Skip to the content.

Context

Several models or providers can produce the same result at different and individually variable latencies, and the request sits on the user’s critical path where a slow tail hurts the experience.

Example

Zeeguu races calls where the latency tail hurts a waiting user and a second attempt is cheap.

Problem

How can worst-case latency stay low when any single provider can be intermittently slow?

Forces

Multiple LLM providers offer similar capabilities but with varying latency. When speed matters for user experience, relying on a single provider creates a bottleneck.

Solution

Dispatch the same request to multiple providers simultaneously and use the first response that arrives. To cap the redundant cost, race only the few providers that have historically been fastest rather than all of them.

This composes with live retrieval: when a user is unsure of a translation and asks for alternatives, the UI shows the already-raced cached results immediately, then fetches more on demand, streaming them in as they arrive.

Consequences

Known Uses

Notes

In itself this is not LLM-specific. It is the hedged requests pattern from distributed systems (Dean and Barroso, The Tail at Scale, CACM 2013), also known as request racing or tied requests. Two things give it an LLM flavour here.

  1. First, the hedge is across independent, competing vendors (Anthropic, DeepSeek, Google Translate) rather than replicas of a single service.
  2. Second, each redundant call costs real money per token, so the losing responses are not thrown away: they are kept and surfaced to future readers of the same text as alternative translations, turning the redundant work into a feature.

💬 Open an issue about this pattern