Skip to content
Branch: master
Find file Copy path
Find file Copy path
1 contributor

Users who have contributed to this file

629 lines (499 sloc) 28.4 KB


This RFC describes a practical Byzantine fault tolerant (PBFT) consensus algorithm for Hyperledger Sawtooth. The algorithm uses the Consensus API described in a pending Sawtooth RFC, and is adapted for blockchain use from a 1999 paper by Miguel Castro and Barbara Liskov [1].


The motivation for this RFC is to add a new, voting-based consensus mechanism with Byzantine fault tolerance to the capabilities of Hyperledger Sawtooth. This PBFT consensus algorithm ensures safety and liveness of a network, provided at most floor((n - 1)/3) nodes are faulty, where n is the total number of nodes in the network [1]. The PBFT algorithm is also inherently crash fault tolerant. Another advantage of PBFT is that blocks committed by nodes are final, so there are no forks in the network. This is verified by a "Consensus Seal," which is appended to blocks when they are finalized and checked upon receipt of a new block. In the future, this base PBFT algorithm could be extended to other PBFT-style algorithms, which decrease the number of necessary nodes in the network and reduce the total number of inter-node communication steps required.

Guide-level explanation

The Byzantine Generals Problem

Malicious users and code bugs can cause nodes in a network to exhibit arbitrary (Byzantine) behavior [1]. Byzantine behavior can be described by nodes that send conflicting information to other nodes in the network [2].

Example: Consider a scenario where there is a binary decision to be made: "to be" or "not to be." Say that there are nine nodes in a network. Four nodes send a "to be" message to all the other nodes, and four send "not to be" to all other nodes. However, the last node is faulty and sends "to be" to half of the nodes, and "not to be" to the other half. Because of this faulty node, the network ends up in an existential crisis. The goal of a Byzantine fault tolerant system is to resolve crises like this.

Practical Byzantine Fault Tolerance

Algorithms that attempt to conquer the aforementioned Byzantine faults are called Byzantine fault tolerant algorithms. Such algorithms include: Zyzzyva, RBFT, MinBFT, and PBFT. This RFC focuses on the base PBFT algorithm.

This consensus algorithm is voting-based, meaning [3]:

  • Only a single node (the primary) can commit blocks to the chain at any given time
  • One or more nodes in the network maintains a global view of the network
  • Adding and removing nodes from the network is difficult
  • There are many peer-to-peer messages passed in between nodes which specifically relate to consensus (see Message Types)

PBFT can be thought of in terms of state machine replication. The generic (non-blockchain-specific) algorithm works as follows:

  1. A client sends a message (request) to all the nodes in the network
  2. A series of messages is sent between nodes to determine if the request is valid, or has been tampered with.
  3. Once a number of nodes agree that the request is valid, then the instructions (operations) in the request are executed, and a result (reply) is returned to the client.
  4. The client waits for a number of replies that match, then accepts the result.

This generic algorithm has been modified and extended for blockchain use; see Normal Case Operation for a more detailed step-by-step walkthrough of this.


Term Definition
Node Machine running all the components necessary for a working blockchain (including the Validator, the REST API, at least one transaction processor, and the PBFT algorithm itself). In this RFC, unless otherwise specified, it can be assumed that Node refers to the PBFT component of the machine.
Server Synonym for node.
Replica Synonym for node.
Validator Component of a node responsible for interactions with the blockchain. Interactions with the validator are abstracted by the Consensus API.
Block A part of the blockchain, containing some operations and a link to the previous block.
Primary Node in charge of making the final consensus decisions and committing to the blockchain. Additionally is responsible for publishing the blocks given to it by the Consensus API, and starting the consensus process.
Secondary Auxiliary node used for consensus.
Client Machine that sends requests to and receives replies from the network of nodes. PBFT has no direct interaction with clients; the Validator bundles all client requests into blocks and sends them through the Consensus API to the consensus algorithm.
Checkpoint Point in time when logs can get garbage collected.
Checkpoint period How many client requests in between each checkpoint.
Block duration How many seconds to wait in between the creation of each block.
Message Block, with additional information (see Data Structures).
Working block The block that has been initialized but not finalized, and is currently being committed to.
Low water mark The sequence number of the last stable checkpoint.
High water mark Low water mark plus the desired maximum size of nodes' message logs.
View The scope of PBFT when the current primary is in charge. The view changes when the primary is deemed faulty, as described in View Changes.
n The total number of nodes in the network.
f The maximum number of faulty nodes.
v The current view number.
p The primary server number; p = v mod n.

