Whenever a node performs a write operation, the node must know whether it should wait for the data to be written to stable storage before successfully returning from the operation.
As a part of performing a write operation, the data modification is first made to an in-memory cache. It is then written to the filesystem's data buffers. And, finally, the contents of the data buffers are synchronized to stable storage (typically, a hard drive).
You can control how much of this process the master node will wait to complete before it returns from the write operation with a normal status. There are three different levels of synchronization durability that you can require:
The data is written to the host's in-memory cache, but the master node does not wait for the data to be written to the file system's data buffers, or for the data to be physically transferred to stable storage. This is the fastest, but least durable, synchronization policy.
The data is written to the in-memory cache, and then written to the file system's data buffers, but the data is not necessarily transferred to stable storage before the operation completes normally.
The data is written to the in-memory cache, then transferred to the file system's data buffers, and then synchronized to stable storage before the write operation completes normally. This is the slowest, but most durable, synchronization policy.
Notice that in all cases, the data is eventually written to stable storage (assuming some failure does not occur to prevent it). The only question is, how much of this process will be completed before the write operation returns and your application can proceed to its next operation.
See the next section for an example of setting durability policies.