fix: add local fallback for risk scores on Cloud Run

Author: Siriapps

.dockerignore

@@ -3,7 +3,10 @@ docs/
 catboost_info/
 cache/
 scripts/
-backend/model/output/
+backend/model/output/*
+!backend/model/output/intersection_scores.parquet
+!backend/model/output/shap_top3.pkl
+!backend/model/output/hourly_risk_factors.parquet
 backend/__pycache__/
 backend/.env
 **/__pycache__/

backend/main.py

@@ -442,8 +442,8 @@ def compute_route(payload: RouteRequest):
                     "time_band": safety.get("time_band"),
                     "segment_risks": safety.get("segment_risks", []),
                 }
-            except Exception:
-                pass
+            except Exception as score_err:
+                print(f"[route] score_coordinates failed for route: {score_err}", flush=True)
             result_routes.append({
                 "distance_meters": route.get("distanceMeters"),
                 "duration": route.get("duration"),

backend/model/output/intersection_scores.parquet

@@ -682,6 +682,12 @@ def main():
     tod_df.to_parquet(OUTPUT_DIR / "hourly_risk_factors.parquet", index=False)
     print(f"  Saved hourly_risk_factors.parquet ({len(tod_df)} rows).", flush=True)
 
+    # Save intersection_scores.parquet locally (used as fallback on Cloud Run)
+    scores_local = df[["node_id", "predicted_risk"]].copy()
+    scores_local["predicted_risk"] = scores_local["predicted_risk"].fillna(0)
+    scores_local.to_parquet(OUTPUT_DIR / "intersection_scores.parquet", index=False)
+    print(f"  Saved intersection_scores.parquet locally ({len(scores_local)} rows).", flush=True)
+
     if args.export_csv:
         df.to_csv(OUTPUT_DIR / "chicago_intersection_training_dataset.csv", index=False)
         print(f"  Saved dataset CSV.", flush=True)

backend/model/train_model.py

@@ -92,7 +92,20 @@ def get_risk_map() -> dict[str, float]:
         _tod_map = None
         _prepared_graph = None
 
-    # Try GCS first (local key file or Cloud Run default credentials)
+    # Try local file first (bundled in Docker image from training output)
+    local_path = OUTPUT_DIR / "intersection_scores.parquet"
+    if local_path.exists():
+        try:
+            import pandas as pd
+            scores = pd.read_parquet(local_path)
+            _risk_map = dict(zip(scores["node_id"].astype(str), scores["predicted_risk"].fillna(0).astype(float)))
+            _cache_loaded_at = time.time()
+            print(f"[risk_cache] loaded {len(_risk_map)} nodes from local file", flush=True)
+            return _risk_map
+        except Exception as e:
+            print(f"[risk_cache] local parquet load failed ({e}), trying GCS", flush=True)
+
+    # Try GCS (local key file or Cloud Run default credentials)
     try:
         import pandas as pd
         fs = _get_gcs_fs()