[streamdiffusion] Request timeout while fetching IP adapter style image URL — unhandled in param update path

[livepeer-tessa]

Summary

When updating pipeline parameters that include an ip_adapter_style_image_url, the worker logs an error: Error updating params: Request timeout while fetching image from URL. This error is surfaced as last_error in the pipeline status but appears to be swallowed/unhandled — the pipeline remains ONLINE but may silently use stale/no style image.

First observed: 2026-03-20 (stream str_9Kk5GrU9P842aFd4, gateway_request_id=d2256c54)
Source: Grafana/Loki {container="live-video-to-video_streamdiffusion-sdxl_*"}

Error Details

From pipeline status event:

{
  "last_error_time": 1774030166223,
  "last_error": "Error updating params: Request timeout while fetching image from URL"
}

The ip_adapter_style_image_url in use was:
https://storage.googleapis.com/lp-ai-assets/ipadapter_style_imgs/textures/vortex.jpeg

Behavior

• Pipeline continues running (remains ONLINE) despite the error
• The style image is not updated — behavior unclear: may use previous image or none
• No retry logic observed; user is not notified

Impact

• Silent degradation of output quality (IP adapter style not applied correctly)
• Users have no visibility into the failure unless checking raw status events
• Could contribute to confusing visual artifacts / unexpected outputs

Suggested Fix

1. Add retry with timeout when fetching ip_adapter_style_image_url (e.g., 3 retries with 5s timeout each).
2. Surface the error clearly — emit a user-facing error event, not just last_error in status.
3. Consider pre-fetching / caching style images at startup rather than on every param update.
4. Fallback behavior: if URL fetch fails, log clearly and skip IP adapter for this update rather than silently proceeding.

Environment

• Pipeline: streamdiffusion-sdxl
• Worker: fal.ai
• URL: GCS bucket (generally reliable; timeout may indicate transient network issue or large image)