It's best to keep object records small when using ZODB. This does mean you need to do some planning about object structure. Objects that inherit from "persistent.Persistent" are kept as separate records in ZODB, so you can break up a large object into several smaller ones by attaching persistent attributes to another object. If you just make a big structure out of nonpersistent objects, ZODB will have all the downsides of plain old pickle (like slow loading time for large objects) indeed, but its entire purpose is to allow you to not do this.
Ignoring ZRS- As your number of clients and transactions go up you're still bottlenecked in a single server. That's what I mean by doesn't scale. For various reasons (e.g. objects can refer to each other) you're basically stuck. A scalable database provide various means of growing as your load grows and ZODB does not.
There are a few problems with pickling. First of all it is slow. Under some assumptions there are faster ways of marshalling in Python. Secondly your granularity of access is the entire object. You can't just get a certain field out of a large object. Thirdly because objects in the client cache are pickled you are spending a whole lot of time serializing/de-serializing them when you don't really need to do that. In one of my applications that happens to account for 80% of the execution time.
I haven't spent a lot of time with SqlAlchemy but I think an ORM that maps well to some performant database is a better approach in Python.