MapR-DB operates directly on the file system

MapR-DB tables are implemented directly in the MapR file system (MapR-FS).

One of the resulting advantages is that MapR-DB has no layers to pass through when performing operations on data. MapR-DB runs inside of the MFS process, which reads from and writes to disks directly. In contrast, Apache HBase running on the Hadoop file system (HDFS) must communicate with the HDFS process, which in turn must communicate with the ext3 file system, which itself ultimately writes data to disks. The approach taken by MapR-DB eliminates such process hops, duplicate caching, and needless abstractions, with the consequence of optimizing I/O operations on your data.

Another advantage is the absence of compaction delays that arise due to I/O storms as logged operations are merged with structures on disk. MapR-DB, like several other NoSQL databases, is a log-based database. Periodically, logged operations must be written to disk. In MapR-DB, tablets (called regions in Apache HBase) and smaller structures within them are stored partially as b-trees which together with write-ahead log (WAL) files comprise log-structured-merge trees. Write-ahead logs for the smaller structures within tablets are periodically restructured by rolling merge operations on the b-trees. Because MapR-DB performs these merges at small scales, applications running against MapR-DB see no significant effects on latency while the merges are taking place.