What Is Consensus

Imagine you're withdrawing money from an ATM while simultaneously your spouse is transferring funds from the same account using a mobile app. Both operations hit different servers in different data centers. How does the banking system ensure that your shared account balance remains correct? How do all the servers agree on the final state without allowing you to overdraw?

This seemingly simple scenario encapsulates one of the most profound challenges in computer science: the consensus problem. Understanding consensus is not merely an academic exercise—it's the intellectual foundation upon which reliable distributed systems are built. Every database that replicates data, every blockchain that maintains a ledger, every configuration service that coordinates distributed applications—all rely on solving the consensus problem.

In this page, we'll develop a rigorous understanding of what consensus means, why it matters, and how it manifests in real-world systems.

At its heart, the consensus problem asks: How can a group of distributed processes agree on a single value, even when some processes may fail?

This question, while deceptively simple to state, captures a profound challenge. In a distributed system, processes communicate only by exchanging messages over an unreliable network. Messages may be delayed, lost, or delivered out of order. Processes may crash at any moment. Despite all this uncertainty, we need all correctly functioning processes to eventually agree on some value—and that value must be meaningful.

The Abstract Formulation:

Consider a system of n processes, each starting with an initial proposed value. The consensus problem requires designing a protocol such that all processes eventually agree on exactly one of the proposed values. This simple statement hides enormous complexity, as we'll discover.

Why 'Agreement' Is Harder Than It Sounds:

In a centralized system, achieving agreement is trivial—a single process decides, and that's the answer. But distributed systems exist precisely because we want to avoid single points of failure. The moment we introduce multiple processes that must agree, we face fundamental coordination challenges:

1. No global clock: Processes cannot simply timestamp their actions and pick the earliest
2. Asynchronous communication: We cannot distinguish a slow process from a failed one
3. Partial failures: Some processes may fail while others continue operating
4. Network partitions: Processes may temporarily be unable to communicate with each other

These challenges aren't engineering obstacles to overcome with better hardware—they're fundamental limits of distributed computation that any consensus protocol must grapple with.

Any correct consensus protocol must satisfy three fundamental properties. These properties are not merely desirable—they define what it means to solve consensus. Understanding them deeply is essential for reasoning about the correctness of distributed systems.

The formal properties are:

Understanding Each Property Deeply:

Agreement ensures consistency. Without agreement, different parts of your distributed system would have conflicting views of reality. In a database, this would mean different replicas showing different data. In a configuration service, different servers would have different configurations. Agreement is the property that makes consensus useful for building reliable systems.

Validity ensures meaningfulness. A trivial 'protocol' that always decides 0 regardless of inputs would technically satisfy agreement (all processes decide 0) and termination (immediately). But it's useless—it ignores the actual proposals. Validity forces the protocol to actually consider the inputs and choose one of them.

Termination ensures progress. A protocol that runs forever, never deciding, isn't solving anything. Termination guarantees that the system eventually makes forward progress. This is crucial for practical systems—we can't wait indefinitely for decisions.

Consensus isn't an abstract theoretical concept—it's the foundation of practical distributed systems. Nearly every reliable distributed service you use daily relies on consensus in some form. Let's examine the major applications:

The Replicated State Machine Pattern in Depth:

The replicated state machine (RSM) pattern deserves special attention because it's the canonical application of consensus. The idea is profound in its simplicity:

1. Model your service as a deterministic state machine
2. Use consensus to agree on the sequence of inputs (commands) to the state machine
3. Every replica applies the same sequence of commands in the same order
4. Since the state machine is deterministic, all replicas end up in the same state

This pattern transforms the problem of replicating arbitrary state into the problem of replicating a log of commands. And replicating a log is exactly what consensus gives us—agreement on an ordered sequence of values.

When studying consensus, it's important to distinguish between single-instance consensus (agreeing on one value) and multi-instance consensus (agreeing on a sequence of values). Both are fundamental, but they have different characteristics.

Single-Instance Consensus:

This is the basic theoretical construct—a protocol that allows a group of processes to agree on a single value. Examples include the basic Paxos algorithm (single-decree Paxos) and simple leader election. Single-instance consensus is typically useful for one-time decisions:

• Electing a leader for a term
• Deciding whether to commit or abort a transaction
• Agreeing on a single configuration value

Multi-Instance Consensus (Replicated Logs):

Practical systems usually need to agree on not one value but a continuous sequence of values—a log. Each position in the log represents an instance of consensus. This is the foundation of replicated state machines.

