sectionColossus [24df3d4e]
- GFS single master could not hold everything
-
Reed Solomon vs full replication
- parity bits are added for error correction, whereas full replication only uses
-
Frontend has changed
- Curator -> horizontally scalable metadata servers
- Master metadata database
- Custodians -> horizontal scalers
- GFS was built on the assumption that file sizes were always going to outnumber metadata, but the biggest problem is that that may not be true, small blobs double their size with metadata
1. Client Library
1. Client Library
- Still the most complex part, plays the same role as the GFS client
2. Control Plane and Metadata Database
2. Control Plane and Metadata Database
- Metadata Database has moved to BigTable, but curators fill the role of serving metadata (horizonitally scalable metadata servers), and custodians manage the data (by moving files around), D file servers retain the same process as mapping over disks