **kwargs) # return first success or re-raise exception
wait(...).poll(predicate, func, *args, **kwargs)  # return first success or raise StopIteration

The decorator variants are partial applications of the corresponding methods.

backoff = wait(0.1) * 2
@backoff.repeating
@backoff.retrying(exception)
@backoff.polling(predicate)

But in the real world:

• the function may not exist or be succinctly written as a lambda
• the predicate may not exist or be succinctly written as a lambda
• logging may be required
• there may be complex handling of different exceptions or results

So consider the block form, just as decorators don't render with blocks superfluous. Also note wait objects are re-iterable provided their original delays were.

async

Waiters also support async iteration. throttle optionally accepts an async iterable. repeat, retry, and poll optionally accept coroutine functions.

statistics

Waiter objects have a stats attribute for aggregating statistics about the calls made. The base implementation is an attempt counter. The interface of the stats object itself is considered provisional, but can be extended by overriding the Stats class attribute. The add method also allows customization of the iterable values; elapsed time is the default.

Installation

% pip install waiter

Tests

100% branch coverage.

% pytest [--cov]