can we extend the deduplicator to two cache buckets?
since we're fetching two buckets when reading states, then we can reconstruct deduplicated states from those two. OTOH, when fetching states from the past, we don't need to fetch two buckets.
Pick a position in this tradeoff.