Skip to main content
Version: Next

ADR 60: ABCI 1.0 Integration (Phase I)

Changelog

  • 2022-08-10: Initial Draft (@alexanderbez, @tac0turtle)
  • Nov 12, 2022: Update PrepareProposal and ProcessProposal semantics per the initial implementation PR (@alexanderbez)

Status

ACCEPTED

Abstract

This ADR describes the initial adoption of ABCI 1.0, the next evolution of ABCI, within the Cosmos SDK. ABCI 1.0 aims to provide application developers with more flexibility and control over application and consensus semantics, e.g. in-application mempools, in-process oracles, and order-book style matching engines.

Context

Tendermint will release ABCI 1.0. Notably, at the time of this writing, Tendermint is releasing v0.37.0 which will include PrepareProposal and ProcessProposal.

The PrepareProposal ABCI method is concerned with a block proposer requesting the application to evaluate a series of transactions to be included in the next block, defined as a slice of TxRecord objects. The application can either accept, reject, or completely ignore some or all of these transactions. This is an important consideration to make as the application can essentially define and control its own mempool allowing it to define sophisticated transaction priority and filtering mechanisms, by completely ignoring the TxRecords Tendermint sends it, favoring its own transactions. This essentially means that the Tendermint mempool acts more like a gossip data structure.

The second ABCI method, ProcessProposal, is used to process the block proposer's proposal as defined by PrepareProposal. It is important to note the following with respect to ProcessProposal:

  • Execution of ProcessProposal must be deterministic.
  • There must be coherence between PrepareProposal and ProcessProposal. In other words, for any two correct processes p and q, if q's Tendermint calls RequestProcessProposal on up, q's Application returns ACCEPT in ResponseProcessProposal.

It is important to note that in ABCI 1.0 integration, the application is NOT responsible for locking semantics -- Tendermint will still be responsible for that. In the future, however, the application will be responsible for locking, which allows for parallel execution possibilities.

Decision

We will integrate ABCI 1.0, which will be introduced in Tendermint v0.37.0, in the next major release of the Cosmos SDK. We will integrate ABCI 1.0 methods on the BaseApp type. We describe the implementations of the two methods individually below.

Prior to describing the implementation of the two new methods, it is important to note that the existing ABCI methods, CheckTx, DeliverTx, etc, still exist and serve the same functions as they do now.

PrepareProposal

Prior to evaluating the decision for how to implement PrepareProposal, it is important to note that CheckTx will still be executed and will be responsible for evaluating transaction validity as it does now, with one very important additive distinction.

When executing transactions in CheckTx, the application will now add valid transactions, i.e. passing the AnteHandler, to its own mempool data structure. In order to provide a flexible approach to meet the varying needs of application developers, we will define both a mempool interface and a data structure utilizing Golang generics, allowing developers to focus only on transaction ordering. Developers requiring absolute full control can implement their own custom mempool implementation.

We define the general mempool interface as follows (subject to change):

type Mempool interface {
// Insert attempts to insert a Tx into the app-side mempool returning
// an error upon failure.
Insert(sdk.Context, sdk.Tx) error

// Select returns an Iterator over the app-side mempool. If txs are specified,
// then they shall be incorporated into the Iterator. The Iterator must
// closed by the caller.
Select(sdk.Context, [][]byte) Iterator

// CountTx returns the number of transactions currently in the mempool.
CountTx() int

// Remove attempts to remove a transaction from the mempool, returning an error
// upon failure.
Remove(sdk.Tx) error
}

// Iterator defines an app-side mempool iterator interface that is as minimal as
// possible. The order of iteration is determined by the app-side mempool
// implementation.
type Iterator interface {
// Next returns the next transaction from the mempool. If there are no more
// transactions, it returns nil.
Next() Iterator

// Tx returns the transaction at the current position of the iterator.
Tx() sdk.Tx
}

We will define an implementation of Mempool, defined by nonceMempool, that will cover most basic application use-cases. Namely, it will prioritize transactions by transaction sender, allowing for multiple transactions from the same sender.

