Retry Pattern with Exponential Backoff

Also Known As

TL;DR

Explanation

Transient failures (network blips, rate limit 429s, database deadlocks) are often self-resolving. The Retry pattern retries the operation automatically, but naive immediate retries can overwhelm a struggling service. Exponential backoff doubles the wait on each attempt (1s, 2s, 4s, 8s...); jitter adds randomness to spread retries from many clients (prevents thundering herd). Best practices: cap maximum retries (3–5), cap maximum delay (30–60s), only retry idempotent operations (GET, PUT — not POST without idempotency keys), use a circuit breaker to stop retrying when the service is clearly down. Guzzle's retry middleware and symfony/http-client both support retry configuration. Log every retry with the delay and reason.

Diagram

flowchart TD
    REQ[Send Request] --> TRY{Success?}
    TRY -->|Yes| DONE[Done]
    TRY -->|No| CHECK{Retryable
error?}
    CHECK -->|No - 400 400 auth| FAIL[Fail immediately]
    CHECK -->|Yes - 500 429 timeout| WAIT[Wait with<br/>exponential backoff<br/>1s -> 2s -> 4s -> 8s]
    WAIT --> ATTEMPTS{Max attempts
reached?}
    ATTEMPTS -->|No| REQ
    ATTEMPTS -->|Yes| GIVEUP[Give up<br/>alert / dead letter]
style DONE fill:#238636,color:#fff
style FAIL fill:#f85149,color:#fff
style GIVEUP fill:#f85149,color:#fff
style WAIT fill:#d29922,color:#fff

Common Misconception

Why It Matters

Common Mistakes

• Retrying non-idempotent operations without checking — retrying a payment charge double-charges the customer.
• Constant retry interval without exponential backoff — hammers a recovering service at full rate.
• No jitter — all clients retry at the same time, creating a thundering herd.
• No maximum retry limit — infinite retries hold resources and can cause cascading failures.

Code Examples

// Retry without backoff or limit — thundering herd:
for ($i = 0; $i < 10; $i++) {
    try {
        return $this->http->post('/payments', $data);
    } catch (Exception $e) {
        sleep(1); // Same interval every time — all clients retry simultaneously
    }
}
// Fix: sleep(2 ** $i + random_int(0, 1000) / 1000) — exponential backoff with jitter

// Guzzle retry middleware
$handler = HandlerStack::create();
$handler->push(Middleware::retry(
    fn($retries, $req, $res) => $retries < 3 && ($res?->getStatusCode() === 429),
    fn($retries) => (int)(1000 * 2 ** $retries) // exponential ms delay
));

References

• ↗ https://aws.amazon.com/blogs/architecture/exponential-backoff-and-jitter/

Tags

Related Terms