fix(ml): race condition when loading models (#3207)

* sync model loading, disabled model ttl by default

* disable revalidation if model unloading disabled

* moved lock

machine-learning/app/config.py

@@ -13,7 +13,7 @@ class Settings(BaseSettings):
     facial_recognition_model: str = "buffalo_l"
     min_tag_score: float = 0.9
     eager_startup: bool = True
-    model_ttl: int = 300
+    model_ttl: int = 0
     host: str = "0.0.0.0"
     port: int = 3003
     workers: int = 1

machine-learning/app/main.py

@@ -25,7 +25,7 @@ app = FastAPI()
 
 
 def init_state() -> None:
-    app.state.model_cache = ModelCache(ttl=settings.model_ttl, revalidate=True)
+    app.state.model_cache = ModelCache(ttl=settings.model_ttl, revalidate=settings.model_ttl > 0)
 
 
 async def load_models() -> None:

machine-learning/app/models/cache.py

@@ -1,4 +1,3 @@
-import asyncio
 from typing import Any
 
 from aiocache.backends.memory import SimpleMemoryCache
@@ -48,13 +47,10 @@ class ModelCache:
         """
 
         key = self.cache.build_key(model_name, model_type.value)
+        async with OptimisticLock(self.cache, key) as lock:
             model = await self.cache.get(key)
             if model is None:
-            async with OptimisticLock(self.cache, key) as lock:
-                model = await asyncio.get_running_loop().run_in_executor(
-                    None,
-                    lambda: InferenceModel.from_model_type(model_type, model_name, **model_kwargs),
-                )
+                model = InferenceModel.from_model_type(model_type, model_name, **model_kwargs)
                 await lock.cas(model, ttl=self.ttl)
         return model