tendermint/spec/abci/apps.md

Applications

Please ensure you've first read the spec for ABCI Methods and Types

Here we cover the following components of ABCI applications:

• Connection State - the interplay between ABCI connections and application state and the differences between CheckTx and DeliverTx.
• Transaction Results - rules around transaction results and validity
• Validator Set Updates - how validator sets are changed during InitChain and EndBlock
• Query - standards for using the Query method and proofs about the application state
• Crash Recovery - handshake protocol to synchronize Tendermint and the application on startup.
• State Sync - rapid bootstrapping of new nodes by restoring state machine snapshots

State

Since Tendermint maintains four concurrent ABCI connections, it is typical for an application to maintain a distinct state for each, and for the states to be synchronized during Commit.

Commit

Application state should only be persisted to disk during Commit.

Before Commit is called, Tendermint locks and flushes the mempool so that no new messages will be received on the mempool connection. This provides an opportunity to safely update all three states to the latest committed state at once.

When Commit completes, it unlocks the mempool.

WARNING: if the ABCI app logic processing the Commit message sends a /broadcast_tx_sync or /broadcast_tx_commit and waits for the response before proceeding, it will deadlock. Executing those broadcast_tx calls involves acquiring a lock that is held during the Commit call, so it's not possible. If you make the call to the broadcast_tx endpoints concurrently, that's no problem, it just can't be part of the sequential logic of the Commit function.

Consensus Connection

The Consensus Connection should maintain a DeliverTxState - the working state for block execution. It should be updated by the calls to BeginBlock, DeliverTx, and EndBlock during block execution and committed to disk as the "latest committed state" during Commit.

Updates made to the DeliverTxState by each method call must be readable by each subsequent method - ie. the updates are linearizable.

Mempool Connection

The Mempool Connection should maintain a CheckTxState to sequentially process pending transactions in the mempool that have not yet been committed. It should be initialized to the latest committed state at the end of every Commit.

The CheckTxState may be updated concurrently with the DeliverTxState, as messages may be sent concurrently on the Consensus and Mempool connections. However, before calling Commit, Tendermint will lock and flush the mempool connection, ensuring that all existing CheckTx are responded to and no new ones can begin.

After Commit, CheckTx is run again on all transactions that remain in the node's local mempool after filtering those included in the block. To prevent the mempool from rechecking all transactions every time a block is committed, set the configuration option mempool.recheck=false. As of Tendermint v0.32.1, an additional Type parameter is made available to the CheckTx function that indicates whether an incoming transaction is new (CheckTxType_New), or a recheck (CheckTxType_Recheck).

Finally, the mempool will unlock and new transactions can be processed through CheckTx again.

Note that CheckTx doesn't have to check everything that affects transaction validity; the expensive things can be skipped. In fact, CheckTx doesn't have to check anything; it might say that any transaction is a valid transaction. Unlike DeliverTx, CheckTx is just there as a sort of weak filter to keep invalid transactions out of the blockchain. It's weak, because a Byzantine node doesn't care about CheckTx; it can propose a block full of invalid transactions if it wants.

Info Connection

The Info Connection should maintain a QueryState for answering queries from the user, and for initialization when Tendermint first starts up (both described further below). It should always contain the latest committed state associated with the latest committed block.

QueryState should be set to the latest DeliverTxState at the end of every Commit, ie. after the full block has been processed and the state committed to disk. Otherwise it should never be modified.

Snapshot Connection

The Snapshot Connection is optional, and is only used to serve state sync snapshots for other nodes and/or restore state sync snapshots to a local node being bootstrapped.

Transaction Results

ResponseCheckTx and ResponseDeliverTx contain the same fields.

The Info and Log fields are non-deterministic values for debugging/convenience purposes that are otherwise ignored.

The Data field must be strictly deterministic, but can be arbitrary data.

Gas

Ethereum introduced the notion of gas as an abstract representation of the cost of resources used by nodes when processing transactions. Every operation in the Ethereum Virtual Machine uses some amount of gas, and gas can be accepted at a market-variable price. Users propose a maximum amount of gas for their transaction; if the tx uses less, they get the difference credited back. Tendermint adopts a similar abstraction, though uses it only optionally and weakly, allowing applications to define their own sense of the cost of execution.

In Tendermint, the ConsensusParams.Block.MaxGas limits the amount of gas that can be used in a block. The default value is -1, meaning no limit, or that the concept of gas is meaningless.

