Generation Clock (Epoch)

The **Generation Clock** (also known as **Term**, **Epoch**, or **Generation Number**) is a critical logical clock pattern in distributed systems. It provides a mechanism to identify which node in a cluster is the *current* legitimate authority, preventing data corruption caused by "Zombie Leaders" (stale leaders that haven't yet realized they've been replaced).

1. The Problem: The "Zombie Leader"

In a distributed system, leadership is often managed via [Leases](LeasePattern). If a leader node experiences a long **Stop-the-World GC Pause** or a **Network Partition**, its lease may expire while it is "asleep."

1. **Lease Expires:** The cluster detects the leader is missing and elects a **New Leader**.

2. **The Zombie Wakes Up:** The **Old Leader** wakes up and *thinks* it still has a few seconds left on its lease (based on its own local clock).

3. **Conflict:** Both nodes attempt to write to shared resources (e.g., storage or database), leading to split-brain corruption.

2. The Solution: Fencing Tokens

The Generation Clock solves this by attaching a **Fencing Token** to every request.

The Protocol

1. **Increment:** Every time a new leader is elected, a central coordinator (or consensus cluster) increments a monotonic counter—the **Generation Clock**.

2. **Attach:** The current leader is given this value (e.g., `Term: 5`). The leader must include this number in every message it sends to followers or shared resources.

3. **Validate:** The receiving resource (the "Fenced" component) tracks the highest generation number it has ever seen.

* If a request arrives with `Term: 5` and the resource is at `4`, it accepts the write and updates its state to `5`.

* If the "Zombie" arrives with `Term: 4`, the resource **rejects** the request because `4 < 5`.

3. Terminology in Industry

While the pattern is identical, different systems use unique names for the Generation Clock:

| System | Name | Usage |

| :--- | :--- | :--- |

| **Raft** | **Term** | Incremented for every election cycle. Used to ignore stale RPCs. |

| **ZooKeeper** | **Epoch** | Incremented whenever a new leader starts a session. |

| **Kafka** | **Controller Epoch** | Ensures only one controller node manages the cluster. |

| **Cassandra** | **Generation** | Stored in the Gossip state to detect node restarts. |

4. Why "Clock"?

It is called a **Logical Clock** because it measures **Causality**, not seconds. In distributed environments where physical clocks are unreliable (due to drift and skew), the Generation Clock provides a "Happens-Before" relationship: a message with a higher generation number is mathematically guaranteed to be more recent than one with a lower number.