Enhance the caching policy

Currently, it caches each node with a node address instead of a key-value pair.
It's helpful because it reduces the disk IO queries, but the cache can be redundant, depending on the usages.
If the real usage is querying some keys repeatedly, caching key-value pairs is enough, and caching all nodes is just a memory wasting.
However, if the assumption is wrong, caching a key-value pair would not be helpful because of the high missing rates. We should investigate the real scenario before changing the policy.