Skip to content

Mutable State

The CAP Theorem (The Classic)

The Three Pillars:

  • Consistency (C): Every read receives the most recent write or an error. (Essentially, "all nodes see the same data at the same time").
  • Availability (A): Every request receives a (non-error) response, without the guarantee that it contains the most recent write.
  • Partition Tolerance (P): The system continues to operate despite an arbitrary number of messages being dropped (or delayed) by the network between nodes.

The "Pick Two" Reality:

In a distributed system, Network Partitions (P) are inevitable. Therefore, you aren't really picking between C, A, and P. You are picking how the system behaves when a partition happens:

  1. CP (Consistency + Partition Tolerance): If a partition occurs, the system shuts down the "out of sync" nodes to ensure data stays correct. It sacrifices Availability.
    • Example: HBase, MongoDB (usually), Banking systems.
  2. AP (Availability + Partition Tolerance): If a partition occurs, nodes keep answering requests even if they can't talk to each other. It sacrifices Consistency (users might see old data).
    • Example: Cassandra, DynamoDB, CouchDB.

The PACELC Theorem (The Extension)

The Formula: (P + A or C) ELSE (E) (L or C)

IF there is a Partition (P):

  • How does the system trade off Availability vs Consistency?

ELSE (E) (during normal operation):

  • How does the system trade off Latency vs Consistency?

Common PACELC Classifications:

  1. PA/EL (The "Fast" Systems):
    • During a partition, they stay Available.
    • Otherwise (Else), they prioritize low Latency.
    • Reality: These are "Eventual Consistency" systems. They are always fast, but never strictly consistent.
    • Example: Cassandra, DynamoDB.
  2. PC/EC (The "Strict" Systems):
    • During a partition, they choose Consistency (and go down).
    • Otherwise (Else), they choose Consistency (waiting for replicas to sync before responding).
    • Reality: These systems are always "correct" but can be slow and fragile.
    • Example: BigTable, HDFS.

  3. PC/EL (The "Hybrid"):

    • Chooses Consistency during a partition, but prioritizes Latency during normal times.
    • Does not make sense. You are trying to be "Strictly Consistent" only when the network breaks, but "Eventually Consistent" when it's fine. Mathematically, that doesn't work.

    The "Fake" PC Reality:

    • If the system is EL, some nodes already have old (stale) data while you are just minding your own business.
    • If a Partition (P) hits right then, the system can't magically become Consistent (C) because the data was already messy before the break happened.
    • You can't "protect" consistency during a crisis if you weren't maintaining it during peace time.