Someone else may be able to shed more light on this and correct me (please do!) but thats my concern and would love to hear if it was a valid one or not.
Can you just do the same thing with JSON, Yaml, or your favorite serialization method of choice that does not have catastrophic consequences if parsing gets subverted?
Sure. You can even extend the `json` module to serialize custom data types not defined in the JSON spec.
This snippet is really helpful: http://www.djangosnippets.org/snippets/513/
One may think that if only the internal model of the user input is pickled, and the user only gets to update this internal model through web requests in defined ways, thus never getting to upload pickles or construct arbitrary complicated python objects, it should be relatively safe.
However, what remains is that if someone manages to blindly inject SQL, he can inject executable code into your application, instead of just injecting data.
When you upvote and need to sign in, the code to execute the upvote is stored to be executed once you sign in.
I love it
That's basically what App Engine does right there. It stores by an id, then the value is your object serialized as a protocol buffer.
The App Engine datastore allows you to look up by other properties than id because it indexes every field of your object in the background without you having to deal with it. You can query by pair-wise or more-wise indexes too, but you have to create those manually.
The point about schema evolution is important.
I've taken to using a pickled python object in a text field in every model. That flexibility to store arbitrary data in a hash is amazing. I've used it to, for example, represent twitter accounts that are oauthed, with password, following the business account, or just nothing special. 4 kinds of accounts, with logic in the model to deal with it properly, all without evolving the schema. I only started with the first 2, and added the latter two without a problem.
I'm pushing this a bit further now, and transitioning away from models altogether, and experimenting with a non-relational database like couchdb. I'm debating moving to google app engine too, but found some details (like getting on your own domain) to be pretty annoying.