ResilientDB - PBFT Commit Algorithm

Allow this article to be a comprehensive introduction into the Practical Byzantine Fault Tolerance Consensus Protocol (PBFT) through the lens of the ResilientDB codebase. We will detail the sequenc...

Allow this article to be a comprehensive introduction into the Practical Byzantine Fault Tolerance Consensus Protocol (PBFT) through the lens of the ResilientDB codebase. We will detail the sequences and algorithms which encompass the protocol and encourage the readers to explore the rigorous math proofs afterwards.

The true power of this protocol becomes evident quite quick; it enables a democratic (i.e. through consensus) and decentralized (i.e. through replication) computational model!

Blockchain Consensus

The aim of a BFT consensus protocol for the application of blockchain is to ensure that all non-faulty replicas, that is replicas acting in bad faith, have the same lineariazability guarantee; where they assign the same order for a given client transaction1. This is a tolerance more resilient than achieving consensus in the presence of benign crash-faults, which Paxos2 and Viewstamped Replication3 can only sustain.

Since any BFT protocol employs cryptographic constructs to prevent malicious replicas from impersonating non-faulty replicas via validity of message senders, we will assume the use of those same constructs for authenticated communication in PBFT.

Practical Byzantine Fault Tolerance Consensus

Figure 1.

PBFT (Castro and Liskov 1999) follows a primary-replica model, as is the case in many traditional consensus protocols, where $\mathcal{P}$ (the primary replica) recieves transactional requests $\langle \tau \rangle_{\mathcal{C}}$ from $\mathcal{C}$ (clients) and through initiating consensus, coordinates the replication of transactions in all the nodes $\mathcal{R} \in R$ (replicas).

PBFT guarantees consensus in three phases if:

1. the network is reliable
2. $\mathcal{P}$ is not byzantine,

Throughout the normal case of PBFT, Replicas will exchange state information, necessary not only for the replication described above, but also for the progression of the system (livelihood) in the detection of byzantine behavior or faults that prevent successful completion of consensus rounds.

Some Notation

Keep in mind that PBFT operates in a Views ($v = 1, 2, 3, \dots , n$) meaning that we view the system through the lens of a single replica $\mathcal{P}$, named the Primary, with an identity $id(\mathcal{P}) = v {\rm mod}(n)$. The View and Primary are closely related. Next, each view contains many rounds $\rho$ of consensus.

• $n = \mathbf{card}(\mathcal{R})$ which is the number of replicas in the system.

• $g = \mathbf{card}(\mathcal{G})$ which is the number of Good replicas in the system.

• $f = \mathbf{card}(\mathcal{F})$ which is the number of Byzantine replicas in the system.

Therefore we can see that $g + f = n$. For Consensus to succeed at most $f$ replicas can act as byzantine where $n \geq 3 f + 1$ if you are a pessimist, or $g \geq \frac{3}{4} n$ if it helps to look at it from a different perspective.

Byzantine Commit Algorithm of PBFT

Client Submitting Transactions

In the normal case of PBFT we begin when:

A client $\mathcal{C}$ transmits a signed transactional request $\langle \tau \rangle_{\mathcal{C}}$ to the primary $\mathcal{R}$ in current view $v$.

ResilientDB Client Submitting Transactions
graph LR;
A(Client)
B(Primary)
A -- ClientQueryBatch --> B



In ResilientDB, transactions are batched to increase system throughput. This process begins in the client executable .runcl; where g_client_thread_cnt amount of client worker threads are spawned in client/client_main.cpp, via POSIX system calls and a spin their ClientThread::run() member function. Together, they generate a synthetic workload for scientific experiments; concurrently aggregating transactions into repeated ClientQueryBatch messages, each representing one large transactional request.

Pre-Prepare Phase

Subsequently, $\mathcal{P}$ checks on the condition that the client signature is valid in $\langle \tau \rangle_{\mathcal{C}}$, upon which it broadcasts a PRE-PREPARE message (defined as $m := PrePrepare(\langle \tau \rangle_{\mathcal{C}}, v, \rho)$) to all the replicas $\mathcal{R} \in R$.

