https://blueprints.launchpad.net/zaqar/+spec/support-more-backoff-functions

Sometimes, because of network/internet connection issue or the load of the target services/applications, the notification sent from Zaqar to the subscriber may fail. That means there is no retries if endpoint has some problem that can’t receive the notification. For users or applications, it’s missing important message without notification. To solve this problem, Zaqar has supported notification delivery policy in pike release [1].

Now that Zaqar has supported only linear retry backoff function in Pike release,it increases the delay time linearly from minimum delay to maximum delay. We will support more retry backoff functions in Zaqar now.

Fortunately, AWS SNS service has supported the policy, that would be a good reference for Zaqar[2].

For clarification what we want to do, there are some questions to be answered:

- How many retry backoff functions Will Zaqar support?

You can choose from four retry backoff functions as follows:

- Linear.
- Arithmetic.
- Geometric.
- Exponential.

- What difference between these four retry backoff functions?

As we can see the picture in chapter Backoff Phase in [2], The screen shot shows how each retry backoff function affects the delay associated with messages during the backoff period. The vertical axis represents the delay in seconds associated with each of the 10 retries. The horizontal axis represents the retry number. The minimum delay is 5 seconds, and the maximum delay is 260 seconds.

- The formula of these four retry backoff functions

Generally, we assume that the minimum delay is MIN, the maximum delay is MAX, and the total retry number is NUM. From three element above, we can conclude the formula of these four retry backoff functions.

Linear.

LINEAR_INTERVAL = ( MAX - MIN ) / NUM time1 = MIN time2 = MIN + LINEAR_INTERVAL*1 time3 = MIN + LINEAR_INTERVAL*2 timen = MIN + LINEAR_INTERVAL*(n-1) timeN = MAX

The Linear retry backoff functions has already been implemented in Pike release, we only list here for reference and comparing.

Geometric

we assume the common ratio of a geometric series is K: time1 = MIN time2 = K * MIN time3 = K^2 * MIN timen = K^(n-1) * MIN timeN = K^(NUM-1) * MIN = MAX

so :

K^(NUM-1) = MAX/MIN lg[ K^(NUM-1) ] = lg( MAX/MIN ) (NUM-1) * lg(K) = lg( MAX/MIN ) lg(K) = lg( MAX/MIN ) / (NUM-1) K = 10^[ lg( MAX/MIN ) / (NUM-1) ]

Finally:

timen = {10^[ lg( MAX/MIN ) / (NUM-1) ]}^(n-1) * MIN

Exponential.

we assume the power exponent of a exponential series is k, coefficient is p:

the general function of power exponent of a exponential series is: y = p * k^x

time1 = MIN = p*k^1 timeN = MAX = p*k^NUM MAX/MIN = k^(NUM-1) lg(MAX/MIN) = (NUM-1)*lg(k) lg(k) = lg(MAX/MIN)/ (NUM-1) k = 10^[lg(MAX/MIN)/ (NUM-1)] p = MIN/{10^[lg(MAX/MIN)/ (NUM-1)]}

Finally:

timen = p*k^n timen = MIN/{10^[lg(MAX/MIN)/ (NUM-1)]} * {10^[lg(MAX/MIN)/ (NUM-1)]} ^n

Arithmetic.

we assume the common difference of a arithmetic series is d:

time1 = MIN time2 = time1 + d*1 time3 = time2 + d*2 time4 = time3 + d*3

timen = timen-1 + d*(n-1)

timeN = timeN-1 + d*(NUM-1) = MAX

So: timeN - time1 = NUM*(NUM-1)/2*d

d = 2*(MAX-MIN)/[NUM*(NUM-1)]

timen - time1 = n*(n-1)/2 * { 2*(MAX-MIN)/[NUM*(NUM-1)] }

Finally:

timen = n*(n-1)/2 * { 2*(MAX-MIN)/[NUM*(NUM-1)] } + MIN

- What changes will bring into Zaqar?

- First, we will define three functions according to formula of these retry backoff functions.
- Second, change the logic of notification, when fail is existing when sending notification to endpoint, then use the policy which has binded to this subscription to make retries. In Backoff Phase, we will select the retry backoff functions according to user choice to resend notification.

N/A

N/A

- Primary assignee:
- gecong (ge.cong@zte.com.cn)

Q-2

- Add three functions according to formula of these retry backoff functions.
- Change the Backoff Phase of notification process for applying the retry backoff functions.
- UTs for this feature.