Cassandra

• Leaderless
• high write throughput
• append only
  • modifications are new rows with new timestamp
  • deletions are tombstones
• partition by id, sort by how you want to receive data
  • With Cassandra, you often start with: “What query will my application execute?”
    • Suppose the query is: “Give me the newest 50 messages for channel 123 during September”
    • partition key = (channel_id, month)
    • clustering key = message_id
• Generally AP for CAP, but consistancy guarantee is tunable
• Use bloom filters to see if the required data is in the SSTable