Reference-level explanation

The PBFT consensus algorithm is written in Rust, and implements the Engine trait described in the Consensus API [3].

Data Structures

Consensus Messages

These are the messages that will be sent from node to node which specifically relate to consensus. The content of all consensus-related messages are serialized using Protocol Buffers (Protobuf).

From the Consensus API [3]:

// A consensus-related message sent between peers
message ConsensusPeerMessage {
  // Interpretation is left to the consensus engine implementation
  string message_type = 1;

  // The opaque payload to send to other nodes
  bytes content = 2;

  // Used to identify the consensus engine that produced this message
  string name = 3;
  string version = 4;

Consensus messages sent by the PBFT algorithm will have one of the following types (contained in the message_type field of ConsensusPeerMessage):

  • PrePrepare
  • Prepare
  • Commit
  • Checkpoint
  • ViewChange

Message Types

By definition, nodes in a PBFT network need to send a significant number of messages to each other. Most messages have similar contents, shown by PbftMessage. Auxiliary messages related to view changes are also shown. Furthermore, PBFT uses some of the message types defined in the Consensus API (referred to as updates), such as blockchain-related updates like BlockNew and BlockCommit, and the system update Shutdown.

The following Protobuf-style definitions are used to represent all consensus-related messages in the PBFT system:

// PBFT-specific block information (don't need to keep sending the whole payload
// around the network)
message PbftBlock {
  bytes block_id = 1;

  bytes signer_id = 2;

  uint64 block_num = 3;

  bytes summary = 4;

// 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, Checkpoint)
message PbftMessage {
  // Message information
  PbftMessageInfo info = 1;

  // The actual message
  PbftBlock block = 2;
// View change message, for when a node suspects the primary node is faulty
message PbftViewChange {
  // Message information
  PbftMessageInfo info = 1;

  // Set of `2f + 1` Checkpoint messages, proving correctness of stable
  // Checkpoint mentioned in info's `seq_num`
  repeated PbftMessage checkpoint_messages = 2;

Algorithm Operation

The PBFT algorithm will operate within the framework described by the Consensus API [3]. The start method contains an event loop which handles all incoming messages, in the form of Updates. The most important form of Update to the functionality of PBFT is Update::PeerMessage, but other updates like BlockNew, BlockCommit, BlockValid, BlockInvalid, and Shutdown are considered.

Peer Messages

When a message arrives from a peer, it must be interrogated for its type, and then the system must create a corresponding language-specific object of that message type. This is made easier by the fact that all consensus messages are Protobuf-serialized. Generally, once a message is converted into an appropriate object, it needs to be checked for content to make sure everything is legitimate. Some of these checks are performed by the validator (such as checking that a message's signature is valid), but some (making sure messages match, and that there are the correct number of them) need to be handled by the PBFT consensus engine.

On-chain Settings

The following on-chain settings are configurable, using the settings transaction family, with their respective defaults:

  • sawtooth.consensus.pbft.peers (required): JSON-formatted string of {<public-key1>: <node-id1>, <public-key2>: <node-id2>} mappings. For instance, {"02f86c2e26cf89117b10db0cfeff3fa26139f0c03ee8d6e3a77a814fca8c6de215":0, "03b515f97b2597f82ff0d5004a066d6d4f8052c82540a0b52cafcf5a719c8b5552":1} defines a 2-node network. The public keys refer to the validator's public key, usually located in /etc/sawtooth/keys/

  • sawtooth.consensus.pbft.block_duration=200: How long to wait before trying to publish a block

  • sawtooth.consensus.pbft.checkpoint_period=100: How many Commit messages in between each checkpoint

  • sawtooth.consensus.pbft.view_change_timeout=4000: How long to wait before deeming a primary node faulty

