For instance, if I were to store an adjacency list representation of a graph in LMDB, I would use vertex ID as the key, and the value is a vector of 64-bit IDs. To insert an edge into the graph will simply require appending 8 bytes to the end of the vector. Re-reading and re-writing the entire value will be somewhat excessive (and will actually change the asymptotic runtime of the insertion).
This is also one of those situations where a batch append might be useful. For instance, if I have an edge list on file which is too large to retain in memory. To convert it to an adjacency list in LMDB will require batch-append. Batch-writes are insufficient since that will require me to cache at least, large parts of the adjacency list in memory, and the original edge list could be ordered arbitrarily.
The actual graph representation I am thinking of is somewhat more intelligent than a straight adjacency list, but similar concepts hold.
I have tried looking through the lmdb source to try to implement append myself, but the put logic is somewhat complicated...
Yucheng