Add Vision image fetch failure observability

clip/main.py

@@ -1,6 +1,8 @@
 from __future__ import annotations
 
 import os
+import logging
+import time
 from typing import List, Optional
 
 import torch
@@ -9,7 +11,7 @@ from fastapi import FastAPI, HTTPException, UploadFile, File, Form
 from pydantic import BaseModel, Field
 import numpy as np
 
-from common.image_io import fetch_url_bytes, bytes_to_pil, ImageLoadError
+from common.image_io import fetch_url_bytes, bytes_to_pil, ImageLoadError, image_url_log_context
 
 MODEL_NAME = os.getenv("MODEL_NAME", "ViT-B-32")
 MODEL_PRETRAINED = os.getenv("MODEL_PRETRAINED", "openai")
@@ -25,6 +27,7 @@ TAGS: List[str] = [
 ]
 
 app = FastAPI(title="Skinbase CLIP Service", version="1.0.0")
+logger = logging.getLogger("clip")
 
 model, _, preprocess = open_clip.create_model_and_transforms(MODEL_NAME, pretrained=MODEL_PRETRAINED)
 tokenizer = open_clip.get_tokenizer(MODEL_NAME)
@@ -44,6 +47,21 @@ class EmbedRequest(BaseModel):
     pretrained: Optional[str] = None
 
 
+def _log_image_load_failure(error: ImageLoadError, url: str, elapsed_ms: float) -> None:
+    fields = {
+        "event": "clip_image_fetch_failed",
+        "stage": error.stage,
+        "error_type": error.category,
+        "exception_class": error.original_exception_class or error.__class__.__name__,
+        "elapsed_ms": round(elapsed_ms, 1),
+        **image_url_log_context(url),
+    }
+    for key in ("remote_status", "content_type"):
+        if key in error.metadata:
+            fields[key] = error.metadata[key]
+    logger.warning("%s", " ".join(f"{key}={value!r}" for key, value in fields.items()))
+
+
 @app.get("/health")
 def health():
     return {"status": "ok", "device": DEVICE, "model": MODEL_NAME, "pretrained": MODEL_PRETRAINED}
@@ -143,10 +161,12 @@ async def analyze_file(
 def embed(req: EmbedRequest):
     if not req.url:
         raise HTTPException(400, "url is required")
+    started_at = time.perf_counter()
     try:
         data = fetch_url_bytes(req.url)
         return _embed_image_bytes(data, backend=req.backend, model_name=req.model, pretrained=req.pretrained)
     except ImageLoadError as e:
+        _log_image_load_failure(e, req.url, (time.perf_counter() - started_at) * 1000)
         raise HTTPException(400, str(e))