  • sawtooth.consensus.pbft.message_timeout=100: How long to wait to receive a message. If no message is received in this time, the Consensus API returns with a RecvTimeoutError::Timeout

  • sawtooth.consensus.pbft.max_log_size=1000: The maximum number of messages that can be in the log

Node Information Storage

Every node will keep track of the following state information:

  • Its own id

  • Its current sequence number and view number

  • Whether it's a primary or secondary node

  • Which step of the algorithm it's on

  • Mode of operation (Normal, ViewChanging, Checkpointing)

  • The maximum number of faulty nodes allowed in the network

  • The block that it's currently working on

  • Log of every peer message that has been sent to it (used to determine if it has received enough matching messages to proceed to the next stage of the algorithm; can be garbage collected every so often).

  • List of its connected peers. This is provided at startup from on-chain settings specified by the user. The length of this peer list will be used to calculate f, the maximum number of faulty nodes this network can tolerate. Currently, only static networks will be supported (that is, there will be no adding or removal of peers).


In order to keep the algorithm explanation below concise, we'll define some predicates here.

  • prepared is true for the current node if the following messages are present in its log:

    • The original BlockNew message
    • A PrePrepare message matching the original message (in the current view)
    • 2f + 1 matching Prepare messages from different nodes that match PrePrepare message above (including its own)
  • committed is true if for the current node:

    • prepared is true
    • This node has accepted 2f + 1 Commit messages, including its own

Normal Case Operation

Explanation of Message Types

  • PrePrepare: Sent from primary node to all nodes in the network, notifying them that a new message (BlockNew) has been received from the validator.

  • Prepare: Broadcast from every node once a PrePrepare is received for the current working block; used as verification of the PrePrepare message, and to signify that the block is ready to be checked.

  • Commit: Broadcast from every node once a BlockValid update is received for the current working block; used to determine if there is consensus that nodes should indeed commit the block contained in the original message.

  • Checkpoint: Sent by any node that has completed checkpoint_period Commit messages.

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


States: PBFT follows a state-machine replication pattern, where these states are defined:

  • NotStarted: The algorithm has not been started yet. No BlockNew updates have been received. In this stage, Checkpointing will occur if checkpoint_period blocks have been committed to the chain. Ready to receive a BlockNew update for the next block.

  • PrePreparing: A BlockNew has been received through the Consensus API, and its Consensus Seal has been verified. Ready to receive a PrePrepare message for the block corresponding to the BlockNew message just received.

  • Preparing: A PrePrepare message has been received and is valid. Ready to receive Prepare messages corresponding to this PrePrepare.

  • Checking: The predicate prepared is true; meaning this node has a BlockNew, a PrePrepare, and 2f + 1 corresponding Prepare messages. Ready to receive a BlockValid update.

  • Committing: A BlockValid has been received. Ready to receive Commit messages.

  • Finished: The predicate committed is true and the block has been committed to the chain. Ready to receive a BlockCommit update.

These states may be interrupted at any time if the view change timer expires, forcing the node into ViewChanging mode.

State Transitions: The following state transitions are defined; listed with their causes:

  • NotStartedPrePreparing: Receive a BlockNew update for the next block.

  • PrePreparingPreparing: Receive a PrePrepare message corresponding to the BlockNew.

  • PreparingChecking: prepared predicate is true.

  • CheckingCommitting: Receive a BlockValid update corresponding to the current working block.

  • CommittingFinished: committed predicate is true.

  • FinishedNotStarted: Receive a BlockCommit update for the current working block.

The states, state transitions, and actions that the algorithm takes are represented in the following diagram:

PBFT states


At the beginning of the Engine's start method, some initial setup is required:

  • Create the node for processing messages
  • Establish timers and counters for checkpoint periods and block durations, which are loaded from the on-chain settings

Event Loop

In Normal mode (when the primary node is not faulty), the PBFT consensus algorithm operates as follows, inside the event loop of the start method:

  1. Receive a BlockNew message from the Consensus API, representative of several batched client requests. The primary node checks the legitimacy of the message and assigns this message a sequence number, then broadcasts a PrePrepare message to all nodes. Legitimacy is checked by looking at the signer_id of the block in the BlockNew message, and making sure the previous_id is valid as the current chain head. The Consensus Seal is checked here as well, and all nodes tentatively update their working blocks. Secondary nodes ignore BlockNew messages; only append them to their logs.