A PRE-PREPARE message contains the following:

• Hash of the client request $\langle \tau \rangle_{\mathcal{C}}$ for optimizing network communication.

• Sequence number $\rho$ associated with the order of the transaction requested.

This “proposal” initiates the two-phase Byzantine commit algorithm in each replica $\mathcal{R}$ (Prepare + Commit Phases).

ResilientDB Pre-Prepare Phase
graph LR;
B(Primary)
C(Replicas)
D{"validate()"}
B ==> D
D -- BatchRequests --> C


In ResilientDB, replicas perform consensus in the executable rundb:

There are threads spawned in system/main.cpp, similarly spinning their WorkerThread::run() member function (defined in system/worker_thread.cpp) until the simulation is completed. These worker threads await messages via the work queue (work_queue.dequeue()) and perform specific tasks at each corresponding message in WorkerThread::process().

1. When a message of type ClientQueryBatch arrives, the primary replica will process it via WorkerThread::process_client_batch() (defined in system/worker_thread_pbft.cpp).
• This validates the client signature via the message’s validate() virtual method.
2. The primary will now propose transactions to all the replicas through broadcasting its BatchRequests message, an aggregation of transactions, to all the other replicas representing initiation of the consensus protocol.

Prepare Phase

This is the phase where replicas “prepare” the proposal of the primary replica.

A replica $\mathcal{R}$ enters the prepare phase upon recieving a PRE-PREPARE message $m$ and will support the “proposal” of $\mathcal{P}$ with broadcasting a PREPARE (defined as $Prepare(m)$) to all $\mathcal{R} \in R$ once the following criteria are met:

1. checks if the PRE-PREPARE was actually sent by $\mathcal{P}$

2. verification of the signature of $\mathcal{C}$ on $m$.

3. ensures that the sequence number $\rho$ has not been previously used.

After sending off a PREPARE message to all replicas, a given $\mathcal{R}$ awaits reassurance that a sufficient number of distinct replicas have agreed on $\rho$, until it has recieved $2f$ distinct and identical PREPARE messages. Once this is satisfied, $\mathcal{R}$ has now entered a prepared state for this request, preparing a certificate of $m$ (defined as $Prepare_{\mathcal{R}}(m)$ ) consisting of $m$ and the set of distict PREPARE messages required to achieve this state.

This certificate becomes hugely important in the later section when we discuss the Primary Replacement Algorithm. Remember, $Prepare_{\mathcal{R}}(m)$ is certification of a single replica $\mathcal{R}$’s knowledge that a sufficient number of nodes have conceded to agree to the same proposal. A replica changing to this state is significant.

ResilientDB Prepare Phase
graph TD;
B(Replica)
C(Other Replicas)
F["TxnManager::set_primarybatch"]
B ==> D
D -- PBFTPrepMessage --> C
D == deepCopy ==> E
D == deepCopy ==> F


Upon a Replica recieving a BatchRequests message, it immediately processes it thereafter in the WorkerThread::process_batch function.

• The non-primary replicas first perform a deep copy of a the whole BatchRequests messages in both WorkerThread::store_batch_msg and TxnManager::set_primarybatch (defined in system/txn.cpp).

• Replicas immediately enqueue the PBFTPrepMessage messages to their message queue via the TxnManager::send_pbft_prep_msgs() routine, broadcasting to all nodes in system.

• PBFTPrepMessage corresponds to $Prepare(m)$ message.
graph LR;
C(Replica)
D{"prep_rsp_cnt == 0 ?"}
G["process_pbft_prep_msg()"]
E[OK]
F(Achieved Prepared State)
H["TxnManager::decr_prep_rsp_cnt()"]
C ==> G
D == no ==> H
G --- D
D -- yes --> E
E --> F

• Upon recieving a PREPARE message from another replica, WorkerThread::process_pbft_prep_msg() is invoked, deincrementing a count TxnManager::decr_prep_rsp_cnt() in WorkerThread::prepared until prep_rsp_cnt in the transaction manager reaches 0, signalling that this replica has achieved a prepared state.

