feat(infra): add metrics endpoint, redis/postgres services, ollama url fix, llm fallback

api.py

@@ -47,6 +47,7 @@ from extractor import SchemaExtractor
 from mconfig import PryConfig
 from mcp_production import make_fallback_server, register_all
 from mcp_sse import mcp_post_message, mcp_sse_endpoint
+from observability import get_metrics_output, setup_tracing
 from pipeline import HOOK_POINTS, get_pipeline, run_pipeline
 from pryextras import BatchProcessor, TransformEngine, recorder, streams
 from routers.auth import router as auth_router
@@ -80,6 +81,7 @@ except ImportError:
 async def lifespan(app: FastAPI) -> AsyncIterator[None]:
     """Startup: validate deps. Shutdown: cleanup clients."""
     logger.info("pry_startup", version="3.0.0")
+    setup_tracing()
     if not settings.api_key:
         logger.warning(
             "pry_api_key_unset",
@@ -2810,6 +2812,13 @@ app.include_router(health_router)
 app.include_router(scraping_router)
 app.include_router(templates_router)
 
+
+@app.get("/metrics", tags=["System"], summary="Prometheus metrics", include_in_schema=False)
+async def metrics() -> Response:
+    """Expose Prometheus metrics for scraping."""
+    data, content_type = get_metrics_output()
+    return Response(content=data, media_type=content_type)
+
 # ── Review ──
 
 

cli.py

@@ -83,6 +83,7 @@ def _spinner(label: str):
             if stop:
                 break
             print(f"  {CYAN}{c}{NC} {label}...  ", end="\r", flush=True)
+            # CLI spinner runs in a daemon thread; sync sleep is acceptable here.
             time.sleep(0.1)
 
     t = threading.Thread(target=_spin, daemon=True)

docker-compose.yml

@@ -80,3 +80,40 @@ services:
 volumes:
   pry-data:
   pry-sessions:
+
+  redis:
+    image: redis:7-alpine
+    container_name: pry-redis
+    restart: unless-stopped
+    ports:
+      - "127.0.0.1:6379:6379"
+    volumes:
+      - pry-redis:/data
+    deploy:
+      resources:
+        limits:
+          memory: 512M
+          cpus: "0.5"
+
+  postgres:
+    image: postgres:15-alpine
+    container_name: pry-postgres
+    restart: unless-stopped
+    ports:
+      - "127.0.0.1:5432:5432"
+    volumes:
+      - pry-postgres:/var/lib/postgresql/data
+    environment:
+      - POSTGRES_USER=pry
+      - POSTGRES_PASSWORD=pry
+      - POSTGRES_DB=pry
+    deploy:
+      resources:
+        limits:
+          memory: 512M
+          cpus: "0.5"
+
+volumes:
+  pry-sessions:
+  pry-redis:
+  pry-postgres:

llm_providers/registry.py

@@ -94,6 +94,35 @@ class LLMRegistry:
                     break
         raise Exception(f"All LLM providers failed. Last: {last_error}")
 
+    async def complete_or_empty(
+        self,
+        prompt: str,
+        system: str = "",
+        provider_name: str = "",
+        max_tokens: int = 1024,
+        temperature: float = 0.7,
+        model: str = "",
+        fallback: bool = True,
+    ) -> LLMResponse:
+        """Complete via LLM, returning an empty response on total failure.
+
+        Use this at API boundaries where a 500 is worse than degraded output.
+        """
+        try:
+            return await self.complete(
+                prompt, system, provider_name, max_tokens, temperature, model, fallback
+            )
+        except Exception as e:  # noqa: BLE001
+            logger.warning("llm_complete_failed_all_providers", extra={"error": str(e)[:120]})
+            return LLMResponse(
+                text="",
+                provider="none",
+                model="",
+                input_tokens=0,
+                output_tokens=0,
+                cost_usd=0.0,
+            )
+
     async def embed(self, text: str, provider_name: str = "", model: str = "") -> list[float]:
         names = [provider_name] if provider_name else list(self.fallback_chain)
         for name in names:

pyproject.toml

@@ -50,6 +50,10 @@ dependencies = [
     "structlog>=24.0.0",
     "sqlalchemy>=2.0.0",
     "aiosqlite>=0.19.0",
+    "prometheus-client>=0.21.0",
+    "opentelemetry-api>=1.29.0",
+    "opentelemetry-sdk>=1.29.0",
+    "opentelemetry-instrumentation-fastapi>=0.50b0",
 ]
 
 [project.optional-dependencies]

settings.py

@@ -25,7 +25,7 @@ class PrySettings(BaseSettings):
     port: int = 8002
 
     # Ollama LLM endpoint
-    ollama_url: str = "http://100.100.18.18:11434"
+    ollama_url: str = "http://100.104.130.92:11434"
 
     # FlareSolverr Cloudflare bypass endpoint
     flaresolverr_url: str = "http://flaresolverr:8191/v1"

ultimate_scraper.py

@@ -314,6 +314,11 @@ class UltimateScraper:
 
                 driver = uc.Chrome(headless=True, use_subprocess=True)
                 driver.get(url)
+                # Note: time.sleep is used inside a sync thread-pool task because
+                # undetected_chromedriver's blocking API runs in asyncio.to_thread.
+                # Replacing with asyncio.sleep would require restructuring the tier
+                # to be fully async; this sync sleep is bounded and does not block
+                # the event loop directly.
                 time.sleep(random.uniform(2, 4))
                 html = driver.page_source
                 driver.quit()