Earlier this year, Yahoo, Facebook, and Hbase developers had a roundtable discussion where we agreed to support three types of flush in HDFS (API1, 2, and 3) and the append project aims to implement API2. Here is a proposal to expose these APIs to application users.
1. Three flush APIs
- API1: flushes out from the address space of client into the socket to the data nodes. On the return of the call there is no guarantee that that data is out of the underlying node and no guarantee of having reached a DN. New readers will eventually see this data if there are no failures.
- API2: flushes out to all replicas of the block. The data is in the buffers of the DNs but not on the DN's OS buffers. New readers will see the data after the call has returned.
- API3: flushes out to all replicas and all replicas have done posix fsync equivalent - ie the OS has flushed it to the disk device (but the disk may have it in its cache).
2. Support flush APIs in FS
- FSDataOutputStream#flush supports API1
- FSDataOutputStream implements Syncable interface defined below. If its wrapped output stream (i.e. each file system's stream) is Syncable, FSDataOutputStream#hflush() and hsync() call its wrapped output stream's hflush & hsync.
- In each file system, if only hflush() is implemented, hsync() by default calls hflush(). If only hsync() is implemented, hflush() by default calls flush().