A distributed system can only provide two of three guarantees simultaneously: Consistency, Availability, Partition tolerance.

Trade-off

      Consistency
          |
          |
   AP ----+---- CP
          |
          |
      Availability

In practice, networks WILL partition -> Partition tolerance is non-negotiable.
You choose between C and A during the partition.

Real picks

• CP (consistency > availability): HBase, ZooKeeper, etcd, banking ledgers.
• AP (availability > consistency): Cassandra, DynamoDB, CouchDB, social feeds.
• PACELC: extends CAP for the no-partition case — what trade-off do you make when the network is healthy?

Pitfalls

• Thinking “in CAP, I pick 2” is wrong — partitions are inevitable, so you always end up picking C or A under partition.
• “Eventual consistency” needs to be quantified — lag bound in seconds? SLA via timestamp.

Analogy

Two armies separated by fog. “Consistency” = both agree on the battle plan before acting. “Availability” = each army keeps fighting even if it can’t reach the other.

Interview tip: Name the picks real systems make. Cassandra = AP; ZooKeeper = CP.