Simplifying Raft with Chaining

Raft is a consensus algorithm for deciding a sequence of commands to execute on a replicated state machine. Raft is famed for its understandability (relative to other consensus algorithms such as Paxos) yet some aspects of the protocol still require careful treatment. For instance, determining when it is safe for a leader to commit commands from previous leaders or when it is safe for servers to delete or overwrite commands... [Read More]

Neither Non-equivocation nor Transferability alone is enough for tolerating minority corruptions in asynchrony

In this post, we explore a theorem of Clement, Junqueira, Kate, and Rodrigues from PODC 2012 regarding the limits of non-equivocation. Informally, this theorem says that neither Non-equivocation nor Transferability alone is enough for tolerating minority corruptions in asynchrony. [Read More]

Benign Hotstuff

In this post we describe a simple variant of Paxos (or Raft or any Lock-Commit) that is inspired by looking through the lens of HotStuff and Blockchain protocols. The most noticeable difference is that while Paxos and Raft aim to maintain a stable Primary/Leader (and change views infrequently), in Benign Hotstuff, the Primary is rotated every round! A more subtle difference is that in Paxos and Raft each block of... [Read More]

Living with Asynchrony: the Gather protocol

A very useful tool in Asynchronus distributed computing is Reliable Broadcast, or simply called Broadcast. It allows a leader to send a message, knowing that all parties will eventually receive the same message, even if a malicious adversary control $f$ parties and $f<n/3$. Broadcast is deterministic and takes just a constant number of rounds. [Read More]

Good-case Latency of Byzantine Broadcast: the Synchronous Case

In our first post, we presented a summary of our good-case latency results for Byzantine broadcast (BB) and state machine replication (SMR), where the good case measures the latency to commit given that the broadcaster or leader is honest. In our second post, we discussed our results for partial synchrony, and described a new BFT SMR protocol named (5f-1)-SMR that can commit a decision within $2$ rounds in the good... [Read More]

Raft does not Guarantee Liveness in the face of Network Faults

Last month, Cloudflare published a postmortem of a recent 6-hour outage caused by a partial switch failure which left etcd unavailable as it was unable to establish a stable leader. This outage has understandably led to discussion online about exactly what liveness guarantees are provided by the Raft consensus algorithm in the face of network failures. [Read More]
Tags: raft dist101

A Simple and Succinct Zero Knowledge Proof

Many people have popularized the idea that succinct proofs and zero-knowledge proofs are a type of moon math. In this post, our goal is to present a simple proof system that can provide an introduction and intuition to this space. Perhaps surprisingly, the only tool we will use is the Theorem that non-trivial degree-at-most-$d$ polynomials over a field have at most $d$ roots. [Read More]

The Lock-Commit Paradigm: Multi-shot and Mixed Faults

In this follow up post to our basic Lock-Commit post, we show a multi-shot synchronous protocol for uniform consensus that can tolerate $f$ omission failures, given 2f < n. We then extend it to one that tolerates both $f$ omission failures and $k$ crash failures given k+2f < n. [Read More]
Tags: dist101

The Lock-Commit Paradigm

In this post, we explore one of the most celebrated and widely used techniques for reaching consensus: the Lock-Commit paradigm. This approach is a key technique of DLS88, Lamport’s Paxos, and many subsequent protocols. Protocols like Raft, PBFT, Tendermint, SBFT, Casper, HotStuff, etc., are all based on this paradigm. [Read More]
Tags: dist101

BFT Protocol Forensics

An important property satisfied by any Byzantine fault tolerant consensus protocol is agreement, which requires non-faulty replicas to not decide on conflicting values. Depending on the network model, typical consensus protocols tolerate only a fraction of Byzantine replicas. In particular, under partial synchrony or asynchrony, no consensus protocol with $n$ replicas can tolerate more than $n/3$ Byzantine faults. If the number of Byzantine replicas exceed this number, the protocols do... [Read More]

Living with Asynchrony: Bracha's Reliable Broadcast

In this series of posts, we explore what can be done in the Asynchronous model. This model seems challenging because the adversary can delay messages by any bounded time. By the end of this series, you will see that almost everything that can be done in synchrony can be obtained in asynchrony. [Read More]