Skip to content

Latest commit

 

History

History
468 lines (380 loc) · 24.2 KB

abci++_basic_concepts.md

File metadata and controls

468 lines (380 loc) · 24.2 KB
Raw
order title
1
Overview and basic concepts

Outline

Overview and basic concepts

ABCI++ vs. ABCI

↑ Back to Outline

The Application's main role is to execute blocks decided (a.k.a. finalized) by consensus. The decided blocks are the main consensus's ouput to the (replicated) Application. With ABCI, the application only interacts with consensus at decision time. This restricted mode of interaction prevents numerous features for the Application, including many scalability improvements that are now better understood than when ABCI was first written. For example, many ideas proposed to improve scalability can be boiled down to "make the block proposers do work, so the network does not have to". This includes optimizations such as transaction level signature aggregation, state transition proofs, etc. Furthermore, many new security properties cannot be achieved in the current paradigm, as the Application cannot require validators to do more than executing the transactions contained in finalized blocks. This includes features such as threshold cryptography, and guaranteed IBC connection attempts.

ABCI++ addresses these limitations by allowing the application to intervene at three key places of consensus execution: (a) at the moment a new proposal is to be created, (b) at the moment a proposal is to be validated, and (c) at the moment a (precommit) vote is sent/received. The new interface allows block proposers to perform application-dependent work in a block through the PrepareProposal method (a); validators to perform application-dependent work and checks in a proposed block through the ProcessProposal method (b); and applications to require their validators do more than just validate blocks through the ExtendVote and VerifyVoteExtension methods (c). Furthermore, ABCI++ coalesces {BeginBlock, [DeliverTx], EndBlock} into FinalizeBlock, as a simplified, efficient way to deliver a decided block to the Application.

Method overview

↑ Back to Outline

Methods can be classified into four categories: consensus, mempool, info, and state-sync.

Consensus/block execution methods

The first time a new blockchain is started, Tendermint calls InitChain. From then on, method FinalizeBlock is executed upon the decision of each block, resulting in an updated Application state. During the execution of an instance of consensus, which decides the block for a given height, and before method FinalizeBlock is called, methods PrepareProposal, ProcessProposal, ExtendVote, and VerifyVoteExtension may be called several times. See Tendermint's expected behavior for details on the possible call sequences of these methods.

  • InitChain: This method initializes the blockchain. Tendermint calls it once upon genesis.

  • PrepareProposal: It allows the block proposer to perform application-dependent work in a block before proposing it. This enables, for instance, batch optimizations to a block, which has been empirically demonstrated to be a key component for improved performance. Method PrepareProposal is called every time Tendermint is about to broadcast a Proposal message, but no previous proposal has been locked at the Tendermint level. Tendermint gathers outstanding transactions from the mempool, generates a block header, and uses them to create a block to propose. Then, it calls RequestPrepareProposal with the newly created proposal, called raw proposal. The Application can make changes to the raw proposal, such as modifying transactions, and returns the (potentially) modified proposal, called prepared proposal in the ResponsePrepareProposal call. The logic modifying the raw proposal can be non-deterministic.

  • ProcessProposal: It allows a validator to perform application-dependent work in a proposed block. This enables features such as immediate block execution, and allows the Application to reject invalid blocks. Tendermint calls it when it receives a proposal and the Tendermint algorithm has not locked on a value. The Application cannot modify the proposal at this point but can reject it if it is invalid. If that is the case, Tendermint will prevote nil on the proposal, which has strong liveness implications for Tendermint. As a general rule, the Application SHOULD accept a prepared proposal passed via ProcessProposal, even if a part of the proposal is invalid (e.g., an invalid transaction); the Application can ignore the invalid part of the prepared proposal at block execution time.

  • ExtendVote: It allows applications to force their validators to do more than just validate within consensus. ExtendVote allows applications to include non-deterministic data, opaque to Tendermint, to precommit messages (the final round of voting). The data, called vote extension, will be broadcast and received together with the vote it is extending, and will be made available to the Application in the next height, in the rounds where the local process is the proposer. Tendermint calls ExtendVote when it is about to send a non-nil precommit message. If the Application does not have vote extension information to provide at that time, it returns a 0-length byte array as its vote extension.

  • VerifyVoteExtension: It allows validators to validate the vote extension data attached to a precommit message. If the validation fails, the whole precommit message will be deemed invalid and ignored by Tendermint. This has a negative impact on Tendermint's liveness, i.e., if vote extensions repeatedly cannot be verified by correct validators, Tendermint may not be able to finalize a block even if sufficiently many (+2/3) validators send precommit votes for that block. Thus, VerifyVoteExtension should be used with special care. As a general rule, an Application that detects an invalid vote extension SHOULD accept it in ResponseVerifyVoteExtension and ignore it in its own logic. Tendermint calls it when a process receives a precommit message with a (possibly empty) vote extension.