Commit Phase

This is the phase where replicas “commit” to the proposal of the primary replica.

Once a replica $\mathcal{R}$ achieves a prepared state it immediately sends off a COMMIT message (defined as $Commit(m)$) to all replicas. A given $\mathcal{R}$ awaits reassurance, until it has recieved $2f + 1$ identical and distict COMMIT messages, that a sufficient number of replicas - have knowledge that others - have agreed on $\rho$. After writing a certificate of commitment $Commit_{\mathcal{R}}(m)$, consisting of $m$ and the set of distict COMMIT messages, then $\mathcal{R}$ has now achieved a committed state for this request.

ResilientDB Commit Phase
graph LR;
A(Replica)
B(Other Replicas)
A -- PBFTCommitMessage --> B



Once again in WorkerThread::process_pbft_prep_msg, we see once a replica is sufficiently prepared (i.e. TxnManager::is_prepared()), it may now broadcasts PBFTCommitMessage messages via TxnManager::send_pbft_commit_msgs().

graph TD;
B(Replica)
C(Other Replicas)
E["validate_msg()"]
B ==> D
D -- PBFTPrepMessage --> C
D == validate ==> E
D == deepCopy ==> F


Upon recieving a COMMIT message from another replica, WorkerThread::process_pbft_commit_msg() is invoked.

• Validating the signature and copying the COMMIT message via TxnManager::add_commit_msg().
graph LR;
C(Replica)
D{"commit_rsp_cnt == 0 ?"}
G["committed_local()"]
E[OK]
F(Achieved Committed State)
H["TxnManager::decr_commit_rsp_cnt()"]
C ==> G
D == no ==> H
G --- D
D -- yes --> E
E --> F

• deincrementing a count TxnManager::decr_commit_rsp_cnt() in WorkerThread::committed_local() until commit_rsp_cnt in the transaction manager reaches 0, signalling that this replica has achieved a committed state.

Execution

Finally:

• Locally, $\mathcal{R}$ schedules and executes $\tau$ as the $\rho$-th transactions, appending to its ledger $\mathcal{L}_{ \mathcal{R} }$ .

• Subsequently $\mathcal{R}$ sends a message (defined as $r := Inform(\langle \tau \rangle_{\mathcal{C}}, v, \rho)$ ) to $\mathcal{C}$, indicating the result of this operation.

A client considers its request completed once it has recieved $f + 1$ distinct $Inform(m)$ responses.

When a replica is in a committed state, it enqueues a ExecuteMessage MESSAGE.

ResilientDB Execute Phase

Later we see the execute thread pick this message up in WorkerThread::process_execute_msg and enlists a transaction manager for each transaction in a batch and performs the run_txn() routine. For a YCSB Query, this virtual routine is defined in YCSBTxnManager::run_txn and is as simple as looping each operation in the ycsb_request and calling the database API (db->Put(KEY, VALUE)) and when complete, capturing the stats associated with how long this took.

References

1. Gupta S, Hellings J ,Sadoghi M. (2021). Fault-Tolerant Distributed Transactions on Blockchain, Synthesis Lectures on Data Management, February 2021, Vol. 16, No. 1 , Pages 1-268 (https://doi.org/10.2200/S01068ED1V01Y202012DTM065).

2. Lamport L (1998) The Part-time Parliament. ACM Trans Comput Syst 16(2):133–169

3. Oki BM, Liskov BH (1988) Viewstamped Replication: A New Primary Copy Method to Support Highly-Available Distributed Systems. In: Proceedings of the Seventh Annual ACM Symposium on Principles of Distributed Computing, ACM, New York, NY, USA, PODC ’88, pp 8–17