PBFT Architecture

The Sawtooth PBFT algorithm is a voting-based consensus algorithm with Byzantine fault tolerance, which ensures safety and liveness. The network can tolerate a certain number of “bad” nodes. As long as this number is not exceeded, the network will work properly. In addition, blocks committed by nodes are final, so there are no forks in the network.

The nodes on the network send many messages to reach consensus, commit blocks, and maintain a healthy leader node, called a primary node. The network switches to a new primary (called a view change) at regular intervals, as well as when the current primary is faulty.

  • The primary node builds and publishes blocks.

  • The other nodes (called secondary nodes) vote on blocks and the health of the leader.

Sawtooth PBFT runs on each node in the network as a consensus engine, a separate process that handles consensus-related functionality and communicates with the validator through the consensus API.

The following sections describe Sawtooth PBFT architecture:

  • Network overview: Describes PBFT fault tolerance, view changes, sequence numbers, and the information stored by each node

  • Consensus messages: Explains consensus message structures and message types

  • Sawtooth PBFT operation: Shows how the algorithm handles initialization, normal mode (block processing), and view changes

Network Overview

Fault Tolerance

A PBFT network consists of nodes that are ordered from 0 to n-1, where n is the total number of nodes in the network. The on-chain setting sawtooth.consensus.pbft.members lists all PBFT member nodes and determines the node order.

The PBFT algorithm guarantees network safety as long as the number of faulty nodes remains below the required percentage. The maximum number of faulty nodes that the network can tolerate is determined by the formula f = \frac{n - 1}{3}. In other words, no more than one third of the nodes (rounded down) can be unreachable or dishonest at any given time.

For example, a four-node network can tolerate one faulty node. (PBFT requires a minimum of four nodes in order to maintain Byzantine fault tolerance.) Increasing the size of the network reduces the likelihood that all \frac{n - 1}{3} nodes would be faulty at the same time.

View Changes: Choosing a New Primary

A view is the period of time that a given node is the primary, so a view change means switching to a different primary node. The next primary is selected in a round-robin (circular) fashion, according to the order of nodes listed in the on-chain setting sawtooth.consensus.pbft.members.

In a four-node network, for example, the first node (node 0) is the primary at view 0, the second node (node 1) is the primary at view 1, and so on. When the network gets to view 4, it will return to node 0 as the primary.

The algorithm uses the formula p = v mod n to determine the next primary. In this formula, p is the primary, v is the view number, and n is the total number of nodes in the network. For example, if a four-node network is at view 7, the formula 7 mod 4 determines that node 3 is the primary.

The Sawtooth PBFT algorithm changes the primary at regular intervals, as well as when the secondary nodes determine that the current primary is faulty. See View Changing Mode for a description of this process.

Sequence Numbers

In addition to moving through a series of views, the network moves through a series of sequence numbers. In Sawtooth PBFT, a node’s sequence number is the same as the block number of the next block in the chain. For example, a node that is on sequence number 10 has already committed block 9 and is evaluating block 10.

Also, each message includes a sequence number that indicates which block the message is for. For example, a message with sequence number 10 applies to block number 10.

Information Storage

Each node stores several key pieces of information as part of its state:

  • List of PBFT member nodes in the network (from sawtooth.consensus.pbft.members)

  • Current view number, which identifies the primary node

  • Current sequence number, which is also the number of the block being processed

  • The current head of the chain

  • If in normal mode, the step of the algorithm it’s on (see Normal Mode)

  • Log of all blocks it has received

  • Log of all messages it has received

Network Configuration

Sawtooth PBFT configures the network with on-chain settings, which are processed by the Settings transaction processor (or an equivalent).

These settings list each node in the network, set the view-change interval (how often the primary changes), and specify other items such as the block publishing frequency, timeout periods, and message log size. For more information, see Configuring PBFT.

Consensus Messages

All PBFT consensus messages are serialized as protobufs (protocol buffers).

When a node receives a new consensus message from another node, it parses the protobuf into a native Rust struct (ParsedMessage) that allows for easier handling of the message. After parsing, the node performs a series of checks to ensure the validity of the message:

  1. The signer_id of the PBFT message must match the signer_id of the PeerMessage that contains it.

  2. The message must be from a known PBFT member (the signer_id of the PBFT message must be in the sawtooth.consensus.pbft.members list)

The first check verifies that the PBFT message was created and signed by the node that it claims to be from. The PeerMessage, which acts as a wrapper for the PBFT message, contains a signature of the PBFT message and a signer_id; this signature and ID are verified by the validator to ensure that the contents of the PeerMessage are legitimate. PBFT then ensures that the signer_id of the PBFT message matches the one the validator used, which guarantees the origin of the PBFT message.

