Ref: https://learn.cantrill.io/courses/1820301/lectures/41301416

• Summary:
• Contents:

CAP Theorem

• DB transactions have different properties depending on the DB model and architecture
• CAP = Consistency, Availability, Partition tolerance → DB transaction properties
  1. Consistency: read operations always return the most recent data, or if that's not immediately possible, they return an error
     • 💡 If you write a new item on the DB and immediately afterwards read the DB, the DB with the new item will be returned (or you get an error and don't receive anything)
     • ‼️ Non-consistent DBs can return DB states that are NOT the most recent!
  2. Availability: every request to the DB will always receive a non-error response, but no guarantee that the response will contain the most recent data
  3. Partition tolerance (or partition resilience): a DB system can be made of multiple NW partitions, and will continue to operate EVEN IF there are faults in the NW
     • e.g. errors between NW nodes, or messages between nodes getting dropped
• 🔧 CAP Theorem → A DB system can not have all 3 CAP properties, it must CHOOSE 2
  • 💡 Every pair of CAP properties brings its own pros, cons and tradeoffs
  • Consider a DB cluster in a NW. If some of the NW nodes fail, you have 2 options:
    1. Reject read operations and return an error
       • Reduces availability, but ensures consistency
       • Non-error answers are guaranteed to contain the latest write (i.e. latest DB state)
    2. Accept read operations and return something
       • Improves availability, but risks consistency
       • The answer might not contain the latest write
    • 💡 If a DB is spread in multiple NW nodes, you can't have both consistency and availability. There's no way around this.
  • Additional reference: https://en.wikipedia.org/wiki/CAP_theorem
• ACID and BASE are two DB transaction models that choose different tradeoffs
  • ACID → Ensures consistency
  • BASE → Opts for high availability

ACID Transaction Model

• DB transactions are Atomic, Consistent, Isolated and Durable (ACID)

  

  1. Atomic = Either all parts of a transaction are successful, or none are
     • Transactions succeed or fail as a whole
     • 💡 Consider a $10 transaction from bank account A to bank account B
       • Part 1: take out $10 from account A + Part 2: put $10 in account B
       • Partial success is not desirable → it could mean money is lost (if only part 1 succeeds) or money is duplicated (if only part 2 succeeds)
  2. Consistent = Transactions move the DB from one valid state to another valid state
     • Nothing in between is allowed, DB can never be in an invalid state
     • Valid states determined as per rules of the DB
       • 💡 e.g. in a SQL DB, all rows must have values for all columns. DB can't be in a state where that's not the case
     • Additional reference: https://en.wikipedia.org/wiki/ACID#Consistency
  3. Isolated = Different transactions executed in parallel don't interfere with each other
     • End state is the same as if all transactions had been sequentially executed
     • DB can function with multiple apps or users accessing the DB simultaneously
  4. Durable = Once committed, transactions are stored on non-volatile memory
     • Transactions remain committed, resilient to system failures (outages or crashes)
     • 💡 Operation reported as successful by the DB → data stored securely
       • If system now fails, or power fails, or a NW/node restarts, data is unaffected

• ‼️ 👎 ACID transactions limit the ability of a DB to scale!

  • Transactions have strict and rigid characteristics

• Most SQL DBs (RDBMS) use ACID-based transactions

  • Useful for e.g. finances → very rigid structure for data transactions

BASE Transaction Model

• DB transactions are Basically Available, Soft state and Eventually consistent (BASE)

  

  1. Basically Available = R/W operations available “as much as possible”, without consistency guarantees
     • Data availability increased by spreading and replicating data across all of the different NW nodes
     • The DB does its best to be consistent, but no guarantee (not the main focus)
     • 💡 Reads and Writes are “kinda, maybe”
  2. Soft state = Consistency is not enforced in DB, instead offloaded onto consumers
     • Developers must care for consistency and state in the apps that use BASE DBs
     • By default, returned data might not be the latest data, so…:
       • EITHER apps need to tolerate this fact…
       • OR apps must specify use of consistent operations, if the DB allows them
  3. Eventually consistent = Immediate consistency is not enforced
     • Consistency might happen… eventually! 😅
       • A read will match the latest write as long as we wait long enough
       • 💡 How long must we wait? Unspecified! We only know that eventually the read will be consistent with the latest write!
     • Additional reference: https://en.wikipedia.org/wiki/Eventual_consistency
  • 💡 All BASE characteristics are related (soft state implies eventual consistency, etc)

• ‼️ Many BASE DBs allow immediately consistent operations with guaranteed latest state!

  • Default = Eventually consistent operations with no guarantee for latest state
  • To guarantee latest state, app must have awareness and explicitly/specifically ask the DB for immediately consistent reads

• 👍 BASE-modeled DBs are highly scalable and can deliver really high performance

  • No need to worry about pesky consistency in the DB, apps can handle that
    • 💡 BASE transactions might initially sound pretty horrible due to all the inconsistency, but they actually bring lots of benefits with scalability and performance

• BASE-modeled DBs are NoSQL DBs