import time, traceback

def every(delay, task):
  next_time = time.time() + delay
  while True:
    time.sleep(max(0, next_time - time.time()))
    try:
      task()
    except Exception:
      traceback.print_exc()
      # in production code you might want to have this instead of course:
      # logger.exception("Problem while executing repetitive task.")
    # skip tasks if we are behind schedule:
    next_time += (time.time() - next_time) // delay * delay + delay

def foo():
  print("foo", time.time())

every(5, foo)

如果你想在不阻塞剩余代码的情况下这样做，你可以使用这个让它在自己的线程中运行:

import threading
threading.Thread(target=lambda: every(5, foo)).start()

该解决方案结合了其他解决方案中很少结合的几个特性:

Exception handling: As far as possible on this level, exceptions are handled properly, i. e. get logged for debugging purposes without aborting our program. No chaining: The common chain-like implementation (for scheduling the next event) you find in many answers is brittle in the aspect that if anything goes wrong within the scheduling mechanism (threading.Timer or whatever), this will terminate the chain. No further executions will happen then, even if the reason of the problem is already fixed. A simple loop and waiting with a simple sleep() is much more robust in comparison. No drift: My solution keeps an exact track of the times it is supposed to run at. There is no drift depending on the execution time (as in many other solutions). Skipping: My solution will skip tasks if one execution took too much time (e. g. do X every five seconds, but X took 6 seconds). This is the standard cron behavior (and for a good reason). Many other solutions then simply execute the task several times in a row without any delay. For most cases (e. g. cleanup tasks) this is not wished. If it is wished, simply use next_time += delay instead.

我认为这取决于你想做什么，你的问题没有详细说明。

对我来说，我想在一个已经多线程的进程中做一个昂贵的操作。所以我让leader流程检查时间，只有她做昂贵的操作(检查点深度学习模型)。为了做到这一点，我增加了计数器，以确保5秒、10秒、15秒过去，每5秒保存一次(或使用math.floor的模块化算术):

def print_every_5_seconds_have_passed_exit_eventually():
    """
    https://stackoverflow.com/questions/3393612/run-certain-code-every-n-seconds
    https://stackoverflow.com/questions/474528/what-is-the-best-way-to-repeatedly-execute-a-function-every-x-seconds
    :return:
    """
    opts = argparse.Namespace(start=time.time())
    next_time_to_print = 0
    while True:
        current_time_passed = time.time() - opts.start
        if current_time_passed >= next_time_to_print:
            next_time_to_print += 5
            print(f'worked and {current_time_passed=}')
            print(f'{current_time_passed % 5=}')
            print(f'{math.floor(current_time_passed % 5) == 0}')

starting __main__ at __init__
worked and current_time_passed=0.0001709461212158203
current_time_passed % 5=0.0001709461212158203
True
worked and current_time_passed=5.0
current_time_passed % 5=0.0
True
worked and current_time_passed=10.0
current_time_passed % 5=0.0
True
worked and current_time_passed=15.0
current_time_passed % 5=0.0
True

对我来说，检查if语句是我所需要的。在我已经复杂的多处理器多gpu代码中拥有线程，调度器并不是我想要添加的复杂性，如果我可以避免它，似乎我可以。检查worker id很容易确保只有一个进程在做这件事。

注意，我使用True print语句来确保模块化的算术技巧有效，因为检查确切的时间显然是行不通的!但令我惊喜的是，地板竟然起了作用。

如果漂移不是一个问题

import threading, time

def print_every_n_seconds(n=2):
    while True:
        print(time.ctime())
        time.sleep(n)
    
thread = threading.Thread(target=print_every_n_seconds, daemon=True)
thread.start()

异步输出。

#Tue Oct 16 17:29:40 2018
#Tue Oct 16 17:29:42 2018
#Tue Oct 16 17:29:44 2018

如果正在运行的任务需要相当多的时间，那么间隔就变成2秒+任务时间，所以如果您需要精确的调度，那么这并不适合您。

注意daemon=True标志意味着这个线程不会阻止应用程序关闭。例如，在运行测试等待此头停止后，pytest将无限期挂起的问题。

这里是另一个不使用任何额外库的解决方案。

def delay_until(condition_fn, interval_in_sec, timeout_in_sec):
    """Delay using a boolean callable function.

    `condition_fn` is invoked every `interval_in_sec` until `timeout_in_sec`.
    It can break early if condition is met.

    Args:
        condition_fn     - a callable boolean function
        interval_in_sec  - wait time between calling `condition_fn`
        timeout_in_sec   - maximum time to run

    Returns: None
    """
    start = last_call = time.time()
    while time.time() - start < timeout_in_sec:
        if (time.time() - last_call) > interval_in_sec:
            if condition_fn() is True:
                break
            last_call = time.time()

如果您的程序还没有事件循环，请使用sched模块，它实现了一个通用的事件调度器。

import sched, time

def do_something(scheduler): 
    # schedule the next call first
    scheduler.enter(60, 1, do_something, (scheduler,))
    print("Doing stuff...")
    # then do your stuff

my_scheduler = sched.scheduler(time.time, time.sleep)
my_scheduler.enter(60, 1, do_something, (my_scheduler,))
my_scheduler.run()

如果您已经在使用事件循环库，如asyncio、trio、tkinter、PyQt5、gobject、kivy等，则只需使用现有事件循环库的方法来调度任务。

我用这个方法使每小时产生60个事件，其中大多数事件在整分钟后的相同秒数内发生:

import math
import time
import random

TICK = 60 # one minute tick size
TICK_TIMING = 59 # execute on 59th second of the tick
TICK_MINIMUM = 30 # minimum catch up tick size when lagging

def set_timing():

    now = time.time()
    elapsed = now - info['begin']
    minutes = math.floor(elapsed/TICK)
    tick_elapsed = now - info['completion_time']
    if (info['tick']+1) > minutes:
        wait = max(0,(TICK_TIMING-(time.time() % TICK)))
        print ('standard wait: %.2f' % wait)
        time.sleep(wait)
    elif tick_elapsed < TICK_MINIMUM:
        wait = TICK_MINIMUM-tick_elapsed
        print ('minimum wait: %.2f' % wait)
        time.sleep(wait)
    else:
        print ('skip set_timing(); no wait')
    drift = ((time.time() - info['begin']) - info['tick']*TICK -
        TICK_TIMING + info['begin']%TICK)
    print ('drift: %.6f' % drift)

info['tick'] = 0
info['begin'] = time.time()
info['completion_time'] = info['begin'] - TICK

while 1:

    set_timing()

    print('hello world')

    #random real world event
    time.sleep(random.random()*TICK_MINIMUM)

    info['tick'] += 1
    info['completion_time'] = time.time()

根据实际情况，你可能会得到长度的刻度:

60,60,62,58,60,60,120,30,30,60,60,60,60,60...etc.

但在60分钟结束时，你会有60个滴答;而且它们中的大多数都将出现在您喜欢的正确偏移时间。

在我的系统中，我得到了< 1/20秒的典型漂移，直到需要纠正。

该方法的优点是具有较好的时钟漂移分辨率;这可能会导致问题，如果你做的事情，比如每tick追加一个项目，你希望每小时追加60个项目。未能考虑漂移可能导致次要指标，如移动平均线，将数据考虑得过于深入过去，从而导致错误的输出。