Correct usage of session with asynchronous servers in SQLAlchemy

Question

Background:

We have a Python web application which uses SqlAlchemy as ORM. We run this application with Gunicorn(sync worker) currently. This application is only used to respond LONG RUNNING REQUESTS (i.e. serving big files, please don't advise using X-Sendfile/X-Accel-Redirect because the response is generated dynamically from Python app).

With Gunicorn sync workers, when we run 8 workers only 8 request is served simulatenously. Since all of these responses are IO bound, we want to switch to asyncronous worker type to get better throughput.

We have switched the worker type from sync to eventlet in Gunicorn configuration file. Now we can respond all of the requests simultaneously but another mysterious (mysterious to me) problem has occured.

In the application we have a scoped session object in module level. Following code is from our orm.py file:

uri = 'mysql://%s:%s@%s/%s?charset=utf8&use_unicode=1' % (\
    config.MYSQL_USER,
    config.MYSQL_PASSWD,
    config.MYSQL_HOST,
    config.MYSQL_DB,
)

engine = create_engine(uri, echo=False)

session = scoped_session(sessionmaker(
    autocommit=False,
    autoflush=False,
    bind=engine,
    query_cls=CustomQuery,
    expire_on_commit=False    
))

Our application uses the session like this:

from putio.models import session

f = session.query(File).first()
f.name = 'asdf'
session.add(f)
session.commit()

While we using sync worker, session was used from 1 request at a time. After we have switched to async eventlet worker, all requests in the same worker share the same session which is not desired. When the session is commited in one request, or an exception is happened, all other requests fail because the session is shared.

In documents of SQLAlchemy, said that scoped_session is used for seperated sessions in threaded environments. AFAIK requests in async workers run in same thread.

Question:

We want seperate sessions for each request in async worker. What is the correct way of using session with async workers in SQLAlchemy?

Answer 1

Use scoped_session's scopefunc argument.