Arvados

Background¶

Currently, the API server provides permission signatures for collections that have replication_desired=0. This means a client can:

1. Write some data
2. Create a collection "A" with replication_desired=0
3. Wait until the blocks are old enough to be deleted by keep-balance
4. Retrieve collection "A" and create a new collection "B" with the same manifest
5. Change replication_desired on collection "A" to 2

After this, collections "A" and "B" refer to blocks which keep-balance was allowed to delete.

(As long as the underlying storage devices don't fail, it should never be possible for a client to obtain a signed locator for a block that doesn't exist.)

Collections with replication=0 might be useful, but proper support will include:

• improving clients so they don't try to retrieve data from these collections
• improving API so it doesn't provide locator signatures for these collections

In the meantime, we should avoid situations where data seems to be safe but isn't.

Proposed fix¶

In keep-balance, when a collection has replication_desired=0, pretend it's 1.