Nodes in Swarm may wish to retain certain chunks for various reasons. For example, they might want to keep certain content available in Swarm either because their operator owns it or because they have been paid to do so. In this document, I propose an extension to Swarm's local database that would facilitate the management of such content.
The problem with the naive implementation of simply flagging sticky content is that it is not clear under what conditions can the flag be removed. Instead of flags, we can use reference counters, whereby chunks with a non-zero reference count can be considered sticky/flagged. Fortunately, Swarm references, both encrypted and unencrypted are guaranteed to be cycle-free. Thus, reference counters can be initialized by resetting all to zero and then simply traversing references from a given root while increasing the reference counter for each encountered chunk. Similarly, dereferencing requires a decrement at each chunk encountered during a traversal.
For practical reasons, I suggest that stickiness is defined by accessibility from a single root, stored locally. Thus, we can effectively reuse already existing tools for managing sticky content.
Unfinished traversals can leave the database in an inconsistent state. However, since concurrent traversals do not result in race conditions, if the state of all unfinished traversal is available, they can simply be finished, thus restoring consistency. The state of a traversal is described by a root reference and a path descriptor. In the path descriptor we find ordinal numbers for references found in intermediate chunks as well as those found in manifests. In both cases, in addition to unambiguously marking the location of the reference, we need to indicate whether or not that particular reference is encrypted or not.
While optimizations are obviously possible, the simplest way of updating reference counters in case of change of root reference is increasing counters in a tranversal from the new root and then dereferencing the old root.
If the states of ongoing traversals is lost, consistency can be restored by setting all reference counters to zero and traversing from the root.
In order not to require a full scan of all chunks, reference counters can have a label identifying the consistency epoch. If a new epoch is started, reference counters with the old label should be treated as zero.