tx-lifecycle.md

Transaction Lifecycle

This document describes the lifecycle of a transaction from creation to committed state changes. Transaction definition is described in a different doc. The transaction will be referred to as Tx. {synopsis}

Pre-requisite Readings

• Anatomy of a Cosmos SDK Application {prereq}

Creation

Transaction Creation

One of the main application interfaces is the command-line interface. The transaction Tx can be created by the user inputting a command in the following format from the command-line, providing the type of transaction in [command], arguments in [args], and configurations such as gas prices in [flags]:

[appname] tx [command] [args] [flags]

This command will automatically create the transaction, sign it using the account's private key, and broadcast it to the specified peer node.

There are several required and optional flags for transaction creation. The --from flag specifies which account the transaction is originating from. For example, if the transaction is sending coins, the funds will be drawn from the specified from address.

Gas and Fees

Additionally, there are several flags users can use to indicate how much they are willing to pay in fees:

• --gas refers to how much gas, which represents computational resources, Tx consumes. Gas is dependent on the transaction and is not precisely calculated until execution, but can be estimated by providing auto as the value for --gas.
• --gas-adjustment (optional) can be used to scale gas up in order to avoid underestimating. For example, users can specify their gas adjustment as 1.5 to use 1.5 times the estimated gas.
• --gas-prices specifies how much the user is willing pay per unit of gas, which can be one or multiple denominations of tokens. For example, --gas-prices=0.025uatom, 0.025upho means the user is willing to pay 0.025uatom AND 0.025upho per unit of gas.
• --fees specifies how much in fees the user is willing to pay in total.
• --timeout-height specifies a block timeout height to prevent the tx from being committed past a certain height.

The ultimate value of the fees paid is equal to the gas multiplied by the gas prices. In other words, fees = ceil(gas * gasPrices). Thus, since fees can be calculated using gas prices and vice versa, the users specify only one of the two.

Later, validators decide whether or not to include the transaction in their block by comparing the given or calculated gas-prices to their local min-gas-prices. Tx will be rejected if its gas-prices is not high enough, so users are incentivized to pay more.

CLI Example

Users of application app can enter the following command into their CLI to generate a transaction to send 1000uatom from a senderAddress to a recipientAddress. It specifies how much gas they are willing to pay: an automatic estimate scaled up by 1.5 times, with a gas price of 0.025uatom per unit gas.

appd tx send <recipientAddress> 1000uatom --from <senderAddress> --gas auto --gas-adjustment 1.5 --gas-prices 0.025uatom

Other Transaction Creation Methods

The command-line is an easy way to interact with an application, but Tx can also be created using a gRPC or REST interface or some other entrypoint defined by the application developer. From the user's perspective, the interaction depends on the web interface or wallet they are using (e.g. creating Tx using Lunie.io and signing it with a Ledger Nano S).

Addition to Mempool

Each full-node (running Tendermint) that receives a Tx sends an ABCI message, CheckTx, to the application layer to check for validity, and receives an abci.ResponseCheckTx. If the Tx passes the checks, it is held in the nodes' Mempool, an in-memory pool of transactions unique to each node) pending inclusion in a block - honest nodes will discard Tx if it is found to be invalid. Prior to consensus, nodes continuously check incoming transactions and gossip them to their peers.

Types of Checks

The full-nodes perform stateless, then stateful checks on Tx during CheckTx, with the goal to identify and reject an invalid transaction as early on as possible to avoid wasted computation.

Stateless checks do not require nodes to access state - light clients or offline nodes can do them - and are thus less computationally expensive. Stateless checks include making sure addresses are not empty, enforcing nonnegative numbers, and other logic specified in the definitions.

Stateful checks validate transactions and messages based on a committed state. Examples include checking that the relevant values exist and are able to be transacted with, the address has sufficient funds, and the sender is authorized or has the correct ownership to transact. At any given moment, full-nodes typically have multiple versions of the application's internal state for different purposes. For example, nodes will execute state changes while in the process of verifying transactions, but still need a copy of the last committed state in order to answer queries - they should not respond using state with uncommitted changes.

