First I think it’s good to clarify what I understand Bitcasa is trying to do and if I was their architect how I would do it.

This of course is all guessing and speculation.

They are trying to store all clients data for ever and cache all recently accessed data locally & predictively cache other data.  All while expiring from the cache old infrequently accessed data.

Features they have or will be implementing are:

• All data is backed up to a cloud. All data is deduplicated with all other client data
• All data is NOT accessible by any other client or them
• Files and / or folders can be shared seamlessly with other users.
• Local hard drive space is used as a cache by predictive algorithms

Here is how I would implement each one of these to create an overall architecture similar to what I think Bitcasa is providing.  At a high level I would follow the process of deduplicating all data at the block level across ALL clients, have each client encrypt (with client’s key) the reference to what blocks that make up their files and folders, store those encrypted references in a users file database or key->value store, compress and encrypt (with Bitcasa’s master key) new unknown blocks, then write those blocks back to the Bitcasa cloud storage.  Bitcasa would retain a master hash table of all known blocks and each clint would send a list of it’s blocks and then Bitcasa master block hash table would respond and tell the client to only send new, unknown, blocks back to the storage compressed and encrypted.

So breaking it down here is how the address each feature:

• All data is backed up to a cloud & All data is deduplicated with all other client data.

All un-encrypted blocks (not files) are hashed and those hashes are sent back to the master hash table at Bitcasa and then you get a list of what “new” blocks need to be backed up and duplicate blocks can be thrown away.

• All data is NOT accessible by any other client.

The next level is to take the blocks / hashes that make up a file and create an encrypted user client key hash of what blocks make up that hash. You may even need to take the block level down to a sector level to get away files that could fit in a single block or segment size.

• Files and / or folders can be shared seamlessly with other users.

Because file / folder references are only stored with an encrypted reference to what blocks make up those files you simple need to give the new client the list of those blocks.

• Local hard drive space is used as a cache by predictive algorithms

An additional piece of code is monitoring what files / hashes / blocks are being accessed and knows if they’re cached locally or need to be pulled remotely.  I believe the predictive part is where most of their patents are but unfortunately we won’t be able to find out about them for ~18 months, read Bitcasa gets an early start on IP acquisition.

Well, that’s what I think?

Thoughts?