With sqlalchemy how to dynamically bind to database engine on a per-request basis

Question

I have a Pylons-based web application which connects via Sqlalchemy (v0.5) to a Postgres database. For security, rather than follow the typical pattern of simple web apps (as seen in just about all tutorials), I'm not using a generic Postgres user (e.g. "webapp") but am requiring that users enter their own Postgres userid and password, and am using that to establish the connection. That means we get the full benefit of Postgres security.

Complicating things still further, there are two separate databases to connect to. Although they're currently in the same Postgres cluster, they need to be able to move to separate hosts at a later date.

We're using sqlalchemy's declarative package, though I can't see that this has any bearing on the matter.

Most examples of sqlalchemy show trivial approaches such as setting up the Metadata once, at application startup, with a generic database userid and password, which is used through the web application. This is usually done with Metadata.bind = create_engine(), sometimes even at module-level in the database model files.

My question is, how can we defer establishing the connections until the user has logged in, and then (of course) re-use those connections, or re-establish them using the same credentials, for each subsequent request.

We have this working -- we think -- but I'm not only not certain of the safety of it, I also think it looks incredibly heavy-weight for the situation.

Inside the __call__ method of the BaseController we retrieve the userid and password from the web session, call sqlalchemy create_engine() once for each database, then call a routine which calls Session.bind_mapper() repeatedly, once for each table that may be referenced on each of those connections, even though any given request usually references only one or two tables. It looks something like this:

# in lib/base.py on the BaseController class
def __call__(self, environ, start_response):

    # note: web session contains {'username': XXX, 'password': YYY}
    url1 = 'postgres://%(username)s:%(password)s@server1/finance' % session
    url2 = 'postgres://%(username)s:%(password)s@server2/staff' % session

    finance = create_engine(url1)
    staff = create_engine(url2)
    db_configure(staff, finance)  # see below
    ... etc

# in another file

Session = scoped_session(sessionmaker())

def db_configure(staff, finance):
    s = Session()

    from db.finance import Employee, Customer, Invoice
    for c in [
        Employee,
        Customer,
        Invoice,
        ]:
        s.bind_mapper(c, finance)

    from db.staff import Project, Hour
    for c in [
        Project,
        Hour,
        ]:
        s.bind_mapper(c, staff)

    s.close()  # prevents leaking connections between sessions?

So the create_engine() calls occur on every request... I can see that being needed, and the Connection Pool probably caches them and does things sensibly.

But calling Session.bind_mapper() once for each table, on every request? Seems like there has to be a better way.

Obviously, since a desire for strong security underlies all this, we don't want any chance that a connection established for a high-security user will inadvertently be used in a later request by a low-security user.

Answer 1

I would look at the connection pooling and see if you can't find a way to have one pool per user. You can dispose() the pool when the user's session has expired

Answer 2

Binding global objects (mappers, metadata) to user-specific connection is not good way. As well as using scoped session. I suggest to create new session for each request and configure it to use user-specific connections. The following sample assumes that you use separate metadata objects for each database:

binds = {}

finance_engine = create_engine(url1)
binds.update(dict.fromkeys(finance_metadata.sorted_tables, finance_engine))
# The following line is required when mappings to joint tables are used (e.g.
# in joint table inheritance) due to bug (or misfeature) in SQLAlchemy 0.5.4.
# This issue might be fixed in newer versions.
binds.update(dict.fromkeys([Employee, Customer, Invoice], finance_engine))

staff_engine = create_engine(url2)
binds.update(dict.fromkeys(staff_metadata.sorted_tables, staff_engine))
# See comment above.
binds.update(dict.fromkeys([Project, Hour], staff_engine))

session = sessionmaker(binds=binds)()