Responses contain a GasWanted and GasUsed field. The former is the maximum amount of gas the sender of a tx is willing to use, and the later is how much it actually used. Applications should enforce that GasUsed <= GasWanted - ie. tx execution should halt before it can use more resources than it requested.

When MaxGas > -1, Tendermint enforces the following rules:

• GasWanted <= MaxGas for all txs in the mempool
• (sum of GasWanted in a block) <= MaxGas when proposing a block

If MaxGas == -1, no rules about gas are enforced.

Note that Tendermint does not currently enforce anything about Gas in the consensus, only the mempool. This means it does not guarantee that committed blocks satisfy these rules! It is the application's responsibility to return non-zero response codes when gas limits are exceeded.

The GasUsed field is ignored completely by Tendermint. That said, applications should enforce:

• GasUsed <= GasWanted for any given transaction
• (sum of GasUsed in a block) <= MaxGas for every block

In the future, we intend to add a Priority field to the responses that can be used to explicitly prioritize txs in the mempool for inclusion in a block proposal. See #1861.

CheckTx

If Code != 0, it will be rejected from the mempool and hence not broadcasted to other peers and not included in a proposal block.

Data contains the result of the CheckTx transaction execution, if any. It is semantically meaningless to Tendermint.

Tags include any tags for the execution, though since the transaction has not been committed yet, they are effectively ignored by Tendermint.

DeliverTx

If DeliverTx returns Code != 0, the transaction will be considered invalid, though it is still included in the block.

Data contains the result of the CheckTx transaction execution, if any. It is semantically meaningless to Tendermint.

Both the Code and Data are included in a structure that is hashed into the LastResultsHash of the next block header.

Tags include any tags for the execution, which Tendermint will use to index the transaction by. This allows transactions to be queried according to what events took place during their execution.

See issue #1007 for how the tags will be hashed into the next block header.

Validator Updates

The application may set the validator set during InitChain, and update it during EndBlock.

Note that the maximum total power of the validator set is bounded by MaxTotalVotingPower = MaxInt64 / 8. Applications are responsible for ensuring they do not make changes to the validator set that cause it to exceed this limit.

Additionally, applications must ensure that a single set of updates does not contain any duplicates - a given public key can only appear in an update once. If an update includes duplicates, the block execution will fail irrecoverably.

InitChain

ResponseInitChain can return a list of validators. If the list is empty, Tendermint will use the validators loaded in the genesis file. If the list is not empty, Tendermint will use it for the validator set. This way the application can determine the initial validator set for the blockchain.

EndBlock

Updates to the Tendermint validator set can be made by returning ValidatorUpdate objects in the ResponseEndBlock:

message ValidatorUpdate {
  PubKey pub_key
  int64 power
}

message PubKey {
  string type
  bytes  data
}

The pub_key currently supports only one type:

• type = "ed25519" and data = <raw 32-byte public key>

The power is the new voting power for the validator, with the following rules:

• power must be non-negative
• if power is 0, the validator must already exist, and will be removed from the validator set
• if power is non-0:
  • if the validator does not already exist, it will be added to the validator set with the given power
  • if the validator does already exist, its power will be adjusted to the given power
• the total power of the new validator set must not exceed MaxTotalVotingPower

Note the updates returned in block H will only take effect at block H+2.

Consensus Parameters

ConsensusParams enforce certain limits in the blockchain, like the maximum size of blocks, amount of gas used in a block, and the maximum acceptable age of evidence. They can be set in InitChain and updated in EndBlock.

BlockParams.MaxBytes

The maximum size of a complete Amino encoded block. This is enforced by Tendermint consensus.

This implies a maximum tx size that is this MaxBytes, less the expected size of the header, the validator set, and any included evidence in the block.

Must have 0 < MaxBytes < 100 MB.

BlockParams.MaxGas

The maximum of the sum of GasWanted in a proposed block. This is not enforced by Tendermint consensus. It is left to the app to enforce (ie. if txs are included past the limit, they should return non-zero codes). It is used by Tendermint to limit the txs included in a proposed block.

Must have MaxGas >= -1. If MaxGas == -1, no limit is enforced.

BlockParams.TimeIotaMs

The minimum time between consecutive blocks (in milliseconds). This is enforced by Tendermint consensus.

Must have TimeIotaMs > 0 to ensure time monotonicity.

EvidenceParams.MaxAgeDuration

This is the maximum age of evidence in time units. This is enforced by Tendermint consensus.

If a block includes evidence older than this (AND the evidence was created more than MaxAgeNumBlocks ago), the block will be rejected (validators won't vote for it).

Must have MaxAgeDuration > 0.