The second check ensures that only nodes that are accepted members of the PBFT network are able to participate in the consensus process.

Any messages that fail to parse or pass the required checks are ignored. If a message is successfully parsed and passes verification, it is passed to a handler for that specific message type (see Message Types), where it may go through further checks, be stored in the message log, or trigger some actions.

Message Definitions

Most Sawtooth PBFT messages use the PbftMessage protobuf, as shown below. The PbftNewView, PbftSignedVote, and PbftSeal protobufs are structurally different from PbftMessage and are used for messages that require different sets of data to be exchanged.

Sawtooth PBFT also uses some of the message types defined in the consensus API, such as BlockNew and BlockCommit. These messages are called “updates” to distinguish them from the PBFT-specific messages. For more information on the consensus API’s Update messages, see the Consensus API RFC.

// Represents all common information used in a PBFT message
message PbftMessageInfo {
  // Type of the message
  string msg_type = 1;

  // View number
  uint64 view = 2;

  // Sequence number
  uint64 seq_num = 3;

  // Node who signed the message
  bytes signer_id = 4;

// A generic PBFT message (PrePrepare, Prepare, Commit, ViewChange, SealRequest)
message PbftMessage {
  // Message information
  PbftMessageInfo info = 1;

  // The block this message is for
  bytes block_id = 2;

// A message sent by the new primary to signify that the new view should be
// started
message PbftNewView {
  // Message information
  PbftMessageInfo info = 1;

  // A list of ViewChange messages to prove this view change (must contain at
  // least 2f messages)
  repeated PbftSignedVote view_changes = 2;

message PbftSignedVote {
  // Serialized ConsensusPeerMessage header
  bytes header_bytes = 1;

  // Signature of the serialized ConsensusPeerMessageHeader
  bytes header_signature = 2;

  // Serialized PBFT message
  bytes message_bytes = 3;

message PbftSeal {
  // Message information
  PbftMessageInfo info = 1;

  // ID of the block this seal verifies
  bytes block_id = 2;

  // A list of Commit votes to prove the block commit (must contain at least 2f
  // votes)
  repeated PbftSignedVote commit_votes = 3;

Message Types

A Sawtooth PBFT message has one of the following types:

  • PrePrepare: Sent by the primary node after it has published a new block

  • Prepare: Broadcast by every node in the Preparing phase

  • Commit: Broadcast by every node in the Committing phase

  • ViewChange: Sent by any node that suspects that the primary is faulty

  • NewView: Sent by the node that will be the new primary to complete a view change

  • Seal: Proves that a block was committed after 2f + 1 nodes agreed to commit it

  • SealRequest: Sent by a node that is requesting a consensus seal for the block that was committed at a given sequence number

PBFT Operation

The Sawtooth PBFT algorithm starts with initialization, then operates in one of two modes:


The original PBFT definition includes a checkpointing procedure that is responsible for garbage collection of the log. Sawtooth PBFT does not implement this checkpointing procedure; instead, it cleans the log periodically during its normal operation. For more information, see Log Pruning.


When the Sawtooth PBFT consensus engine starts, it does the following:

  • Loads its configuration

  • Initializes its state and message log

  • Establishes timers and counters

Normal Mode

In Normal mode, nodes check blocks and approve them to be committed to the blockchain. The Sawtooth PBFT algorithm usually operates in normal mode unless a view change is necessary.


The normal mode proceeds as follows:

  1. All nodes begin in the PrePreparing phase; the purpose of this phase is for the primary to publish a new block and endorse the block with a PrePrepare message.

    • The primary node will send a request to its validator to initialize a new block. After a configurable timeout (determined by the sawtooth.consensus.pbft.block_publishing_delay setting), the primary will send a request to the validator to finalize the block and broadcast it to the network.

    • After receiving the block in a BlockNew update and ensuring that the block is valid, all nodes will store the block in their PBFT logs.

    • After receiving the BlockNew update, the primary will broadcast a PrePrepare message for that block to all of the nodes in the network. When the nodes receive this PrePrepare message, they will make sure it is valid; if it is, they will add it to their respective logs and move on to the Preparing phase.

  2. In the Preparing phase, all secondary nodes (not the primary) broadcast a Prepare message that matches the accepted PrePrepare message. Each node will then add its own Prepare to its log, then accept Prepare messages from other nodes and add them to its log. Once a node has 2f + 1 Prepare messages in its log that match the accepted PrePrepare, it will move on to the Committing phase.

