> ## Documentation Index
> Fetch the complete documentation index at: https://infercrane.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Troubleshooting

> Diagnose configuration, provisioning, readiness, routing, rollout, and cleanup failures.

# Troubleshooting

Start with `infercrane doctor`, then `status`, `events`, `inspect`, and `explain`. Use `operation ID` when a lifecycle operation is retrying. `inspect` exposes provider request/resource IDs needed to reconcile external inventory.

* **Provisioning appears stuck:** read the durable progress message before changing anything. `provider is allocating capacity` and `provider capacity and secure worker bootstrap` mean the runtime is not reachable yet; `model artifact and runtime readiness` means the worker is reachable but its model is still loading. Check `operation ID`, `inspect`, and provider inventory. Do not submit another deployment with a different idempotency key.
* **Deployment is degraded:** run the general explanation and verify expected model readiness on each replica.
* **Not scaling:** use `explain scaling`; check thresholds, consecutive intervals, bounds, and cooldown.
* **Candidate rejected:** use `explain rollout`; compare persisted metrics and policy rather than retrying promotion blindly.
* **Slow first request:** use `explain cold-start`; unavailable provider substages are intentionally not inferred.
* **Delete interrupted:** reconnect and inspect the same durable operation. Confirm provider inventory is empty before considering cleanup complete.
* **AWS deployment is rejected before provisioning:** run `infercrane doctor --aws`; verify the
  complete `INFERCRANE_AWS_*` set, exact requested region/GPU, immutable image digest, role trust,
  private subnet reachability, and instance-profile access to the worker secret. Retrying a
  configuration error does not create capacity.
* **Kubernetes deployment is rejected before provisioning:** run `infercrane doctor --kubernetes`;
  verify the explicit context, namespace RoleBinding, immutable image, GPU label/resource, worker
  Secret reference, and optional KServe CRD. A server-side field conflict is intentional protection;
  inspect `managedFields` and resolve ownership instead of forcing the apply.
* **External fallback does not activate:** run `infercrane external inspect DEPLOYMENT`, then check
  primary health, privacy acknowledgement, the injected secret reference, external `/models`
  inventory, and remaining request/cost reservations. InferCrane fails closed and never replays a
  possibly transmitted request.
* **External fallback returns 429:** the in-memory lease or persisted hard budget is exhausted.
  Inspect the policy before deliberately increasing its ceilings; InferCrane does not fabricate or
  refund provider cost.

Never paste API keys, prompts, generated content, or unredacted provider responses into a public issue.