In order to verify a Tx, full-nodes call CheckTx, which includes both stateless and stateful checks. Further validation happens later in the DeliverTx stage. CheckTx goes through several steps, beginning with decoding Tx.

Decoding

When Tx is received by the application from the underlying consensus engine (e.g. Tendermint), it is still in its encoded []byte form and needs to be unmarshaled in order to be processed. Then, the runTx function is called to run in runTxModeCheck mode, meaning the function will run all checks but exit before executing messages and writing state changes.

ValidateBasic

sdk.Msgs are extracted from Tx, and ValidateBasic, a method of the sdk.Msg interface implemented by the module developer, is run for each one. ValidateBasic should include basic stateless sanity checks. For example, if the message is to send coins from one address to another, ValidateBasic likely checks for nonempty addresses and a nonnegative coin amount, but does not require knowledge of state such as the account balance of an address.

AnteHandler

After the ValidateBasic checks, the AnteHandlers are run. Technically, they are optional, but in practice, they are very often present to perform signature verification, gas calculation, fee deduction and other core operations related to blockchain transactions.

A copy of the cached context is provided to the AnteHandler, which performs limited checks specified for the transaction type. Using a copy allows the AnteHandler to do stateful checks for Tx without modifying the last committed state, and revert back to the original if the execution fails.

For example, the auth module AnteHandler checks and increments sequence numbers, checks signatures and account numbers, and deducts fees from the first signer of the transaction - all state changes are made using the checkState.

Gas

The Context, which keeps a GasMeter that will track how much gas has been used during the execution of Tx, is initialized. The user-provided amount of gas for Tx is known as GasWanted. If GasConsumed, the amount of gas consumed so during execution, ever exceeds GasWanted, the execution will stop and the changes made to the cached copy of the state won't be committed. Otherwise, CheckTx sets GasUsed equal to GasConsumed and returns it in the result. After calculating the gas and fee values, validator-nodes check that the user-specified gas-prices is greater than their locally defined min-gas-prices.

Discard or Addition to Mempool

If at any point during CheckTx the Tx fails, it is discarded and the transaction lifecycle ends there. Otherwise, if it passes CheckTx successfully, the default protocol is to relay it to peer nodes and add it to the Mempool so that the Tx becomes a candidate to be included in the next block.

The mempool serves the purpose of keeping track of transactions seen by all full-nodes. Full-nodes keep a mempool cache of the last mempool.cache_size transactions they have seen, as a first line of defense to prevent replay attacks. Ideally, mempool.cache_size is large enough to encompass all of the transactions in the full mempool. If the the mempool cache is too small to keep track of all the transactions, CheckTx is responsible for identifying and rejecting replayed transactions.

Currently existing preventative measures include fees and a sequence (nonce) counter to distinguish replayed transactions from identical but valid ones. If an attacker tries to spam nodes with many copies of a Tx, full-nodes keeping a mempool cache will reject identical copies instead of running CheckTx on all of them. Even if the copies have incremented sequence numbers, attackers are disincentivized by the need to pay fees.

Validator nodes keep a mempool to prevent replay attacks, just as full-nodes do, but also use it as a pool of unconfirmed transactions in preparation of block inclusion. Note that even if a Tx passes all checks at this stage, it is still possible to be found invalid later on, because CheckTx does not fully validate the transaction (i.e. it does not actually execute the messages).

Inclusion in a Block

Consensus, the process through which validator nodes come to agreement on which transactions to accept, happens in rounds. Each round begins with a proposer creating a block of the most recent transactions and ends with validators, special full-nodes with voting power responsible for consensus, agreeing to accept the block or go with a nil block instead. Validator nodes execute the consensus algorithm, such as Tendermint BFT, confirming the transactions using ABCI requests to the application, in order to come to this agreement.

The first step of consensus is the block proposal. One proposer amongst the validators is chosen by the consensus algorithm to create and propose a block - in order for a Tx to be included, it must be in this proposer's mempool.

State Changes

The next step of consensus is to execute the transactions to fully validate them. All full-nodes that receive a block proposal from the correct proposer execute the transactions by calling the ABCI functions BeginBlock, DeliverTx for each transaction, and EndBlock. While each full-node runs everything locally, this process yields a single, unambiguous result, since the messages' state transitions are deterministic and transactions are explicitly ordered in the block proposal.