  • FinalizeBlock: It delivers a decided block to the Application. The Application must execute the transactions in the block deterministically and update its state accordingly. Cryptographic commitments to the block and transaction results, returned via the corresponding parameters in ResponseFinalizeBlock, are included in the header of the next block. Tendermint calls it when a new block is decided.

  • Commit: Instructs the Application to persist its state. It is a fundamental part of Tendermint's crash-recovery mechanism that ensures the synchronization between Tendermint and the Applicatin upon recovery. Tendermint calls it just after having persisted the data returned by ResponseFinalizeBlock. The Application can now discard any state or data except the one resulting from executing the transactions in the decided block.

Mempool methods

  • CheckTx: This method allows the Application to validate transactions. Validation can be stateless (e.g., checking signatures ) or stateful (e.g., account balances). The type of validation performed is up to the application. If a transaction passes the validation, then Tendermint adds it to the mempool; otherwise the transaction is discarded. Tendermint calls it when it receives a new transaction either coming from an external user (e.g., a client) or another node. Furthermore, Tendermint can be configured to call re-CheckTx on all outstanding transactions in the mempool after calling Commitfor a block.

Info methods

  • Info: Used to sync Tendermint with the Application during a handshake that happens upon recovery, or on startup when state-sync is used.

  • Query: This method can be used to query the Application for information about the application state.

State-sync methods

State sync allows new nodes to rapidly bootstrap by discovering, fetching, and applying state machine (application) snapshots instead of replaying historical blocks. For more details, see the state sync documentation.

New nodes discover and request snapshots from other nodes in the P2P network. A Tendermint node that receives a request for snapshots from a peer will call ListSnapshots on its Application. The Application returns the list of locally available snapshots. Note that the list does not contain the actual snapshots but metadata about them: height at which the snapshot was taken, application-specific verification data and more (see snapshot data type for more details). After receiving a list of available snapshots from a peer, the new node can offer any of the snapshots in the list to its local Application via the OfferSnapshot method. The Application can check at this point the validity of the snapshot metadata.

Snapshots may be quite large and are thus broken into smaller "chunks" that can be assembled into the whole snapshot. Once the Application accepts a snapshot and begins restoring it, Tendermint will fetch snapshot "chunks" from existing nodes. The node providing "chunks" will fetch them from its local Application using the LoadSnapshotChunk method.

As the new node receives "chunks" it will apply them sequentially to the local application with ApplySnapshotChunk. When all chunks have been applied, the Application's AppHash is retrieved via an Info query. To ensure that the sync proceeded correctly, Tendermint compares the local Application's AppHash to the AppHash stored on the blockchain (verified via light client verification).

In summary:

  • ListSnapshots: Used by nodes to discover available snapshots on peers.

  • OfferSnapshot: When a node receives a snapshot from a peer, Tendermint uses this method to offer the snapshot to the Application.

  • LoadSnapshotChunk: Used by Tendermint to retrieve snapshot chunks from the Application to send to peers.

  • ApplySnapshotChunk: Used by Tendermint to hand snapshot chunks to the Application.

Other methods

Additionally, there is a Flush method that is called on every connection, and an Echo method that is used for debugging.

More details on managing state across connections can be found in the section on Managing Application State.

Next-block execution vs. same-block execution

↑ Back to Outline

In the original ABCI protocol, the only moment when the Application had access to a block was after it was decided. This led to a block execution model, called next-block execution, where some fields hashed in a block header refer to the execution of the previous block, namely:

  • the Merkle root of the Application's state
  • the transaction results
  • the consensus parameter updates
  • the validator updates

With ABCI++, an Application may be configured to keep using the next-block execution model, by executing the decided block in FinalizeBlock. However, the new methods introduced — PrepareProposal and ProcessProposal — disclose the entire proposed block to the Application, allowing for its immediate exectution. An Application implementing immediate execution may additionally wish to store certain data resulting from the block's execution in the same block that has just been executed. This brings about a new execution model, called same-block execution. An Application implementing this execution model, upon receiving a raw proposal via RequestPrepareProposal and potentially modifying its transaction list, fully executes the resulting prepared proposal as though it was the decided block (immediate execution), and the results of the block execution are used as follows:

  • The block execution may generate a set of events. The Application should store these events and return them back to Tendermint during the FinalizeBlock call if the block is finally decided.
  • The Merkle root resulting from executing the prepared proposal is provided in ResponsePrepareProposal and thus refers to the current block. Tendermint will use it in the prepared proposal's header.
  • Likewise, the transaction results from executing the prepared proposal are provided in ResponsePrepareProposal and refer to the transactions in the current block. Tendermint will use them to calculate the results hash in the prepared proposal's header.
  • The consensus parameter updates and validator updates are also provided in ResponsePrepareProposal and reflect the result of the prepared proposal's execution. They come into force in height H+1 (as opposed to the H+2 rule in next-block execution model).

