Neo4j stores nodes, edges, and properties on disk in stores that are specific to each type—for example, nodes are stored in the node store.[5, s.11] They are also stored in two types of caches—the file system (FS) and the node/relationship caches. The FS cache is divided into regions for each type of store, and data is evicted on a least-frequently-used (LFU) policy.
Data is written in transactions assembled from commands and sorted to obtain a predictable update order. Commands are sorted at the time of creation, with the aim of preserving consistency. Writes are added to the transaction log and either marked as committed or rolled back (in the event of a failure). Changes are then applied (in sorted order) to the store files on disk.
It is important to note that transactions in Neo4j dictate the state and are therefore idempotent by nature.[5, s.34] They do not directly modify the data. Reapplying transactions for a recovery event simply replays the transactions as of...