EvidenceParams.MaxAgeNumBlocks

This is the maximum age of evidence in blocks. This is enforced by Tendermint consensus.

If a block includes evidence older than this (AND the evidence was created more than MaxAgeDuration ago), the block will be rejected (validators won't vote for it).

Must have MaxAgeNumBlocks > 0.

EvidenceParams.MaxNum

This is the maximum number of evidence that can be committed to a single block.

The product of this and the MaxEvidenceBytes must not exceed the size of a block minus it's overhead ( ~ MaxBytes).

The amount must be a positive number.

EvidenceParams.ProofTrialPeriod

This is the duration in terms of blocks that an indicted validator has to prove a correct lock change in the event of amnesia evidence when a validator voted more than once across different rounds.

This must be positive and less then EvidenceParams.MaxAgeNumBlocks.

Updates

The application may set the ConsensusParams during InitChain, and update them during EndBlock. If the ConsensusParams is empty, it will be ignored. Each field that is not empty will be applied in full. For instance, if updating the Block.MaxBytes, applications must also set the other Block fields (like Block.MaxGas), even if they are unchanged, as they will otherwise cause the value to be updated to 0.

InitChain

ResponseInitChain includes a ConsensusParams. If its nil, Tendermint will use the params loaded in the genesis file. If it's not nil, Tendermint will use it. This way the application can determine the initial consensus params for the blockchain.

EndBlock

ResponseEndBlock includes a ConsensusParams. If its nil, Tendermint will do nothing. If it's not nil, Tendermint will use it. This way the application can update the consensus params over time.

Note the updates returned in block H will take effect right away for block H+1.

Query

Query is a generic method with lots of flexibility to enable diverse sets of queries on application state. Tendermint makes use of Query to filter new peers based on ID and IP, and exposes Query to the user over RPC.

Note that calls to Query are not replicated across nodes, but rather query the local node's state - hence they may return stale reads. For reads that require consensus, use a transaction.

The most important use of Query is to return Merkle proofs of the application state at some height that can be used for efficient application-specific lite-clients.

Note Tendermint has technically no requirements from the Query message for normal operation - that is, the ABCI app developer need not implement Query functionality if they do not wish too.

Query Proofs

The Tendermint block header includes a number of hashes, each providing an anchor for some type of proof about the blockchain. The ValidatorsHash enables quick verification of the validator set, the DataHash gives quick verification of the transactions included in the block, etc.

The AppHash is unique in that it is application specific, and allows for application-specific Merkle proofs about the state of the application. While some applications keep all relevant state in the transactions themselves (like Bitcoin and its UTXOs), others maintain a separated state that is computed deterministically from transactions, but is not contained directly in the transactions themselves (like Ethereum contracts and accounts). For such applications, the AppHash provides a much more efficient way to verify lite-client proofs.

ABCI applications can take advantage of more efficient lite-client proofs for their state as follows:

• return the Merkle root of the deterministic application state in ResponseCommit.Data.
• it will be included as the AppHash in the next block.
• return efficient Merkle proofs about that application state in ResponseQuery.Proof that can be verified using the AppHash of the corresponding block.

For instance, this allows an application's lite-client to verify proofs of absence in the application state, something which is much less efficient to do using the block hash.

Some applications (eg. Ethereum, Cosmos-SDK) have multiple "levels" of Merkle trees, where the leaves of one tree are the root hashes of others. To support this, and the general variability in Merkle proofs, the ResponseQuery.Proof has some minimal structure:

message Proof {
  repeated ProofOp ops
}

message ProofOp {
  string type = 1;
  bytes key = 2;
  bytes data = 3;
}

Each ProofOp contains a proof for a single key in a single Merkle tree, of the specified type. This allows ABCI to support many different kinds of Merkle trees, encoding formats, and proofs (eg. of presence and absence) just by varying the type. The data contains the actual encoded proof, encoded according to the type. When verifying the full proof, the root hash for one ProofOp is the value being verified for the next ProofOp in the list. The root hash of the final ProofOp in the list should match the AppHash being verified against.

Peer Filtering

When Tendermint connects to a peer, it sends two queries to the ABCI application using the following paths, with no additional data:

• /p2p/filter/addr/<IP:PORT>, where <IP:PORT> denote the IP address and the port of the connection
• p2p/filter/id/<ID>, where <ID> is the peer node ID (ie. the pubkey.Address() for the peer's PubKey)

If either of these queries return a non-zero ABCI code, Tendermint will refuse to connect to the peer.

Paths

Queries are directed at paths, and may optionally include additional data.