If the Application is configured to keep the next-block execution model, it will not provide any data in ResponsePrepareProposal, other than a potentially modified transaction list. The Application may nevertheless choose to perform immediate execution even in next-block execution mode, however same-block execution mode requires immediate execution.

The long term plan is for the execution model to be set in a new boolean parameter same_block in ConsensusParams. Once this parameter is introduced, it must not be changed once the blockchain has started, unless the Application developers really know what they are doing. However, modifying ConsensusParams structure cannot be done lightly if we are to preserve blockchain compatibility. Therefore we need an interim solution until soft upgrades are specified and implemented in Tendermint. This somewhat unsafe solution consists in Tendermint assuming same-block execution if the Application fills the above mentioned fields in ResponsePrepareProposal.

Tendermint proposal timeout

Immediate execution requires the Application to fully execute the prepared block before returning from PrepareProposal, this means that Tendermint cannot make progress during the block execution. This stands on Tendermint's critical path: if the Application takes a long time executing the block, the default value of TimeoutPropose might not be sufficient to accommodate the long block execution time and non-proposer nodes might time out and prevote nil. The proposal, in this case, will probably be rejected and a new round will be necessary.

The Application is the best suited to provide a value for TimeoutPropose so that the block execution time upon PrepareProposal fits well in the propose timeout interval. Thus, the Application can adapt the value of TimeoutPropose at every height via TimeoutParams.Propose, contained in ConsensusParams.

Deterministic State-Machine Replication

↑ Back to Outline

ABCI++ applications must implement deterministic finite-state machines to be securely replicated by the Tendermint consensus engine. This means block execution must be strictly deterministic: given the same ordered set of transactions, all nodes will compute identical responses, for all successive FinalizeBlock calls. This is critical because the responses are included in the header of the next block, either via a Merkle root or directly, so all nodes must agree on exactly what they are.

For this reason, it is recommended that application state is not exposed to any external user or process except via the ABCI connections to a consensus engine like Tendermint Core. The Application must only change its state based on input from block execution (FinalizeBlock calls), and not through any other kind of request. This is the only way to ensure all nodes see the same transactions and compute the same results.

Some Applications may choose to implement immediate execution, which entails executing the blocks that are about to be proposed (via PrepareProposal), and those that the Application is asked to validate (via ProcessProposal). However, the state changes caused by processing those proposed blocks must never replace the previous state until FinalizeBlock confirms the block decided.

Additionally, vote extensions or the validation thereof (via ExtendVote or VerifyVoteExtension) must never have side effects on the current state. They can only be used when their data is provided in a RequestPrepareProposal call.

If there is some non-determinism in the state machine, consensus will eventually fail as nodes disagree over the correct values for the block header. The non-determinism must be fixed and the nodes restarted.

Sources of non-determinism in applications may include:

  • Hardware failures
    • Cosmic rays, overheating, etc.
  • Node-dependent state
    • Random numbers
    • Time
  • Underspecification
    • Library version changes
    • Race conditions
    • Floating point numbers
    • JSON or protobuf serialization
    • Iterating through hash-tables/maps/dictionaries
  • External Sources
    • Filesystem
    • Network calls (eg. some external REST API service)

See #56 for the original discussion.

Note that some methods (Query, FinalizeBlock) return non-deterministic data in the form of Info and Log fields. The Log is intended for the literal output from the Application's logger, while the Info is any additional info that should be returned. These are the only fields that are not included in block header computations, so we don't need agreement on them. All other fields in the Response* must be strictly deterministic.

Events

↑ Back to Outline

Method FinalizeBlock includes an events field at the top level in its Response*, and one events field per transaction included in the block. Applications may respond to this ABCI++ method with an event list for each executed transaction, and a general event list for the block itself. Events allow applications to associate metadata with transactions and blocks. Events returned via FinalizeBlock do not impact Tendermint consensus in any way and instead exist to power subscriptions and queries of Tendermint state.

