Happy Apache Cassandra 2: File Store Format

h3. Model Review

Cassandra data model

Column Family

Super Column Family

h3. File Format

h4. Keyspace

• Each Keyspace(Eg. Lobs)in separated directory
• Each ColumnFamily(Eg. object) in separated sstable files
  • ColumnFamilyName-version-#-Data.db
  • ColumnFamilyName-version-#-Index.db
  • ColumnFamilyName-version-#-Filter.db

h4. Index File 

• Each entry is [Key, Position], order by key
• Index Summary(pre-loaded into memory), every 1 entry in 128 entries
• Key Cache, cache the entry after access it
  

h4. data file

• Each entry is [Row], order by key
• Each Row contains columns, order by column name
• Each Row contains column index (in range)

h4. others

• Bloom filter file
  • Given {an element}, answer {is contained in a set} .
  • Given a {key}, answer {is contained in current sstable file}
• Commit log file
  • Used in write, avoid data lost 
• Statistical file

h3. CRUD

h4. write

BigTable

• First write to a disk commit log (sequential) 
• Update to appropriate memtables
• Memtables are flushed to disk into SSTable ( immutable)

h4. read

• Check row cache
• Cassandra will read all the SSTables for that Column Family
  • Bloom Filter for each SSTable to determine whether this SSTable contains the key
  • Use index in SSTable to locate the data (check key cache)
  • Read from data file

h4. update

• Same sequence with write
  • SSTable is immutable
  • Update is write into new SSTable
• Column Reconcile
  • Read columns from all the SSTables
  • Merge/Reduced columns with same name, CollationController
    • Use timestamp, return latest column

h4. delete

• Same sequence with write
  • SSTable is immutable
  • Delete is write into new SSTable
• Delete a column
  • Column flag is set to delete flag
• Delete a Row
  • markedForDeleteAt  is set to delete timestamp
  • Column timestamp is compared with markedForDeleteAt when reducing columns in read

h4. drop

• Drop a Column Family
  • Take snapshots (move to snapshot dir)
  • Remove column family definition
• Drop a Keyspace
  • Take snapshots (move to snapshot dir)
  • Remove keyspace definition

h4. Compaction

Periodically data files are merged sorted into a new file (and creates new index)

• Merge keys 
• Combine columns 
• Discard tombstones

References

http://jonathanhui.com/how-cassandra-read-persists-data-and-maintain-consistency