With the popularity of microservices, distributed systems have become the norm. Although distributed systems can certainly help ensuring scalability and resilience, they come with their fair share of challenges. One particularly I want to talk about today is the retry storm.
What is a retry storm?
A retry storm occurs when a large number of clients simultaneously attempt to retry failed requests, overwhelming the system and exacerbating the initial issue. In this blog post, we'll explore the causes of retry storms and provide practical strategies to avoid them.
Let’s image we have a range of services calling each other:
If each service has a retry policy installed which retries 3 times(resulting in a total of 3+1 requests), this would result in 64 times(!) the normal traffic. If retry policies are executed without context in cause traffic to grow exponentially.
How can we avoid it?
Some measures that can help avoid a retry storm are:
- Jittered Exponential Backoff: Introduce randomness into the backoff strategy to prevent synchronized retries. Jittered exponential backoff involves adding a random factor to the delay, reducing the likelihood of multiple clients retrying simultaneously. This helps distribute the load more evenly across the system.
-
Circuit Breaker Patterns: Implement circuit breaker patterns to proactively identify and handle failed requests. Once a predefined threshold of failures is reached, the circuit breaker temporarily opens, preventing further requests and allowing the system to recover. This helps avoid the accumulation of retry requests during times of instability.
-
Retry Budgets: Set limits on the number of retries within a specific time frame. By defining a retry budget, you prevent clients from overwhelming the system with a continuous stream of retry attempts. This approach encourages a more measured and controlled response to failures.
-
Exponential Backoff with Jittered Reset: To address long-term system instability, incorporate a jittered reset mechanism. After a successful request, reset the exponential backoff to its initial state. This prevents prolonged periods of conservative retry delays, allowing the system to adapt to changing conditions.
If you are using .NET, Polly can help you implement jittered backoff:
More information
Retry Storm antipattern - Performance antipatterns for cloud apps | Microsoft Learn