The expectation is for there to be some number of high level paths differentiating concerns, like /p2p, /store, and /app. Currently, Tendermint only uses /p2p, for filtering peers. For more advanced use, see the implementation of Query in the Cosmos-SDK.

Crash Recovery

On startup, Tendermint calls the Info method on the Info Connection to get the latest committed state of the app. The app MUST return information consistent with the last block it succesfully completed Commit for.

If the app succesfully committed block H but not H+1, then last_block_height = H and last_block_app_hash = <hash returned by Commit for block H>. If the app failed during the Commit of block H, then last_block_height = H-1 and last_block_app_hash = <hash returned by Commit for block H-1, which is the hash in the header of block H>.

We now distinguish three heights, and describe how Tendermint syncs itself with the app.

storeBlockHeight = height of the last block Tendermint saw a commit for
stateBlockHeight = height of the last block for which Tendermint completed all
    block processing and saved all ABCI results to disk
appBlockHeight = height of the last block for which ABCI app succesfully
    completed Commit

Note we always have storeBlockHeight >= stateBlockHeight and storeBlockHeight >= appBlockHeight Note also we never call Commit on an ABCI app twice for the same height.

The procedure is as follows.

First, some simple start conditions:

If appBlockHeight == 0, then call InitChain.

If storeBlockHeight == 0, we're done.

Now, some sanity checks:

If storeBlockHeight < appBlockHeight, error If storeBlockHeight < stateBlockHeight, panic If storeBlockHeight > stateBlockHeight+1, panic

Now, the meat:

If storeBlockHeight == stateBlockHeight && appBlockHeight < storeBlockHeight, replay all blocks in full from appBlockHeight to storeBlockHeight. This happens if we completed processing the block, but the app forgot its height.

If storeBlockHeight == stateBlockHeight && appBlockHeight == storeBlockHeight, we're done. This happens if we crashed at an opportune spot.

If storeBlockHeight == stateBlockHeight+1 This happens if we started processing the block but didn't finish.

If appBlockHeight < stateBlockHeight replay all blocks in full from appBlockHeight to storeBlockHeight-1, and replay the block at storeBlockHeight using the WAL. This happens if the app forgot the last block it committed.

If appBlockHeight == stateBlockHeight, replay the last block (storeBlockHeight) in full. This happens if we crashed before the app finished Commit

If appBlockHeight == storeBlockHeight update the state using the saved ABCI responses but dont run the block against the real app. This happens if we crashed after the app finished Commit but before Tendermint saved the state.

State Sync

A new node joining the network can simply join consensus at the genesis height and replay all historical blocks until it is caught up. However, for large chains this can take a significant amount of time, often on the order of days or weeks.

State sync is an alternative mechanism for bootstrapping a new node, where it fetches a snapshot of the state machine at a given height and restores it. Depending on the application, this can be several orders of magnitude faster than replaying blocks.

Note that state sync does not currently backfill historical blocks, so the node will have a truncated block history - users are advised to consider the broader network implications of this in terms of block availability and auditability. This functionality may be added in the future.

For details on the specific ABCI calls and types, see the methods and types section.

Taking Snapshots

Applications that want to support state syncing must take state snapshots at regular intervals. How this is accomplished is entirely up to the application. A snapshot consists of some metadata and a set of binary chunks in an arbitrary format:

• Height (uint64): The height at which the snapshot is taken. It must be taken after the given height has been committed, and must not contain data from any later heights.

• Format (uint32): An arbitrary snapshot format identifier. This can be used to version snapshot formats, e.g. to switch from Protobuf to MessagePack for serialization. The application can use this when restoring to choose whether to accept or reject a snapshot.

• Chunks (uint32): The number of chunks in the snapshot. Each chunk contains arbitrary binary data, and should be less than 16 MB; 10 MB is a good starting point.

• Hash ([]byte): An arbitrary hash of the snapshot. This is used to check whether a snapshot is the same across nodes when downloading chunks.

• Metadata ([]byte): Arbitrary snapshot metadata, e.g. chunk hashes for verification or any other necessary info.

For a snapshot to be considered the same across nodes, all of these fields must be identical. When sent across the network, snapshot metadata messages are limited to 4 MB.

When a new node is running state sync and discovering snapshots, Tendermint will query an existing application via the ABCI ListSnapshots method to discover available snapshots, and load binary snapshot chunks via LoadSnapshotChunk. The application is free to choose how to implement this and which formats to use, but should provide the following guarantees:

• Consistent: A snapshot should be taken at a single isolated height, unaffected by concurrent writes. This can e.g. be accomplished by using a data store that supports ACID transactions with snapshot isolation.

