This is how I was originally achieving this. As I said, it's very slow. I don't think it would be a good solution on its own because it would require that every client be constantly sampling all DHT nodes, and downloading all metadata and indexing it for a potential future lookup. It's a huge amount of additional load on the DHT.

I think a better solution would be some way for clients to query the DHT for a specific "pieces root", but I don't know if all clients publishing "pieces root" for the torrents they know about would also be a good idea. Some kind of distributed metadata database where clients can query would be ideal.