From Single to Multi-Instance:

A naive approach to multi-instance consensus is to run a separate single-instance protocol for each log position. This works but is inefficient. Practical systems like Multi-Paxos and Raft optimize by:

1. Electing a stable leader who coordinates proposals for multiple instances
2. Batching multiple decisions to amortize the cost of message rounds
3. Pipelining proposals without waiting for previous instances to complete
4. Log compaction and snapshots to prevent unbounded log growth

These optimizations transform the theoretical construct into a practical building block. Understanding this progression from single-instance to optimized multi-instance consensus is key to understanding systems like etcd, Consul, and CockroachDB.

The entire point of solving consensus is to handle failures. If processes never failed, consensus would be trivial—designate one process as the decider, and done. The challenge arises precisely because failures can occur at any time.

Types of Failures:

Distributed systems must contend with various failure modes:

1. Process (node) failures: A server crashes and stops responding
2. Network failures: Messages are delayed, lost, or partitions occur
3. Timing failures: Clocks drift, timeouts expire incorrectly
4. Byzantine failures: Processes behave arbitrarily, including maliciously

Different consensus protocols tolerate different failure types with different thresholds.

The 2f + 1 Threshold:

For crash-stop failures, the famous result is that consensus requires a majority quorum—any two majorities must overlap. With n = 2f + 1 nodes, we can tolerate f failures while maintaining a majority of f + 1 correct nodes. The overlap between any two majorities ensures that at least one node participated in both, preserving consistency.

This is why you often see odd-numbered clusters: 3 nodes tolerate 1 failure, 5 nodes tolerate 2 failures, 7 nodes tolerate 3 failures.

The Critical Insight:

Consensus protocols don't eliminate failure—they mask failure. To users of the system, it appears as if the system never fails, even though individual components fail regularly. This is the magic of consensus: transforming unreliable components into a reliable abstraction.

Consensus is closely related to several other distributed coordination problems. Understanding these relationships helps clarify what consensus provides and how it fits into the broader landscape of distributed computing primitives.

The Equivalence Web:

One of the beautiful results in distributed computing theory is that many of these problems are equivalent in terms of computational power:

Consensus ≈ Atomic Broadcast ≈ Total Order Broadcast ≈ Leader Election (single-instance)

This means if you have a solution for any one of these, you can implement all the others. Consensus is the canonical representative of this equivalence class, which is why it receives so much theoretical attention.

Why This Matters Practically:

When evaluating distributed systems, understanding this equivalence helps you recognize:

• Systems that claim to solve replication without consensus are either making unstated assumptions or providing weaker guarantees
• The fundamental limits that apply to consensus (like FLP impossibility) apply equally to all equivalent problems
• Different systems may emphasize different problems (Zookeeper emphasizes atomic broadcast, Consul emphasizes leader election) but they're all solving the same underlying problem

Before diving into specific protocols in subsequent pages, let's establish a simple mental model for how consensus works at a high level. This will make the details of individual protocols more comprehensible.

The Basic Pattern:

Most consensus protocols follow a similar pattern, though the details vary significantly:

1. Proposal Phase: A process (often a leader) proposes a value
2. Voting/Acceptance Phase: Other processes vote on whether to accept the proposal
3. Commitment Phase: Once enough votes are gathered, the value becomes decided
4. Learning Phase: All processes learn the decided value

The key insight is that no single process can unilaterally decide—decisions require coordination among multiple processes, ensuring that even if some fail, the decision persists.

Why Multiple Phases?

The multi-phase structure isn't arbitrary—it's essential for correctness. Consider what would go wrong with a simpler approach:

• Single phase (just propose): If two proposers propose different values simultaneously, different nodes might accept different values, violating agreement.
• Two phases without ordering: Even with voting, concurrent proposals might both achieve majorities if the voting isn't serialized.

The additional phases introduce ordering and coordination that prevent conflicting decisions. How exactly this ordering is achieved differs between protocols (Paxos uses ballot numbers, Raft uses term numbers and leader leases), but the purpose is the same.

The Role of the Leader:

While leaderless consensus protocols exist, most practical protocols use a leader to improve efficiency:

• The leader serializes proposals, preventing conflicts
• Only one message round-trip is needed in the common case
• Leader failure triggers a leader election, which is itself a consensus instance

This leader-based approach is why systems like Raft and Multi-Paxos are so efficient in practice—they've optimized for the common case of a stable leader.