| 1 | # Exponential Backoff with Jitter and Retry-After Headers |
| 2 | |
| 3 | confidence: high |
| 4 | discovered_by: Farnsworth, Bender (GitHub crawler phase) |
| 5 | date: 2026-05-19 |
| 6 | |
| 7 | ## Pattern |
| 8 | |
| 9 | Implement resilient HTTP retry logic that combines: |
| 10 | 1. Exponential backoff (2^attempt, capped at 60s) for deterministic delay |
| 11 | 2. Random jitter (0.3–1.7s) to prevent thundering herd |
| 12 | 3. Server-provided Retry-After header (HTTP 429, 503) takes precedence |
| 13 | 4. Secondary rate limit detection with enforced minimum backoff (8s + 0.0–5s jitter) |
| 14 | 5. Rate limit state tracking (X-RateLimit-Remaining, X-RateLimit-Reset) |
| 15 | |
| 16 | ## When to Use |
| 17 | |
| 18 | - External HTTP requests to rate-limited APIs (GitHub GraphQL, RSS feeds, third-party crawlers) |
| 19 | - Handling HTTP 429, 500, 502, 503, 504 responses |
| 20 | - Distributed systems where retry storms can amplify load (thundering herd) |
| 21 | - API quota exhaustion scenarios with server-provided retry guidance |
| 22 | |
| 23 | ## Implementation |
| 24 | |
| 25 | ```python |
| 26 | # Exponential backoff calculation |
| 27 | base_delay = min(2**attempt, 60) # Cap at 60 seconds |
| 28 | jitter = random.uniform(0.3, 1.7) |
| 29 | delay = base_delay + jitter |
| 30 | |
| 31 | # Honor Retry-After header (seconds) |
| 32 | if "Retry-After" in response_headers: |
| 33 | retry_after = float(response_headers["Retry-After"]) |
| 34 | delay = max(retry_after, 1.0) |
| 35 | |
| 36 | # Secondary rate limit: enforce minimum |
| 37 | if "secondary rate limit" in response_body.lower(): |
| 38 | delay = max(delay, 8.0 + random.uniform(0.0, 5.0)) |
| 39 | |
| 40 | # Cap total delay to prevent indefinite waits |
| 41 | delay = min(delay, max_delay_seconds) |
| 42 | |
| 43 | # Sleep and retry |
| 44 | time.sleep(delay) |
| 45 | ``` |
| 46 | |
| 47 | ## Examples |
| 48 | |
| 49 | From `scripts/crawl.py` (GitHub GraphQL crawler): |
| 50 | |
| 51 | ```python |
| 52 | def _sleep_before_retry( |
| 53 | self, |
| 54 | attempt: int, |
| 55 | headers: dict[str, str] | None, |
| 56 | body: str, |
| 57 | query: str, |
| 58 | retry_limit: int, |
| 59 | max_delay_seconds: float, |
| 60 | ) -> None: |
| 61 | reset_delay = self._reset_delay(headers) |
| 62 | retry_after = None |
| 63 | if headers and headers.get("Retry-After"): |
| 64 | try: |
| 65 | retry_after = max(float(headers["Retry-After"]), 1.0) |
| 66 | except ValueError: |
| 67 | retry_after = None |
| 68 | |
| 69 | base_delay = min(2**attempt, 60) |
| 70 | jitter = random.uniform(0.3, 1.7) |
| 71 | delay = retry_after or reset_delay or (base_delay + jitter) |
| 72 | if "secondary rate limit" in body.lower(): |
| 73 | delay = max(delay, 8.0 + random.uniform(0.0, 5.0)) |
| 74 | delay = min(delay, max_delay_seconds) |
| 75 | |
| 76 | log(f"Retrying {query} in {delay:.1f}s (attempt {attempt + 1}/{retry_limit}).") |
| 77 | time.sleep(delay) |
| 78 | ``` |
| 79 | |
| 80 | State tracking pattern: |
| 81 | ```python |
| 82 | self.rate_limit_reset = max(self.rate_limit_reset or 0, int(time.time() + retry_after)) |
| 83 | ``` |
| 84 | |
| 85 | Retryable status codes: |
| 86 | ```python |
| 87 | RETRYABLE_STATUSES = {403, 429, 500, 502, 503, 504} |
| 88 | ``` |