A content-addressable data store with object capabilities
What is Nebula?
Nebula is a content-addressable file store that uses a capability model for security.
What problem are you trying to solve?
Users have data: they want to be able to keep track of revisions to this data, and they would like to be able to share this data with other users. Users would also like the ability to cluster into groups to share and collaborate on this data.
Secondary objectives are to build real world experience designing, implementing, and operating capability systems; and to characterise the behaviour of capability systems in the real world.
What are the characteristics of a solution?
- Users must be able to upload and retrieve data.
- Users must be able to view the history of their data.
- Users should be able to share data with other users.
- A user should be able to refer to a piece of data as a leaf in a history tree, as a node in the tree, or as an isolated snapshot with the history information stripped.
- Users should have some assurance as to the integrity and confidentiality of their data: one user should not be able to read another user's file unless permission has been explicitly granted or unless the other user has their own copy of that data.
Towards a solution
Data is referred to by the SHA-256 hash of the contents of the file. For technical reasons, this could be prefixed to reside in some directory tree structure. There are two options for this: use a prefix (such as the first n bytes of the ID, or where each byte is a directory. Example:
Example: given the SHA-256 ID
The first solution (with a prefix of 4) yields the path
The second yields
From a file system performance perspective, the two options are equally difficult to implement, and the latter will provide better performance. As data will always be referenced by a SHA-256 hash (a constraint that must be enforced in the code), this should not be a problem.
References to blobs
Users directly interacting with blobs presents two problems:
- Information leakage: if Alice wants to determine if someone already has a copy of some data, she attempts to read its SHA-256 digest. The server will return the data if she has it. This data is of no consequence to Alice, as she likely already had a copy of the data to produce the hash.
- Managing data is more difficult: in the case where a user asks the server to delete a file that multiple users have, the server has no way to determine what other users might have the data. One user can then remove data that other users may still wish to retain. Alternatively, the server might refuse to delete this data, which means users have no way to remove data from the system.
The solution is to only access blob IDs in the software, and to provide users with UUIDs to reference their data. A UUID contains
- the ID
- the referenced object ID (see below, this may be a SHA-256 ID or another UUID)
- the parent UUID
- UUIDs of any children
In order to provide useful access control, a reference may be a proxy reference: that is, it may refer to another blob reference. This means that a user can grant revocable access to the data without jeopardizing their own access.
Therefore, to know an ID is to have access to that ID. For this reason, users can only see the metadata and none of the IDs. The system needs an API that can traverse the history tree without exposing these IDs to users. Proxy objects either need to be presented with no history information (empty parent and children), or the entire history needs to be proxied. Similarly, a revocation API needs to be able to take into account that the entire history tree may be proxied.
This data must be stored in a persistent data structure, such as a database.
This reference is named an entry. This reference is the only interface users have with blobs.
Constantly referring to a UUID for file revisions is something that users will find awkward. A useful abstraction is a named history: presenting a single reference to a history tree that always provides the newest copy of some data, while still allowing users to traverse the history. This abstraction needs to pair some notion of the owner with a name of their choosing; this pairing is termed a file. Writing to a file creates a new entry with the parent set to the file's current reference, and the file's reference is updated to the new entry's ID.
Users will be identified by a UUID, as will collections of users (termed group). This allows groups and users to be interchangeable.
Challenge: how to deal with removing a user from a group? To know an ID is to have access to the ID, so new IDs will need to be generated for each object owned by a group; this change will need to be communicated to the group. Groups are not granular at this time: access to a group ID means all users can read or write entries and files. Group leadership will probably belong to a single user. This a subject that should be considered for revision in future.
- create, update, delete entries
- note that garbage collection will need to be done when a user entry is removed: if no other entry refers to a blob, that blob should be removed from the store.
- update creates a new entry with the parent of the new entry and the children field of the parent updated appropriately.
- create, update, delete, list files
- grant or revoke access
- this needs to account for the need to proxy histories
- group creation, inviting, transfer of ownership
A demo use case
A demo of the Nebula system would be to build an HTTP front end that uses Codemirror to implement a collaborative editor.
At some point it would be advantageous to sync data. Armstrong proposes the use of a DHT. However, implementing sync in this manner means that any participating node has access to all the blobs where no guarantee is made that peers are securing this data; this presents a large hole for data leakage. Participating nodes must have some sort of authentication. The most straight-forward mechanism for this is to communicate over an interface such as schannel with mutual authentication. This brings the complexity of requiring a signature authority trusted by all users. A synchronisation mechanism must operate in a hostile environment:
- At the core, user data must be protected: just as users expect their data to remain secure on the single node system, so too should their data be secured across all nodes.
- A participant should expect that some participants are actively trying to exploit data leakage.
- Participants must have strong mutual authentication, which implies strong identity. Nodes may be pseudonymous, but they cannot be fully anonymous. Peer reputation is a necessity.
- Communications must occur over a secure channel (see
Cryptography Engineering or
- Alternate (not schannel) alternatives should be explored. One alternative is hosts identified by a UUID and using remote attestation or another form of TPM-based authentication. Particularly interesting would be decentralised authentication and attestation, but it is difficult to see how trust could be bootstrapped this way.
Towards a production version
There seem to be two phases to the project now that a development version exists:
- The first step is to make this usable from a unikernel system.
- Once the unikernel system is built, work can proceed towards enabling this as a hypervisor service.
The unikernel stage aims to make this a standalone system that can operate with only some disk. The SQLite data store is out of the question, which means that some system for figuring out how to serialise and store the metadata is in order. An alternative interface to the current HTTP/REST interface will also need to be devised.
Code / demos
Last updated on 2015-03-30.