  3. The Committing phase is similar to the Preparing phase; nodes broadcast a Commit message to all nodes in the network, wait until there are 2f + 1 Commit messages in their logs, then move on to the Finishing phase. The only major difference between the Preparing and Committing phases is that in the Committing phase, the primary node is allowed to broadcast a message.

  4. Once in the Finishing phase, each node will tell its validator to commit the block for which they have a matching PrePrepare, 2f + 1 Prepare messages, and 2f + 1 Commit messages. The node will then wait for a BlockCommit notification from its validator to signal that the block has been successfully committed to the chain. After receiving this confirmation, the node will update its state as follows:

    • Increment its sequence number by 1

    • Update its current chain head to the block that was just committed

    • Reset its phase to PrePreparing

    Finally, the primary node will initialize a new block to start the process all over again.

This diagram summarizes the four Normal mode phases, the messages sent, and the interactions with the validators. N1 is the primary node; N2, N3, and N4 are secondary nodes.

PBFT normal operation procedure

Log Pruning

Sawtooth PBFT does not implement a checkpointing procedure (garbage collection of the log). Instead, each node cleans the log periodically during normal operation.

Log size is controlled on each node with the --max_log_size option when starting the PBFT consensus engine (see PBFT Command-Line Options). When a block is committed, each node compares the size of its log against the maximum size. If the log exceeds this value, Sawtooth PBFT uses these rules to prune the log:

  • Keep blocks and messages for the sequence number of the block that was just committed, plus those for any higher (newer) sequence numbers

  • Delete blocks and messages for all lower (earlier) sequence numbers

View Changing Mode

A view change switches to a different primary node. A node starts a new view change if any of the following occur:

  • The idle timeout expires - when a node enters the PrePreparing phase, it will start its idle timeout. If the node receives a new block and a matching PrePrepare from the primary for its current sequence number before the timeout expires, it will stop the timeout; if not, the node will initiate a view change when the timeout expires.

  • The commit timeout expires - when a node enters the Preparing phase, it will start its commit timeout. If the node is able to move on to the Finishing phase and send a request to the validator to commit the block before the timeout expires, it will stop the timeout; if not, the node will initiate a view change when the timeout expires.

  • The view change timeout expires - when a node starts a view change to view v, it will start a view change timeout. If the node is able to complete the view change before the timeout expires, it will stop the timeout; if not, it will initiate a new view change to view v + 1.

  • Multiple PrePrepare messages are received for the same view and sequence number but different blocks - this indicates the primary is faulty, since this behavior is invalid.

  • A Prepare message is received from the primary - this indicates the primary is faulty, since this behavior is invalid.

  • f + 1 ViewChange messages are received for the same view - this ensures that a node does not wait too long to start a view change; since only f nodes (at most) can be faulty at any given time, if more than f nodes decide to start a view change, other nodes can safely join them to perform that view change.

To start a view change, a node will do the following:

  1. Update its mode to ViewChanging(v), where v is the view the node is changing to

  2. Stop both the idle and and commit timeouts, since these are not needed again until after the view change

  3. Stop the view change timeout if it’s been started; it will be restarted with a new value later

  4. Broadcast a ViewChange message for the new view

ViewChange messages are accepted and added to the log if they satisfy these criteria:

  • They are for a later view than the node’s current view

  • If the node is in the mode ViewChanging(v), the view in the message must be greater than or equal to v

Once a node has received 2f + 1 ViewChange messages for the new view, it will start its view change timeout; this timeout ensures that the new primary starts the new view in a timely manner. The duration of the timeout is calculated based on a base duration (determined by the sawtooth.consensus.pbft.view_change_duration setting) using the formula: (DesiredViewNumber - CurrentViewNumber) * ViewChangeDuration.

When the primary for the new view receives 2f + 1 ViewChange messages, it will broadcast a NewView message to the network, signifying that the view change is complete. As a proof that this view change is valid, the primary will include 2f + 1 signed ViewChange messages from other nodes in the NewView message (the primary’s own “vote” is implicit), which will be validated by the other nodes.

If a node receives the new primary’s valid NewView message before its view change timeout expires, it will:

  1. Stop the view change timeout

  2. Update its view to match the new value

  3. Revert back to Normal mode

However, if a node’s view change timeout expires before it receives a NewView, it will stop the timeout and initiate a brand new view change for view v + 1 (where v is the view it was attempting to change to before).