rfc-103-incoming-txs-when-catching-up.md

RFC 103: Incoming transactions when node is catching up

Changelog

• 2023-04-20: Initial version (@hvanz)
• 2023-05-02: Update following PR comments (@hvanz)
• 2023-05-04: Update following more PR comments (@hvanz)

Abstract

Once a node starts running and is initialized, it can receive transactions from clients via RPC endpoints or from other nodes via P2P channels in its mempool reactor. The node then validates these transactions using CheckTx calls to the application. This process occurs even when the node is catching up to reach the latest block heights through state sync or block sync, and the application may not be up to date yet with the latest height of the blockchain. If a transaction's validity depends on the application state, it will likely be rejected by the application when received at this stage.

This document aims to answer the following questions:

• How should a node handle incoming transactions while catching up? Should it reject all transactions until it switches to consensus mode, or should it call CheckTx even though the application will likely deem the transaction invalid?
• Should nodes or clients send transactions when they suspect the receiving node is not initialized?

Background

Life cycle of a node

A CometBFT node transitions through the following stages: set up, initialization, catch up, and consensus.

When a node starts executing (for instance, from the CLI with the start command), it first reads its configuration files and sets up all its components, including reactors, stores, P2P connections, and RPC services. Then, the node starts running and turns on all the components. At this point we say that the node is initialized. The mempool reactor is running and the RPC endpoints are open, allowing the node to receive and process transactions from other nodes and clients.

The node could start the consensus protocol immediately to reconstruct the block history, but this could take a very long time. State sync and block sync are much faster mechanisms to bootstrap the node and reach the latest heights of the chain. While a node is performing state sync or block sync, it is in catch up mode. Once synchronization finishes, the node is not necessarily at the latest height, but it is assumed to be fairly close to it, so the node is ready to fully participate in consensus. At this point, it switches to consensus mode, which should never leave unless it is stopped or it crashes.

Transaction validation with CheckTx

The CheckTx method acts as a gatekeeper to the mempool, which only accepts valid transactions. The validity of a transaction is defined deterministically by the application, but it may depend on the application state, which may change at every height.

CheckTx is called:

• from the RPC endpoints broadcast_tx_async, broadcast_tx_sync, and broadcast_tx_commit, and
• when the node receives a Txs message from a peer: for each transaction in Txs there will be a call to CheckTx.

CheckTx puts the transaction in the cache, sends a RequestCheckTx ABCI message to the application, and sets a callback to process a ResponseCheckTx. If the application deems the transaction as valid, CheckTx stores the transaction in the mempool and notifies the consensus reactor that there are transactions available to be put in a new block. Otherwise it does not add the transaction to the mempool and removes it from the cache.

Mempool cache

The mempool reactor implements a cache of received transactions that it uses to discard already-seen transactions, thus avoiding processing duplicates.

Transactions remain in the cache until the application replies that a transaction is invalid. This can occur in three cases.

1. The first time a transaction is received and then rejected by CheckTx.
2. When the block executor updates the mempool, right after finalizing and committing a block: if there was an error while executing a transaction against the application, then it is removed from the cache.
3. When all the transactions in the mempool need to be rechecked after a new block has been delivered to the application. Each transaction will be validated again with CheckTx and removed from the cache if deemed invalid.

In the first and third cases, the reply from the application is a ResponseCheckTx message. In the second case, the reply is a ResponseFinalizeBlock.

All these cases are for removing invalid transactions, so the cache also requires that keep-invalid-txs-in-cache is set to false. The configuration keep-invalid-txs-in-cache is for ignoring invalid, already-seen transactions. This is useful for applications that consider rejected transactions will never be accepted again.

When a transaction is received, cached, and deemed valid, but the mempool is full, it is not kept in the cache. This is done to provide an opportunity for the transaction to be re-validated if it is received again later, potentially when there is sufficient free space in the mempool.

Discussion

Inbound messages

A node that is not in consensus mode, should accept or reject incoming transactions? The mempool reactor currently accepts and validates all incoming transactions. It only implements the cache to discard already-seen transactions.

There are not many reasons at this stage to have an operational mempool reactor and the broadcast_tx_* endpoints open. One reason is to store transactions to be able to validate them and have them ready once the node switches to consensus mode (in particular, the node could be the next validator). Another reason is for participating actively in the gossip protocol for propagating transactions. In any case, the application will most probably reject transactions if its validity depends on the application state. And the reactor only propagates transactions that are validated and stored in the mempool.

Decision Nodes that are catching up cannot guarantee effective handling of incoming transactions. They are not yet prepared to participate in transaction propagation properly. Therefore we see no strong reason at this stage for keeping the mempool reactor operational and the broadcast_tx_* endpoints open. Instead, we believe that nodes should reject incoming transactions until they switch to consensus mode. This approach will provide clearer semantics for the RPC endpoints, thus improving the user experience. Please refer to issue #785, which will address this decision.

Outbound messages

We cannot control or predict what a client may send to the RPC endpoints, but we can dictate what honest nodes must send to their peers.

The gossip algorithm currently implemented is a naive push protocol: a node will forward all the transactions in its mempool to all its peers, except to the nodes from which it received the transaction.

The protocol implements only one minor optimization. For each peer, the broadcastTxRoutine function in the mempool reactor iterates through all transactions in the mempool and sends them one by one to a peer. Just before sending a transaction, if the node suspects that the peer is lagging behind, it waits some time before checking again if the node has caught up. This is the code:

if peerState.GetHeight() < memTx.Height()-1 {
  time.Sleep(PeerCatchupSleepIntervalMS * time.Millisecond)
  continue
}

where:

• peerState is the local state of the peer (updated with the information received by the consensus reactor in PeerState.ApplyNewRoundStepMessage messages),
• memTx.Height() is the height at which transaction memTx.tx was validated (set during the handling of CheckTx responses), and
• PeerCatchupSleepIntervalMS is fixed to 100.

For a historical reference, this optimization has been around since 2015. There is no documentation or comment on the reason it was introduced. A first version with only the condition to check the peer's height was introduced in a commit from September 2015. In December 2015 a huge refactoring commit introduced the line that makes the broadcast routine sleep.

Note that the evidence reactor implements a similar optimization.

Should the code for this optimization be removed from the mempool reactor to make it consistent with how the RPC endpoints are implemented, that is, without filtering any message? Experimental evidence shows that actually the optimization works and improves the behavior of the nodes. The following picture shows the results of an experiment with four interconnected nodes. On the left we see the collected metrics when we run the nodes without the optimization. On the right we see the results of running the nodes with the optimization, that is, without modifying the code. The node in orange called validator04 joins the network at arount height 100 and starts performing block sync. In the graph at the bottom we can see the height of all nodes and in particular how this node starts from height 0 and catches up with the other nodes. Also we can observe that, when the optimization is disabled (left side), while the orange node is catching up, both its mempool size (top graph) and the number of rejected transactions (middle graph) increases significantly compared to the optimizated code (right side).

Decision The results presented above indicate that the optimization is effectively improving the system's performance and should be kept for now. In particular, the decrease in mempool size implies that the memory usage of the catching-up node is lower compared to those with the unoptimized code. Similarly, the decrease in the recheck rate suggests that CPU usage is also lower.

References

Mempool Reactor

1. Receive
2. broadcastTxRoutine
3. Outbound optimization

CListMempool

1. CheckTx
2. Update
3. ReapMaxBytesMaxGas

RPC endpoints

1. broadcast_tx_*