  2. Receive PrePrepare messages and check their legitimacy. PrePrepare messages are legitimate if:

  • signer_id and summary of block inside PrePrepare match the corresponding fields of the original BlockNew block &&

  • View in PrePrepare message corresponds to this server's current view &&

  • This message hasn't been accepted already with a different summary &&

  • Sequence number is within the sequential bounds of the log (low and high water marks)

    Once the PrePrepare is accepted:

    • If primary: double check message matches the BlockNew, then broadcast a Prepare message to all nodes.
    • If secondary: update its own sequence number from the message, then broadcast a Prepare message to all nodes.

    If the PrePrepare is determined to be invalid, then start a view change.

  1. Receive Prepare messages, and check them all against their associated PrePrepare message in this node's message log.

  2. Once the predicate prepared is true for this node, then call check_blocks() on the current working block. If an error occurs (ReceiveError or UnknownBlock), abort (call ignore_block()). Otherwise, wait for a response (BlockValid or BlockInvalid) from the validator. If BlockValid, then broadcast a Commit message to all other nodes. If BlockInvalid, then call ignore_block(), and start a view change.

  3. When the predicate committed is true for this node, then it should commit the block using commit_block(), and advance the chain head.

  4. When a BlockCommit update is received by the primary node, it will call initialize_block().

  5. If block duration has elapsed, then try to summarize_block() with the current working block. If the working block is not ready, (BlockNotReady or InvalidState occurs), then nothing happens. Otherwise, finalize_block() is called with a serialized summary of all the Commit messages this node has received (which functions as the Consensus Seal). This in turn sends out a BlockNew update to the network, starting the next cycle of the algorithm.

View Changes

Sometimes, the node currently in charge (the primary) becomes faulty. This could mean it is either malicious, or experiencing internal problems. In either of these cases, a view change is necessary. View changes are triggered by a timeout: When a secondary node receives a BlockNew message, a timer is started. If the secondary ends up receiving a Commit message, the timer is cancelled, and the algorithm proceeds as normal. If the timer expires, the primary node is considered faulty and a view change is initiated. This ensures Byzantine fault tolerance due to the fact that each step of the algorithm will not proceed to the next unless it receives a certain number of matching messages, and due to the fact that the Validator does not pass on any messages that have invalid signatures [3].

The view change process is as follows:

  1. Any node who discovers the primary as faulty (whose timer timed out) sends a ViewChange message to all nodes, containing the node's current sequence number, its current view, proof of the previous checkpoint, and pending messages that have happened since that previous checkpoint. The node enters ViewChanging mode.

  2. Once a server receives 2f + 1 ViewChange messages (including its own), it changes its own view to v + 1, and resumes Normal operation. The new primary node's ID is p = v mod n.

Garbage Collection

After each checkpoint period (around 100 successfully completed cycles of the algorithm), server log messages can possibly be garbage-collected. When each node reaches a checkpoint, it enters Checkpointing mode and sends out a Checkpoint message to all of the other servers, with that node's current state (described by a PbftBlock). When the current node has 2f + 1 matching Checkpoint messages from different servers, the checkpoint is considered stable and the logs can be garbage collected: All log entries with sequence number less than the one in the Checkpoint message are discarded, and all previous checkpoints are removed. The high and low water marks are updated to reflect the sequence number of the new stable checkpoint. Once garbage collection is complete, the node resumes Normal operation.

Dynamic Networking

Dynamic networks are partially supported in this implementation of PBFT. Due to a current limitation in the Sawtooth Validator, only dynamic connecting is supported, not dynamic disconnecting.

Adding a peer:

  • The first node on the network is the primary.
  • Every node starts out in Connecting mode; no updates besides PeerConnected and PeerDisconnected are processed. When each connection message is received, every node starts a ViewChange timer.
  • Once the primary node has enough peers to be Byzantine fault tolerant, then it broadcasts a tentative PbftNetworkChange message. Other nodes receive this message, and verify that the peers contained in the message are the same as the ones in its peer list. Once this is verified, the node broadcasts a final PbftNetworkChange with that peer list.
  • Upon receipt of 2f + 1 PbftNetworkChange messages, the node accepts the peer list contained in the message, and updates its view, and sequence number based on those in the message information. It stops the ViewChange timer, and enters Normal execution mode.
  • If this node's chain head is behind the chain head from the message, then it begins the process of committing all the blocks between its head and the primary's chain head. These blocks do not need to undergo consensus again; they're on the chain, and thus are final.

