~azzar1/unity/add-show-desktop-key

Notes on mod_python.Session

---------------------------

    Author: Tom Conway

    Date: 12/12/2007

``mod_python`` provides some automagic for cookie based sessions. It

carefully separates most of the session logic from how the session is

stored. The base class `BaseSession` contains most of the logic, and

``mod_python`` itself has three derived classes for storing session

objects in-memory, dbm, and on the filesystem. In each case, the

implementations use apache's locking mechanism to serialize updates to

the store of cookies. This mechanism takes care of mutual exclusion

between the multiple processes of an apache instance, but does not

provide any facility to provide any kind locking for multiple servers

sharing the filesystem for file bases session storage.  There is code

for storing sessions in MySQL (and SQLLite) floating round on the net,

though none has made it in to any distributions. This code uses the

underlying database to take care of the locking.

In the case of IVLE, we wish to be able to share the session objects not

merely between the separate processes of an apache instance, but between

the multiple servers in a load balancing cluster.  There are three high

level strategies we could use to deal with this:

1. Use a static load balancing strategy such as hashing the client's IP

   address to determine which node in the cluster should serve the

   request.

2. Use a SQL backend to store sessions, or create a filesystem based

   storage mechanism that does the necessary locking.

3. Work around the problem by using session objects in a way that avoids

   the locking problems.

Strategy 1 has the advantage that we could use in-memory or dbm session

storage without having to worry about race conditions between servers.

On the other hand, it can run into serious problems if the distribution

of IP addresses is such that load is not balanced. This can be the case

if an ISP uses NAT firewalling (some do!), since all the requests from

that ISP will aparently be coming from a single IP address and will therefore

be routed to the same node in the cluster. As well as the potential for

failing to balance the load, such a scheme, if it works routes an equal

proportion of requests to each node in the cluster. At times when overall

load is light, this may mean that we lose the opportunity to put nodes into

a powersaving mode, when they are superfluous.

Strategy 2, while having the advantage of avoiding race conditions, is likely

to be expensive. The use of a SQL backend is likely to be quite slow, and the

SQL backend itself will be subject to significant load (i.e. at least one op

per request). A filesystem based solution is likely to be quite slow too.

It has to work on a shared filesystem, for which locking is a general issue

(generally, you end up using `mkdir` as the mechanism for creating a lock).

If we want mutable session information, then we will *have* to do something in

this vein.

Strategy 3 is fragile because we need to be careful about how we use

session objects, but if the constraints are simple enough to be practicle

then avoiding the locking issue is highly desirable. A simple constraint

that may be workable is to require that once created, a session object is

treated as read-only until it is deleted. It is possible (though unlikely)

we could create session objects that immediately become orphaned, but we

will not ever create a situation in which the application does anything bad.

If we can make strategy 3 work, then it is easily the best strategy to use.

The main use for session objects in IVLE will be to *cache* authentication and

authorization information. This means that when a user logs in, we authenticate

(the authentication mechanism is not important to our current discussion),

then retrieve the authorization information for that user, and store it in

the session object. For each page access until the user logs out, we can then

use the information from the session object.