[ZODB-Dev] ZEO and replication of BTree based objects

[Jeremy Hylton]

On Wed, 2003-06-11 at 08:21, Steve Alexander wrote:
> I thought it would be fun to instrument the ZODB so that it uses 
> something like syslog to send out a short message for every oid loaded, 
> and give the connection that loaded it.
> 
> This data (oids and when they are loaded, per connection) could be 
> gathered and analysed, and pre-fetching strategies could be tried out 
> against the data to see how effective they are.

Yes.  The Gruber thesis in mentioned my weblog used a benchmark
application (OO7), but traces from real applications would be more
interesting.  The trick, of course, is getting a trace you really care
about.

> I wonder if some of the bayesian / pattern-recognition techniques 
> developed for detecting spam might help?
> 
> The same kind of set-up could be used to analyse a connection in real 
> time, in a separate process, and send messages to the process to load 
> object that it is predicted might be required soon.

Perhaps the neural methods (perceptrons) used for branch prediction
would be more lightweight: http://camino.rutgers.edu/tocs02.pdf

Jeremy