The default app-side mempool implementation, nonceMempool, will operate on a single skip list data structure. Specifically, transactions with the lowest nonce globally are prioritized. Transactions with the same nonce are prioritized by sender address.

type nonceMempool struct {
txQueue *huandu.SkipList
}

Previous discussions1 have come to the agreement that Tendermint will perform a request to the application, via RequestPrepareProposal, with a certain amount of transactions reaped from Tendermint's local mempool. The exact amount of transactions reaped will be determined by a local operator configuration. This is referred to as the "one-shot approach" seen in discussions.

When Tendermint reaps transactions from the local mempool and sends them to the application via RequestPrepareProposal, the application will have to evaluate the transactions. Specifically, it will need to inform Tendermint if it should reject and or include each transaction. Note, the application can even replace transactions entirely with other transactions.

When evaluating transactions from RequestPrepareProposal, the application will ignore ALL transactions sent to it in the request and instead reap up to RequestPrepareProposal.max_tx_bytes from it's own mempool.

Since an application can technically insert or inject transactions on Insert during CheckTx execution, it is recommended that applications ensure transaction validity when reaping transactions during PrepareProposal. However, what validity exactly means is entirely determined by the application.

The Cosmos SDK will provide a default PrepareProposal implementation that simply select up to MaxBytes valid transactions.

However, applications can override this default implementation with their own implementation and set that on BaseApp via SetPrepareProposal.

ProcessProposal

The ProcessProposal ABCI method is relatively straightforward. It is responsible for ensuring validity of the proposed block containing transactions that were selected from the PrepareProposal step. However, how an application determines validity of a proposed block depends on the application and its varying use cases. For most applications, simply calling the AnteHandler chain would suffice, but there could easily be other applications that need more control over the validation process of the proposed block, such as ensuring txs are in a certain order or that certain transactions are included. While this theoretically could be achieved with a custom AnteHandler implementation, it's not the cleanest UX or the most efficient solution.

Instead, we will define an additional ABCI interface method on the existing Application interface, similar to the existing ABCI methods such as BeginBlock or EndBlock. This new interface method will be defined as follows:

ProcessProposal(sdk.Context, abci.ProcessProposalRequest) error {}

Note, we must call ProcessProposal with a new internal branched state on the Context argument as we cannot simply just use the existing checkState because BaseApp already has a modified checkState at this point. So when executing ProcessProposal, we create a similar branched state, processProposalState, off of deliverState. Note, the processProposalState is never committed and is completely discarded after ProcessProposal finishes execution.

The Cosmos SDK will provide a default implementation of ProcessProposal in which all transactions are validated using the CheckTx flow, i.e. the AnteHandler, and will always return ACCEPT unless any transaction cannot be decoded.

DeliverTx

Since transactions are not truly removed from the app-side mempool during PrepareProposal, since ProcessProposal can fail or take multiple rounds and we do not want to lose transactions, we need to finally remove the transaction from the app-side mempool during DeliverTx since during this phase, the transactions are being included in the proposed block.

Alternatively, we can keep the transactions as truly being removed during the reaping phase in PrepareProposal and add them back to the app-side mempool in case ProcessProposal fails.

Consequences

Backwards Compatibility

ABCI 1.0 is naturally not backwards compatible with prior versions of the Cosmos SDK and Tendermint. For example, an application that requests RequestPrepareProposal to the same application that does not speak ABCI 1.0 will naturally fail.

However, in the first phase of the integration, the existing ABCI methods as we know them today will still exist and function as they currently do.

Positive

  • Applications now have full control over transaction ordering and priority.
  • Lays the groundwork for the full integration of ABCI 1.0, which will unlock more app-side use cases around block construction and integration with the Tendermint consensus engine.

Negative

  • Requires that the "mempool", as a general data structure that collects and stores uncommitted transactions will be duplicated between both Tendermint and the Cosmos SDK.
  • Additional requests between Tendermint and the Cosmos SDK in the context of block execution. Albeit, the overhead should be negligible.
  • Not backwards compatible with previous versions of Tendermint and the Cosmos SDK.

Further Discussions

It is possible to design the app-side implementation of the Mempool[T MempoolTx] in many different ways using different data structures and implementations. All of which have different tradeoffs. The proposed solution keeps things simple and covers cases that would be required for most basic applications. There are tradeoffs that can be made to improve performance of reaping and inserting into the provided mempool implementation.

References