• Asynchronous: Taking a snapshot can be time-consuming, so it should not halt chain progress, for example by running in a separate thread.

• Deterministic: A snapshot taken at the same height in the same format should be identical (at the byte level) across nodes, including all metadata. This ensures good availability of chunks, and that they fit together across nodes.

A very basic approach might be to use a datastore with MVCC transactions (such as RocksDB), start a transaction immediately after block commit, and spawn a new thread which is passed the transaction handle. This thread can then export all data items, serialize them using e.g. Protobuf, hash the byte stream, split it into chunks, and store the chunks in the file system along with some metadata - all while the blockchain is applying new blocks in parallel.

A more advanced approach might include incremental verification of individual chunks against the chain app hash, parallel or batched exports, compression, and so on.

Old snapshots should be removed after some time - generally only the last two snapshots are needed (to prevent the last one from being removed while a node is restoring it).

Bootstrapping a Node

An empty node can be state synced by setting the configuration option statesync.enabled = true. The node also needs the chain genesis file for basic chain info, and configuration for light client verification of the restored snapshot: a set of Tendermint RPC servers, and a trusted header hash and corresponding height from a trusted source, via the statesync configuration section.

Once started, the node will connect to the P2P network and begin discovering snapshots. These will be offered to the local application, and once a snapshot is accepted Tendermint will fetch and apply the snapshot chunks. After all chunks have been successfully applied, Tendermint verifies the app's AppHash against the chain using the light client, then switches the node to normal consensus operation.

Snapshot Discovery

When the empty node join the P2P network, it asks all peers to report snapshots via the ListSnapshots ABCI call (limited to 10 per node). After some time, the node picks the most suitable snapshot (generally prioritized by height, format, and number of peers), and offers it to the application via OfferSnapshot. The application can choose a number of responses, including accepting or rejecting it, rejecting the offered format, rejecting the peer who sent it, and so on. Tendermint will keep discovering and offering snapshots until one is accepted or the application aborts.

Snapshot Restoration

Once a snapshot has been accepted via OfferSnapshot, Tendermint begins downloading chunks from any peers that have the same snapshot (i.e. that have identical metadata fields). Chunks are spooled in a temporary directory, and then given to the application in sequential order via ApplySnapshotChunk until all chunks have been accepted.

As with taking snapshots, the method for restoring them is entirely up to the application, but will generally be the inverse of how they are taken.

During restoration, the application can respond to ApplySnapshotChunk with instructions for how to continue. This will typically be to accept the chunk and await the next one, but it can also ask for chunks to be refetched (either the current one or any number of previous ones), P2P peers to be banned, snapshots to be rejected or retried, and a number of other responses - see the ABCI reference for details.

If Tendermint fails to fetch a chunk after some time, it will reject the snapshot and try a different one via OfferSnapshot - the application can choose whether it wants to support restarting restoration, or simply abort with an error.

Snapshot Verification

Once all chunks have been accepted, Tendermint issues an Info ABCI call to retrieve the LastBlockAppHash. This is compared with the trusted app hash from the chain, retrieved and verified using the light client. Tendermint also checks that LastBlockHeight corresponds to the height of the snapshot.

This verification ensures that an application is valid before joining the network. However, the snapshot restoration may take a long time to complete, so applications may want to employ additional verification during the restore to detect failures early. This might e.g. include incremental verification of each chunk against the app hash (using bundled Merkle proofs), checksums to protect against data corruption by the disk or network, and so on. However, it is important to note that the only trusted information available is the app hash, and all other snapshot metadata can be spoofed by adversaries.

Apps may also want to consider state sync denial-of-service vectors, where adversaries provide invalid or harmful snapshots to prevent nodes from joining the network. The application can counteract this by asking Tendermint to ban peers. As a last resort, node operators can use P2P configuration options to whitelist a set of trusted peers that can provide valid snapshots.

Transition to Consensus

Once the snapshot has been restored, Tendermint gathers additional information necessary for bootstrapping the node (e.g. chain ID, consensus parameters, validator sets, and block headers) from the genesis file and light client RPC servers. It also fetches and records the AppVersion from the ABCI application.

Once the node is bootstrapped with this information and the restored state machine, it transitions to fast sync (if enabled) to fetch any remaining blocks up the the chain head, and then transitions to regular consensus operation. At this point the node operates like any other node, apart from having a truncated block history at the height of the restored snapshot.