Page History

...

• The current way the heap backends iterates state values puts a constraint that the keyed state backend needs an efficient way to know the total number of of (NS, K) state value mappings state entries, so that this count can be written before the contiguous state values. This highly depends on the state backend; for example, for RocksDB obtaining this count is non-trivial. The approach that RocksDB adopts, by injecting markers and flags to indicate the end of states and key groups when reading from the file streams is definitely the more general and future-proof approach.
• The RocksDB backend snapshot format has excessive written data due to the key group bytes prepended for each (NS, K) state value mapping. As previously explained, the key group bytes are actually not required to be written with every state mapping in snapshots; they are only required for working state maintained by RocksDB to have an ordering of key value entries by key group. However, the size overhead of this redundancy would be far less prominent than the extra work that RocksDB otherwise requires to truncate and prepend again the key group bytes on each snapshot and restore operation.
• Keyed MapState are written as a single state value with the current heap backend's snapshot layout, while for RocksDB backend, a single state value maps to an entry in the MapState. The advantage in maintaining entries in the MapState as independent key-values in RocksDB is obvious - accessing and updating random keys in the MapState would not require serializing and deserializing the whole map. Trying to merge the separate entries when snapshotting so that the snapshotted version of MapState is a single state value would not work, since a user MapState can have arbitrarily large number of entries that would not fit into memory (again, this is a constraint because the serializer for writing maps, i.e. the MapSerializer, writes in a format that requires the total number of mappings to be known upfront). Moreover, the work of merging the entries on snapshot and flattening them again simply is non-trivial compared to dumping the key value bytes in RocksDB.

...