inline::sqlite-vec missing WAL mode and busy_timeout on vector store connections

[mfleader]

System Info

• llama-stack version: latest main
• Python: 3.12
• OS: Linux (UBI9)
• Deployment: Kubernetes, multiple uvicorn workers

Information

• The official example scripts
• My own modified scripts

🐛 Describe the bug

Bug Description

The inline::sqlite-vec provider opens database connections without setting WAL mode or busy_timeout. With multiple uvicorn workers, concurrent writes fail immediately with sqlite3.OperationalError: database is locked instead of waiting and retrying.

The KV store (src/llama_stack/core/storage/sqlstore/sqlalchemy_sqlstore.py:140-145) already handles this correctly by setting journal_mode=WAL, busy_timeout=5000, and synchronous=NORMAL. The vector store at src/llama_stack/providers/inline/vector_io/sqlite_vec/sqlite_vec.py:111 does not.

Steps to Reproduce

Create a run config with 2 workers:

run.yaml

server:                                                         
  port: 8321                                                                            
  workers: 2                                                                            

Start the server:

llama stack run run.yaml                                                               

Hit the vector store API from two terminals simultaneously:

# Terminal 1                                                                            
curl -X POST http://localhost:8321/v1/vector_stores \                                   
  -H "Content-Type: application/json" \                                                 
  -d '{"name": "test-store-1", "embedding_model": "all-MiniLM-L6-v2"}'                  

# Terminal 2 (run at the same time)                                                     
curl -X POST http://localhost:8321/v1/vector_stores \                                   
  -H "Content-Type: application/json" \                                                 
  -d '{"name": "test-store-2", "embedding_model": "all-MiniLM-L6-v2"}'                  

Error logs

sqlite3.OperationalError: database is locked 

Expected behavior

Concurrent writes should wait up to a timeout and succeed, matching the KV store's concurrency behavior. The fix is two lines in _create_sqlite_connection() after sqlite3.connect(db_path):

connection.execute("PRAGMA journal_mode=WAL")                                           
connection.execute("PRAGMA busy_timeout=5000")                                          

Actual Behavior

The second concurrent write fails immediately with no retry because the default
SQLite journal mode (DELETE) uses exclusive locks and no busy_timeout is set.