At PyCon I learned

At PyCon I learned

PyCon 2016 was held in Portland, Oregon at the end of May. I was one of the lucky 3000 or so Python developers able to attend. During the three day conference I listened to a total of 22 talks. I wanted to share something I learned there and picked one of my favorite talks, Daniel Riti's talk on gracefully degrading service failures.

Remote Calls and Service Failure

The failure of networks and services our systems depend on is inevitable. How do we deal with these situations? Daniel focused on the following three techniques:

I'm going to briefly introduce each one of them.


"Your code can't just wait forever for a response that might never come, sooner or later, it needs to give up. Hope is not a design method." -Michael T. Nygard, Release It!

Let's assume we have a simple web app using Python's requests library to get the current time from a remote time service. Our function could look something like this:

    def get_time():
            response = request.get('http://localhost:3001/time', timeout=3.0)
        except requests.exceptions.Timeout:
            return 'Time not available'
        return response.json().get('datetime')

If the time service is in unhealthy state, timeout forces an error in a given time. In this case, we will give up after a three second wait. Timeouts are extremely easy to use, since all you need to do is add an optional timeout parameter to your request; however, timeout is still applying load to the unhealthy service. Also, setting a timeout too short may mean we never get the response from otherwise healthy service, or in the opposite case keep user waiting without any kind of feedback for too long.


Retry is pretty self explanatory: "If at first you don't succeed, attempt the operation again". Since one request can have a multiplicative effect on a system, it's important to evaluate if the benefit of obtaining a response from a server outweighs potentially increasing load on the service. If you decide to proceed with retries, then always limit the number of retries. It's also a good practice to introduce a delay between retry attempts. There's a wonderful library called retrying that let's you set not only the number of retries, but a whole bunch of other parameters to the request. It's very simple to use:

    from retrying import retry

    @retry(stop_max_attempt_number=3, wait_exponential_multiplier=1000,
    def get_time():
        raise Exception

In the above example we set the total number of attemps to 3, introduce an exponentially growing delay between attempts and add a random jitter value from 0 to 500ms to the wait so that it further spreads out the requests and reduces the simultaneous load on the service.

Let's return to the multiplicative effect for a moment. What this actually means is that if we have a system that has three layers above the database; backend, frontend and javascript and each of these layers are using retries configured to make 4 attempts, then in the worst scenario we end up making 64 attempts from a single user interaction (4 tries ^ 3 levels).

Circuit Breaker Pattern

Circuit breaker pattern differs from the previous two methods in that it actually prevents operations when a depencency is unhealthy. It has three states: closed, half-open, and open:

Circuit Breaker Pattern

Again, there's a great library, called pybreaker that allows you to set parameters like fail_max (=maximum number of failures allowed) and reset_timeout (=how long we want to wait before letting the exploratory request go through). A simple example:

    time_breaker = pybreaker.CircuitBreaker(fail_max=2, reset_timeout=30)
    def get_time():
        raise pybreaker.CircuitBreakerError

The table below explains the pattern further:

Circuit Breaker Pattern

The first two requests go through ok, since the response time is under the timeout value of 3 seconds. The state of the circuit breaker pattern is closed. The third request takes 3s and fails due to the timeout; but, as we let it fail twice before triggering the circuit breaker pattern, the state of it is still closed. The next failing request reaches the max number of failures allowed and sets the circuit breaker pattern to open state. During the next 30 seconds all incoming requests are prevented.

Circuit Breaker Pattern

At the 30s mark, the state of the circuit breaker pattern is set to half-open and if the problem was fixed, like in our example, we allow the next request to go through for exploratory purposes. If it's successful, the state gets set to closed and requests are allowed to go through again. If the exploratory request failed, then another 30s wait would trigger.

The circuit breaker pattern used together with timeouts offers a smooth user experience since users don't have to wait until the timeout triggers. This tool also reduces load on unhealthy services and prevents unhealthy service affecting the whole system.

The advantages of this approach are obvious; however, it's important to pay attention to the parameters. A good approach is to monitor traffic and set parameters based on average response times.

Learn More

Thanks to Daniel for an informative talk. If you want to hear more, please check out his talk on youtube: Remote Calls != Local Calls: Graceful Degradation when Services Fail

All PyCon talks were recorded, so their PyCon 2016 channel is an exellent resource for learning about all Python related topics, highly recommended!