How Does ZooKeeper Keep Data Consistent? The Paxos Algorithm Explained

ZooKeeper is a cluster of multiple servers, each of which can be connected by many clients that can read and modify the data stored on the servers.

ZooKeeper guarantees that the data on every server is exactly the same by using the Paxos algorithm, the foundational distributed consensus algorithm.

Paxos Basic Idea

Imagine a club with two roles: members and councilors (the decision‑making group). Members submit proposals to councilors to change club rules.

Councilors gather, vote on each proposal, and a proposal passes if it receives a majority of votes. Each proposal has a monotonically increasing ID.

Each councilor keeps a notebook containing the current rules and the highest proposal ID it has processed (initially 0). A proposal is accepted only if its ID is greater than the ID in the notebook; the councilor then updates its notebook with the new ID.

When a majority of councilors approve a proposal, it becomes effective and all councilors update their notebooks accordingly. Subsequent queries can be answered directly from the notebook.

In the case of concurrent proposals, only the one that is processed first will have an ID greater than the current notebook ID; the other proposal will be rejected and must be retried.

This mechanism maps to ZooKeeper as follows:

Member – client

Councilor – server

Councilor’s notebook – server’s data

Proposal – data‑change request

Proposal ID – zxid (ZooKeeper Transaction ID)

Proposal execution – applying the data change

ZooKeeper also introduces a leader role, which centralizes the right to initiate proposals, preventing multiple servers from issuing competing proposals simultaneously.

Through voting, a global numbering scheme, and leader coordination, Paxos ensures that at any moment only one write operation is approved, while competing writes must retry after losing the vote.