An Event contains a type and a list of EventAttributes, which are key-value string pairs denoting metadata about what happened during the method's (or transaction's) execution. Event values can be used to index transactions and blocks according to what happened during their execution.

Each event has a type which is meant to categorize the event for a particular Response* or Tx. A Response* or Tx may contain multiple events with duplicate type values, where each distinct entry is meant to categorize attributes for a particular event. Every key and value in an event's attributes must be UTF-8 encoded strings along with the event type itself.

message Event {
  string                  type       = 1;
  repeated EventAttribute attributes = 2;
}

The attributes of an Event consist of a key, a value, and an index flag. The index flag notifies the Tendermint indexer to index the attribute. The value of the index flag is non-deterministic and may vary across different nodes in the network.

message EventAttribute {
  bytes key   = 1;
  bytes value = 2;
  bool  index = 3;  // nondeterministic
}

Example:

 abci.ResponseFinalizeBlock{
  // ...
 Events: []abci.Event{
  {
   Type: "validator.provisions",
   Attributes: []abci.EventAttribute{
    abci.EventAttribute{Key: []byte("address"), Value: []byte("..."), Index: true},
    abci.EventAttribute{Key: []byte("amount"), Value: []byte("..."), Index: true},
    abci.EventAttribute{Key: []byte("balance"), Value: []byte("..."), Index: true},
   },
  },
  {
   Type: "validator.provisions",
   Attributes: []abci.EventAttribute{
    abci.EventAttribute{Key: []byte("address"), Value: []byte("..."), Index: true},
    abci.EventAttribute{Key: []byte("amount"), Value: []byte("..."), Index: false},
    abci.EventAttribute{Key: []byte("balance"), Value: []byte("..."), Index: false},
   },
  },
  {
   Type: "validator.slashed",
   Attributes: []abci.EventAttribute{
    abci.EventAttribute{Key: []byte("address"), Value: []byte("..."), Index: false},
    abci.EventAttribute{Key: []byte("amount"), Value: []byte("..."), Index: true},
    abci.EventAttribute{Key: []byte("reason"), Value: []byte("..."), Index: true},
   },
  },
  // ...
 },
}

Evidence

↑ Back to Outline

Tendermint's security model relies on the use of evidences of misbehavior. An evidence is an irrefutable proof of malicious behavior by a network participant. It is the responsibility of Tendermint to detect such malicious behavior. When malicious behavior is detected, Tendermint will gossip evidences of misbehavior to other nodes and commit the evidences to the chain once they are verified by a subset of validators. These evidences will then be passed on to the Application through ABCI++. It is the responsibility of the Application to handle evidence of misbehavior and exercise punishment.

There are two forms of evidence: Duplicate Vote and Light Client Attack. More information can be found in either data structures or accountability.

EvidenceType has the following protobuf format:

enum EvidenceType {
  UNKNOWN               = 0;
  DUPLICATE_VOTE        = 1;
  LIGHT_CLIENT_ATTACK   = 2;
}

Errors

↑ Back to Outline

The Query, and CheckTx methods include a Code field in their Response*. Field Code is meant to contain an application-specific response code. A response code of 0 indicates no error. Any other response code indicates to Tendermint that an error occurred.

These methods also return a Codespace string to Tendermint. This field is used to disambiguate Code values returned by different domains of the Application. The Codespace is a namespace for the Code.

Methods Echo, Info, and InitChain do not return errors. An error in any of these methods represents a critical issue that Tendermint has no reasonable way to handle. If there is an error in one of these methods, the Application must crash to ensure that the error is safely handled by an operator.

Method FinalizeBlock is a special case. It contains a number of Code and Codespace fields as part of type ExecTxResult. Each of these codes reports errors related to the transaction it is attached to. However, FinalizeBlock does not return errors at the top level, so the same considerations on critical issues made for Echo, Info, and InitChain also apply here.

The handling of non-zero response codes by Tendermint is described below.

CheckTx

When Tendermint receives a ResponseCheckTx with a non-zero Code, the associated transaction will not be added to Tendermint's mempool or it will be removed if it is already included.

ExecTxResult (as part of FinalizeBlock)

The ExecTxResult type delivers transaction results from the Application to Tendermint. When Tendermint receives a ResponseFinalizeBlock containing an ExecTxResult with a non-zero Code, the response code is logged. Past Code values can be queried by clients. As the transaction was part of a decided block, the Code does not influence Tendermint consensus.

Query

When Tendermint receives a ResponseQuery with a non-zero Code, this code is returned directly to the client that initiated the query.