Multiple untrusted tenants, single trusted service
One backup system must be able to provide backup services to multiple clients. The service should be able to support diverse types of clients, including different operating systems and different backup schedules. A client must not be required to trust any other client with anything, but it must trust the backup service with its data.
Additionally, it must be possible to configure the service so that a client does not need to trust itself in the future. E.g., it must be possible to have clients that can add new data without also being able to delete or modify any of their own data. This prevents an attacker who compromises a client from being able to compromise backups made from that client prior to the compromise.
Backed up data should outlive any people involved in creating it. There must not be any single points of failure in accessing (but not necessarily in adding) backed up data. It must be possible to maintain offline copies of backups in relatively inaccessible and secure locations. It must be possible to maintain online copies in places that can't run custom software, e.g., cloud storage. It must be possible to access data long after this software stops being maintained.
Any data format changes must include upgrade paths. Wherever cryptographic, compression, chunking, or other algorithms are used in ways that affect stored data, there must be a clear story for how future versions of the software can transition to newer algorithms as needed. Wherever cryptographic keying material is used, it must be possible to gradually and securely roll over to new keying material.
All software needed for reading backed up data, including all dependencies, must have source code available. Whenever possible, source code and documentation should be stored along with backed up data, to maximize the chance that data will be recoverable after this software stops being maintained. Additionally, data formats and algorithms should be chosen with the intent of making data access without a working copy of this software as easy as possible.
Data must be protected against interrupted processes, power failures, bit rot, or any other form of accidental corruption. Any attempt to access data must either succeed with the correct data, or fail loudly.
It must be possible to verify the authenticity of stored data. It must be possible to detect inauthentic modifications to authentic data, and addition of inauthentic data. It should be possible to detect replay of previously authentic data and deletion of authentic data.
A client must authenticate to a server before the client is authorized to perform any backup operations. A server must authenticate to a client before the client trusts the server with its data. Data in transit between the client and server must be verifiable.
It must be possible to encrypt stored data. [TODO: Get a better understanding of how to maintain confidentiality while also supporting incremental transfer of changed data, and update this section. Consider using random padding. Don't forget about untrusted clients who can inject arbitrary cleartext.]
If availability and durability of data is valued over confidentiality, it must be possible to disable encryption of stored data. Additionally, it should be possible to selectively enable/disable encryption for some storage media but not others.
Data in transit must be encrypted using authenticated, ephemeral key agreement.
It must be possible to monitor the system and receive alerts about any potential problems.
Data integrity should be monitored by reading all stored data on a periodic basis, and alerting on any issues. Read or write errors during normal operation should also trigger alerts. Documentation should strongly recommend SMART monitoring, or any other sensible monitoring that might catch potential integrity or availability issues.
Per-client monitoring should be possible. E.g., the service could alert if a client has not made a backup recently enough. Ideally, it would also be possible to specify per-client backup tests that would trigger an alert on failure. E.g., a test could check that a backup contains a MySQL dump file with a timestamp close to the timestamp of the overall backup.
Per-data-copy monitoring should be possible. E.g., the service could alert if an offline medium has not been updated or verified recently enough.
The ecosystem around the backup software should be healthy. The more people who use and depend on it, the more likely backups will continue to be readable for longer without resorting to code archeology. The more import and export tools, the better. The more supported types of clients, the better. The more supported deployment environments, the better.
These principles should be followed, but they must not interfere with any primary principle. E.g., efficient network transfers must not enable one client to use the backup system as an oracle to violate the confidentiality of another client.
Ideally, the service should be able to run in a distributed mode. It should be possible to add new data to any server, and have that data propagate to all other servers. It should be possible to manage offline storage and cloud storage copies from any server.
Potentially, the service should be able to run in a multi-layer distributed mode, with distinct clusters where no single failure within a cluster affects operation of that cluster, and no single failure of an entire cluster affects operation of the entire system. E.g., a cluster could have multiple storage servers and multiple access servers. Each cluster could have an eventually-consistent, complete copy of all data, striped-with-parity across multiple storage servers. Any access server could operate on data within its cluster, replicate data to/from an access server in another cluster, and accept new backups from clients.
Resources should be used efficiently. Data should be deduplicated. Unchanged data shouldn't be sent over the network unnecessarily. It should be possible to specify a retention policy for when old data is deleted. Resuming an interrupted transfer should be possible.
Ease of use
It should be easy to add or remove a client, to swap drives, etc. Ideally, initial setup would also be easy. There should be a mode where the client and server are bundled together.
Diverse storage media
Hard drives, solid state drives, and cloud storage should be the primary targets. Tapes, optical media, and any other reasonable storage media should also be supported, at least to a limited extent. For example, it might be reasonable to require low seek latency for creating a new backup, but support bulk copying to tapes or optical media after the backup is created. To the extent possible, all media should support integrity verification in a single pass, and extraction of a complete set of or subset of data in a single pass.