Changed The block ingestor no longer exits when getbestblockhash fails. A single failed call (for example HTTP 429 "Too many connections" from a loaded zebrad or zakura backend) used to call Log.Fatal, dropping every in-flight client stream; replicas sharing one backend then exited and reconnected together, adding to the overload. It now logs a warning and retries with a capped exponential backoff (1s to 30s), raises the log to Error after 5 minutes of continuous failures, and logs once when the RPC recovers. Malformed replies stay fatal.