Retrying LLM Requests Isn't Always Safe: Gateway Policies for Partial Streams and Side Effects

Rama_Surasani_ · reddit · 2026-09-18

A Reddit post lays out production policies for LLM gateways when retries are unsafe: bounded jittered retries only before any output, honoring Retry-After on 429s, explicit incomplete states for partial streams (never mixing two models' outputs), and reconciling completed side-effecting tool calls via idempotency keys before retrying. Every attempt should share a trace ID and log tokens, tool-call status, fallback route, latency and cost, with chaos tests injecting timeouts and 5xx errors.

Original post →

More from coding & agent

coding & agent channel →