Fix: ensure data migration is handled by single-process

- Wrap migration logic with get_data_init_lock() to ensure single-process execution
- Prevent race conditions when multiple processes start simultaneously

lightrag/api/lightrag_server.py

@@ -151,9 +151,11 @@ def create_app(args):
         try:
             # Initialize database connections
             await rag.initialize_storages()
+            await initialize_pipeline_status()
+
+            # Data migration regardless of storage implementation
             await rag.check_and_migrate_data()
 
-            await initialize_pipeline_status()
             pipeline_status = await get_namespace_data("pipeline_status")
 
             should_start_autoscan = False

lightrag/lightrag.py

@@ -49,6 +49,7 @@ from lightrag.kg.shared_storage import (
     get_namespace_data,
     get_pipeline_status_lock,
     get_graph_db_lock,
+    get_data_init_lock,
 )
 
 from .base import (
@@ -602,13 +603,16 @@ class LightRAG:
 
     async def check_and_migrate_data(self):
         """Check if data migration is needed and perform migration if necessary"""
+        async with get_data_init_lock(enable_logging=True):
             try:
                 # Check if migration is needed:
                 # 1. chunk_entity_relation_graph has entities and relations (count > 0)
                 # 2. full_entities and full_relations are empty
 
                 # Get all entity labels from graph
-            all_entity_labels = await self.chunk_entity_relation_graph.get_all_labels()
+                all_entity_labels = (
+                    await self.chunk_entity_relation_graph.get_all_labels()
+                )
 
                 if not all_entity_labels:
                     logger.debug("No entities found in graph, skipping migration check")