Removing a peer: (not currently supported by Sawtooth Validator)

  • Peer is removed from this node's local peer list
  • If this change makes the network no longer byzantine fault tolerant, then the node re-enters Connecting mode and waits for another connection. If the network is in the middle of a block consensus process, it is paused until the network has come to consensus on this network change.
  • Primary sends out a tentative PbftNetworkChange, and a similar process to Adding a peer is executed.

Message type:

// Represents the current state of the network, including the active nodes,
// and the current block.
message PbftNetworkChange {
  // The peers in this network configuration
  repeated bytes peers = 1;

  // The current block that the network is on - used to help nodes catch back
  // up to the current chain head after they've been offline
  PbftBlock head = 2;

  // Does this message represent a tentative configuration, or a final
  // configuration?
  bool tentative = 3;

  bytes signer_id = 4;

  PbftMessageInfo info = 5;


One possible downside of PBFT-style consensus algorithms is that in general, 3f + 1 nodes are needed on the network in order to preserve Byzantine fault tolerance, unless special considerations are taken to ensure the authenticity of the peer-to-peer messages and sequence counters [6]. PBFT also requires a large number of consensus-specific messages; the general PBFT algorithm requires five per published block, and this implementation requires three (because it does not directly interact with clients). Additionally, PBFT does not prevent nodes from "leaking" information to bad actors.

A drawback of this specific design of PBFT is that the consensus algorithm operates serially; there is only one working block. Likely in a production environment, this will be unacceptable. A queue of working blocks will be necessary, each keeping track of what stage of the algorithm it is at.

Perhaps the most significant drawback of this algorithm stems from a lack of information provided by the Consensus API. In order to be truly Byzantine fault tolerant, PBFT must have access to the ordering of batches inside of the block, not just the blocks themselves. Currently, the Consensus API does not provide any visibility about batches inside the blocks, so it can't be confirmed that all blocks indeed have the same batch list ordering. One possible solution to this would be to package the batch list inside the block. It is possible that the whole batch list is not necessary, perhaps a hash of it would be sufficient to ensure ordering of batches inside blocks provided by the Consensus API.

Rationale and alternatives

There are many other voting-type consensus algorithms available to anyone who wants to use them, such as those mentioned in Prior art. Many of these algorithms are based on the methods described in this RFC, and most strategies employed by this RFC apply to PBFT's derivations as well.

Prior art


  • The original PBFT paper [1]
  • Improvements on PBFT [4]
  • Speculative Byzantine Fault Tolerance [5]
  • MinBFT and MinZyzzyva [6]

Other consensus implementations

Unresolved questions

  • What should the block consensus payload be comprised of (when calling finalize_block())?
    • This is just the Consensus Seal, containing 2f + 1 Commit messages and proving that the block went through consensus
  • What is the best way to persistently store message logs? Do they need to be secure/encrypted?


[1] M. Castro and B. Liskov. Practical Byzantine Fault Tolerance, Operating Systems Design and Implementation, 1999.

[2] L. Lamport, R. Shostak, and M. Pease. The Byzantine Generals Problem. ACM Transactions on Programming Languages and Systems, 4(3), 1982.

[3] A. Ludvik. Consensus API. Sawtooth RFCs, March 2018.

[4] M. Castro and B. Liskov. Practical Byzantine Fault Tolerance and Proactive Recovery, ACM Transactions on Computer Systems, 20(4), 2002.

[5] R. Kotla, L. Alvisi et al. Zyzzyva: Speculative Byzantine Fault Tolerance. ACM SIGOPS Operating Systems Review, 41(6), 2007.

[6] G. Veronese et al. Efficient Byzantine Fault Tolerance. IEEE Transactions on Computers, 62(1), 2013.

You can’t perform that action at this time.