2026-11-09 · America/Los_Angeles · 社区动态 · #18

“Just retry it” hides three different failures in LLM calls, and only one of them is fixed by retrying

In a lot of LLM code I’ve seen described, error handling is a single retry(3) around the call. I think that lumps together three different kinds of failure that need three different handlers. 1. The call failed (transport / provider errors) Timeouts, rate limits (429), server errors (5xx). Retrying with exponential backoff and jitter is the standard answer, and respecting any Retry-After header the provider sends. This is the only bucket where “just retry” is clearly right. 2. The call succeeded but the output is malformed Invalid JSON, a missing field, a value outside the allowed set. Here,…

热度 43.8 / 100；排名与评分保留该期记录。