CAP Theorem

Distributed system can guarantee at most two of the three properties: Consistency, Availability, and Partition Tolerance.

The three properties

1️⃣ Consistency (C)

• Every read receives the most recent write

• All nodes see the same data at the same time

• Similar to single-machine correctness

📌 Example: After updating a user’s balance, all future reads return the updated balance.

---

2️⃣ Availability (A)

• Every request receives a response (success or failure)

• The system never refuses requests

• Does not guarantee the data is the latest

📌 Example: You always get a response, even if it’s stale.

---

3️⃣ Partition Tolerance (P)

• The system continues to operate despite network failures

• Messages between nodes can be lost or delayed

📌 Example: Data centers cannot talk to each other, but the system keeps running.

---

Why you can’t have all three

In real distributed systems:

• Network partitions are unavoidable → P is mandatory

• During a partition, you must choose:

  • Consistency (C) → reject some requests

  • Availability (A) → serve possibly stale data

So, you choose CP or AP.

---

The valid CAP combinations

✅ CP (Consistency + Partition Tolerance)

• System may become unavailable during partitions

• Ensures correctness over uptime

📌 Examples:

• HBase

• MongoDB (in majority mode)

• ZooKeeper

🧠 Use when: correctness is critical (banking, inventory, pricing)

---

✅ AP (Availability + Partition Tolerance)

• System always responds

• Data may be eventually consistent

📌 Examples:

• Cassandra

• DynamoDB

• Riak

🧠 Use when: uptime > strict accuracy (feeds, metrics, logs)

---

❌ CA (Consistency + Availability)

• Works only when there is no partition

• Not realistic for distributed systems

📌 Example:

• Single-node databases

• Strongly consistent DB in a single data center

---

One important clarification (often asked in interviews)

❗ CAP is about behavior during partitions

• Not about normal operation

• Many systems are CA when no partition exists

• Choice happens only when partition occurs

---

CAP vs Real Systems (modern view)

• Modern systems blur CAP using:

  • Quorums

  • Leader elections

  • Tunable consistency

Example:

• DynamoDB lets you tune R + W > N

• MongoDB lets you tune read/write concerns