Skip to content
SolidwayLLM gateway

Reliability

Failure handling

Three mechanisms run together: choosing a target, moving on when it fails, and taking it out of rotation while it is unwell.

  1. Choosing a target

    Targets are scored on a recent window of latency and error rate. Weights are recalculated on a fixed interval and read from cache on the request path, so scoring does not add measurable work per request.

    • Weighted selection across keys and endpoints
    • Optional session affinity for stateful flows
    • Per-target concurrency and timeout limits
  2. Moving on

    Each attempt is classified before a decision is made. Transient failures retry against the same target with backoff; failures bound to a key move to a different key instead of waiting.

    • Ordered fallback chains with independent budgets
    • Retry with exponential backoff and jitter
    • Failure classification recorded per attempt
  3. Taking a target out

    A failure that describes the target rather than the request holds that target out of rotation. It is re-checked by a single probe once the wait expires, and returns through a ramp rather than at full weight.

    • Hold scoped to a key, or to one model of a key
    • Growing wait with a ceiling per failure class
    • Manual release from the dashboard or the API

How a failure is classified

The class decides whether a target is retried, rotated or held. Request-shaped errors never hold a target.

How a failure is classified
ClassRetryHold
Transient (network, 5xx)Same target, with backoffNo
Rate limitRotate keyYes, short
Credential rejectedRotate key, no backoffYes
Quota exhaustedRotate key, no backoffYes
Model not accessibleRotate key, no backoffYes, that model
Request rejected by providerNoNo

Next step

Talk to us about your architecture

Tell us how your traffic is shaped today. We will walk through where a forward path fits, what the failure policy should look like, and how the gateway would be deployed in your environment.

No pricing on this page. Every deployment is scoped individually.

  • A review of your current provider and key layout
  • A failure policy matched to your error classes
  • A deployment shape for your network