		-----------------------
		|Receive Block Proposal|
		-----------------------
		          |
			  v
		-----------------------
		| BeginBlock	      |
		-----------------------
		          |
			  v
		-----------------------
		| DeliverTx(tx0)      |
		| DeliverTx(tx1)      |
		| DeliverTx(tx2)      |
		| DeliverTx(tx3)      |
		|	.	      |
		|	.	      |
		|	.	      |
		-----------------------
		          |
			  v
		-----------------------
		| EndBlock	      |
		-----------------------
		          |
			  v
		-----------------------
		| Consensus	      |
		-----------------------
		          |
			  v
		-----------------------
		| Commit	      |
		-----------------------

DeliverTx

The DeliverTx ABCI function defined in BaseApp does the bulk of the state transitions: it is run for each transaction in the block in sequential order as committed to during consensus. Under the hood, DeliverTx is almost identical to CheckTx but calls the runTx function in deliver mode instead of check mode. Instead of using their checkState, full-nodes use deliverState:

• Decoding: Since DeliverTx is an ABCI call, Tx is received in the encoded []byte form. Nodes first unmarshal the transaction, using the TxConfig defined in the app, then call runTx in runTxModeDeliver, which is very similar to CheckTx but also executes and writes state changes.

• Checks: Full-nodes call validateBasicMsgs and the AnteHandler again. This second check happens because they may not have seen the same transactions during the addition to Mempool stage
  and a malicious proposer may have included invalid ones. One difference here is that the AnteHandler will not compare gas-prices to the node's min-gas-prices since that value is local to each node - differing values across nodes would yield nondeterministic results.

• MsgServiceRouter: While CheckTx would have exited, DeliverTx continues to run runMsgs to fully execute each Msg within the transaction. Since the transaction may have messages from different modules, BaseApp needs to know which module to find the appropriate handler. This is achieved using BaseApp's MsgServiceRouter so that it can be processed by the module's Protobuf Msg service. For LegacyMsg routing, the Route function is called via the module manager to retrieve the route name and find the legacy Handler within the module.

• Msg service: a Protobuf Msg service, a step up from AnteHandler, is responsible for executing each message in the Tx and causes state transitions to persist in deliverTxState.

• Gas: While a Tx is being delivered, a GasMeter is used to keep track of how much gas is being used; if execution completes, GasUsed is set and returned in the abci.ResponseDeliverTx. If execution halts because BlockGasMeter or GasMeter has run out or something else goes wrong, a deferred function at the end appropriately errors or panics.

If there are any failed state changes resulting from a Tx being invalid or GasMeter running out, the transaction processing terminates and any state changes are reverted. Invalid transactions in a block proposal cause validator nodes to reject the block and vote for a nil block instead.

Commit

The final step is for nodes to commit the block and state changes. Validator nodes perform the previous step of executing state transitions in order to validate the transactions, then sign the block to confirm it. Full nodes that are not validators do not participate in consensus - i.e. they cannot vote - but listen for votes to understand whether or not they should commit the state changes.

When they receive enough validator votes (2/3+ precommits weighted by voting power), full nodes commit to a new block to be added to the blockchain and finalize the state transitions in the application layer. A new state root is generated to serve as a merkle proof for the state transitions. Applications use the Commit ABCI method inherited from Baseapp; it syncs all the state transitions by writing the deliverState into the application's internal state. As soon as the state changes are committed, checkState start afresh from the most recently committed state and deliverState resets to nil in order to be consistent and reflect the changes.

Note that not all blocks have the same number of transactions and it is possible for consensus to result in a nil block or one with none at all. In a public blockchain network, it is also possible for validators to be byzantine, or malicious, which may prevent a Tx from being committed in the blockchain. Possible malicious behaviors include the proposer deciding to censor a Tx by excluding it from the block or a validator voting against the block.

At this point, the transaction lifecycle of a Tx is over: nodes have verified its validity, delivered it by executing its state changes, and committed those changes. The Tx itself, in []byte form, is stored in a block and appended to the blockchain.

Next {hide}

Learn about accounts {hide}