How does Multi-Paxos reduce consensus latency from 2-RTT to 1-RTT compared to Basic Paxos?

In Leslie Lamport's Basic Paxos, every individual state machine command requires two full round-trips over the network: Phase 1 (Prepare / Promise) to establish leadership and discover conflicting proposals, and Phase 2 (Accept / Accepted) to commit the value. For high-throughput transactional databases, paying 2 RTTs per write across geodistributed data centers is prohibitive. Multi-Paxos observes that Phase 1 can be executed ONCE across all future log slots simultaneously during leader election. Once a leader receives Phase 1 Promises for all future instances, it retains a stable ballot number. Subsequent client writes skip Phase 1 entirely and execute Phase 2 directly in a single RTT (Leader sends Accept, followers respond with Accepted), cutting commit latency in half.

What is a Leaseholder and how does it unlock 0-RTT consistent reads in distributed SQL engines?

In standard consensus protocols, serving a strictly consistent read requires either: (1) Running a write consensus round for the read (1 to 2 RTTs), or (2) Executing a ReadIndex heartbeat to a majority of followers to verify leadership hasn't changed. Both approaches introduce network hops. In Spanner and CockroachDB, the leader acquires a time-bounded Leader Lease granted by a majority of replicas. During the lease duration (guaranteed by physical monotonic clocks with bounded drift), no other replica will initiate an election or grant a higher ballot. Because leadership is guaranteed, the leaseholder can serve strongly consistent reads immediately out of local RAM in 0 RTT (sub-millisecond latency) without contacting any remote peer.

How does Multi-Paxos handle log pipelining and out-of-order commits?

Under high concurrency, waiting for slot i to be committed before proposing slot i+1 creates a pipeline bubble. Multi-Paxos allows the leader to pipeline proposals: it issues Phase 2 Accept requests for slots i, i+1, i+2 concurrently without waiting for prior acknowledgments. If slot i+1 receives majority Accepted responses before slot i (due to network reordering), the leader marks slot i+1 committed but delays applying it to the application state machine until slot i is successfully resolved and committed. This preserves the serializable linear log invariant while achieving line-rate pipeline saturation.

What happens when a Leaseholder fails and how is split-brain avoided during lease handover?

If a leaseholder crashes, followers must wait until the current lease timestamp expires before electing a new leader and acquiring a new lease. This brief pause (typically 1 to 3 seconds, matching the lease window) guarantees that two servers can never hold valid leases simultaneously. In systems with hardware TrueTime or Hybrid Logical Clocks (HLC), lease renewals use graceful handover: the current leaseholder coordinates a clean transfer to a new node, shortening the unavailability window to a few milliseconds without ever risking split-brain or dirty reads.

Multi-Paxos & Leaseholder Architecture Studio: 1-RTT Consensus & Local Reads

Verified Specification | Standardized Formulas | Instant Precision

Secure & Private (Zero Data Retention) Free Access • No Sign-Up

Multi-Paxos & Leaseholder Architecture Studio

Architect production distributed consensus engines: simulate Multi-Paxos 1-RTT fast path, benchmark 0-RTT local leaseholder reads, and inspect log instance pipelining.

Leaseholder Active Multi-Paxos 1-RTT Fast Path

0.12 ms

Local Consistent Read Latency (0 RTT)

18.4 ms

Consensus Write Latency (1 RTT)

1,240 Slots

Active Pipelined Log Instances

2,450 ms

Remaining Lease Duration

1. Multi-Paxos Consensus & Lease Configuration

Read Transaction Strategy

Cluster Cross-DC Network Topology

Lease Renewal Interval

2. Pipelined Log Instances (Slots #1040..#1047)

3. Multi-Paxos Wire Protocol Trace

Phase 2 Fast Path OK

Multi-Paxos & Leaseholder Architecture Studio

1. Multi-Paxos Consensus & Lease Configuration

2. Pipelined Log Instances (Slots #1040..#1047)

3. Multi-Paxos Wire Protocol Trace

4. Production Multi-Paxos & Leaseholder Code

🔍 Companion Workbenches & Verification Tools

Linux eBPF bpf_iter & Kernel Task Iterators Studio | In-Kernel Streaming Observability

SQL Query Formatter & Beautifier

Linux io_uring Zero-Copy Network Transmit Studio | IORING_OP_SEND_ZC & splice()

Linux TCP Buffer Sizing, BDP & Autotuning Studio

Why Smart Users Are Switching to This Free Toolkit

The Browser Feature Most People Never Use

Stop Paying for Software You Can Get Free

File Conversion Tricks That Save Hours Every Week

This Free Tool Replaced 5 Paid Apps for Our Team

What Your Browser Can Do Without Any Extensions