In the current Olares release, several Large Language Model (LLM) applications (such as those based on vLLM, Qwen, or Gemma) require downloading multi-gigabyte (GB) model weight files (e.g., GGUF format) from remote repositories like Hugging Face during their first launch or version updates.
For users on lower-bandwidth connections (e.g., 10–20 Mbps), this download process takes considerable time. As a result, the application container often exceeds the default Startup Probe timeout limits, causing the Kubernetes kubelet to flag the container as unhealthy (HTTP probe failed with statuscode: 503) and subsequently terminate (Killing) it.
Because the download sequence in some applications resets from scratch upon restart, the deployment gets trapped in an infinite loop of “downloading → timing out → being killed → restarting,” leaving the application permanently broken.