Ceph: Designing and Implementing Scalable Storage Systems

Deferred writes

Unlike in filestore where every write is written in its entirely to both the journal and finally to disk, in BlueStore, the data part of the write in most cases is written directly to the block device. This removes the double write penalty and on pure spinning disk OSDs dramatically improves performance. However, as mentioned previously, this double write has a side effect of decreasing write latency when the spinning disks are combined with SSD journals. BlueStore can also use flash-based storage devices to lower write latency by deferring writes, first writing data into the RocksDB WAL and then later flushing these entries to disk. Unlike filestore, not every write is written into the WAL, configuration parameters determine ...