- 19-Apr-2022: Initial draft started (@williambanfield).
- 3-May-2022: Initial draft complete (@williambanfield).
The Block data structure is a very central structure within Tendermint. Because
of its centrality, it has gained several fields over the years through accretion.
Not all of these fields may be necessary any more. This document examines which
of these fields may no longer be necessary for inclusion in the block and makes
recommendations about how to proceed with each of them.
The current block structure contains multiple fields that are not required for validation or execution of a Tendermint block. Some of these fields had vestigial purposes that they no longer serve and some of these fields exist as a result of internal Tendermint domain objects leaking out into the external data structure.
In so far as is possible, we should consolidate and prune these superfluous fields before releasing a 1.0 version of Tendermint. All pruning of these fields should be done with the aim of simplifying the structures to what is needed while preserving information that aids with debugging and that also allow external protocols to function more efficiently than if they were removed.
The current block structures are included here to aid discussion.
message Block {
Header header = 1;
Data data = 2;
tendermint.types.EvidenceList evidence = 3;
Commit last_commit = 4;
}message Header {
tendermint.version.Consensus version = 1;
string chain_id = 2;
int64 height = 3;
google.protobuf.Timestamp time = 4;
BlockID last_block_id = 5;
bytes last_commit_hash = 6;
bytes data_hash = 7;
bytes validators_hash = 8;
bytes next_validators_hash = 9;
bytes consensus_hash = 10;
bytes app_hash = 11;
bytes last_results_hash = 12;
bytes evidence_hash = 13;
bytes proposer_address = 14;
}
message Data {
repeated bytes txs = 1;
}message EvidenceList {
repeated Evidence evidence = 1;
}message Commit {
int64 height = 1;
int32 round = 2;
BlockID block_id = 3;
repeated CommitSig signatures = 4;
}message CommitSig {
BlockIDFlag block_id_flag = 1;
bytes validator_address = 2;
google.protobuf.Timestamp timestamp = 3;
bytes signature = 4;
}message BlockID {
bytes hash = 1;
PartSetHeader part_set_header = 2;
}A block is the structure produced as the result of an instance of the Tendermint consensus algorithm. At its simplest, the 'block' can be represented as a Merkle root hash of all of the data used to construct and produce the hash. Our current block proto structure includes far from all of the data used to produce the hashes included in the block.
It does not contain the full AppState, it does not contain the ConsensusParams,
nor the LastResults, nor the ValidatorSet. Additionally, the layout of
the block structure is not inherently tied to this Merkle root hash. Different
layouts of the same set of data could trivially be used to construct the
exact same hash. The thing we currently call the 'Block' is really just a view
into a subset of the data used to construct the root hash. Sections of the
structure can be modified as long as alternative methods exist to query and
retrieve the constituent values.
This digression is aimed at informing what it means to consolidate 'fields' in the 'block'. The discussion of what should be included in the block can be teased apart into a few different lines of inquiry.
- What values need to be included as part of the Merkle tree so that the consensus algorithm can use proof-of-stake consensus to validate all of the properties of the chain that we would like?
- How can we create views of the data that can be easily retrieved, stored, and verified by the relevant protocols?
These two concerns are intertwined at the moment as a result of how we store and propagate our data but they don't necessarily need to be. This document focuses primarily on the first concern by suggesting fields that can be completely removed without any loss in the function of our consensus algorithm.
This document also suggests ways that we may update our storage and propagation mechanisms to better take advantage of Merkle tree nature of our data although these are not its primary concern.
This section proposes a list of data that could be completely removed from the Merkle tree with no loss to the functionality of our consensus algorithm.
Where the change is possible but would hamper external protocols or make debugging more difficult, that is noted in discussion.
This field contains the timestamp included in the precommit message that was issued for the block. The field was once used to produce the timestamp of the block. With the introduction of Proposer-Based Timestamps, This field is no longer used in any Tendermint algorithms and can be completely removed.
The ValidatorAddress is included in each CommitSig structure. This field
is hashed along with all of the other fields of the CommitSigs in the block
to form the LastCommitHash field in the Header. The ValidatorAddress is
somewhat redundant in the hash. Each validator has a unique position in the
CommitSig and the hash is built preserving this order. Therefore, the
information of which signature corresponds to which validator is included in
the root hash, even if the address is absent.
It's worth noting that the validator address could still be included in the
hash even if it is absent from the CommitSig structure in the block by
simply hashing it locally at each validator but not including it in the block.
The reverse is also true. It would be perfectly possible to not include the
ValidatorAddress data in the LastCommitHash but still include the field in
the block.
The BlockID field comprises the PartSetHeader and the hash of the block.
The PartSetHeader is used by nodes to gossip the block by dividing it into
parts. Nodes receive the PartSetHeader from their peers, informing them of
what pieces of the block to gather. There is no strong reason to include this
value in the block. Validators will still be able to gossip and validate the
blocks that they received from their peers using this mechanism even if it is
not written into the block. The BlockID can therefore be consolidated into
just the hash of the block. This is by far the most uncontroversial change
and there appears to be no good reason not to do it. Further evidence that
the field is not meaningful can be found in the fact that the field is not
actually validated to ensure it is correct during block validation. Validation
only checks that the field is well formed.
The ChainID is a string selected by the chain operators, usually a
human-readable name for the network. This value is immutable for the lifetime
of the chain and is defined in the genesis file. It is therefore hashed into the
original block and therefore transitively included as in the Merkle root hash of
every block. The redundant field is a candidate for removal from the root hash
of each block. However, aesthetically, it's somewhat nice to include in each
block, as if the block was 'stamped' with the ID. Additionally, re-validating
the value from genesis would be painful and require reconstituting potentially
large chains. I'm therefore mildly in favor of maintaining this redundant
piece of information. We pay almost no storage cost for maintaining this
identical data, so the only cost is in the time required to hash it into the
structure.
LastResultsHash is a hash covering the result of executing the transactions
from the previous block. It covers the response Code, Data, GasWanted,
and GasUsed with the aim of ensuring that execution of all of the transactions
was performed identically on each node. The data covered by this field should
be also reflected in the AppHash. The AppHash is provided by the application
and should be deterministically calculated by each node. This field could
therefore be removed on the grounds that its data is already reflected elsewhere.
I would advocate for keeping this field. This field provides an additional check
for determinism across nodes. Logic to update the application hash is more
complicated for developers to implement because it relies either on building a
complete view of the state of the application data. The Results returned by
the application contain simple response codes and deterministic data bytes.
Leaving the field will allow for transaction execution issues that are not
correctly reflected in the AppHash to be more completely diagnosed.
Take the case of mismatch of LastResultsHash between two nodes, A and B, where both
nodes report divergent values. If A and B both report
the same AppHash, then some non-deterministic behavior occurred that was not
accurately reflected in the AppHash. The issue caused by this
non-determinism may not show itself for several blocks, but catching the divergent
state earlier will improve the chances that a chain is able to recover.
Both ValidatorsHash and NextValidatorsHash are included in the block
header. Validatorshash contains the hash of the public key and voting power
of each validator in the active set for the current block and NextValidatorsHash
contains the same data but for the next height.
This data is effectively redundant. Having both values present in the block structure is helpful for light client verification. The light client is able to easily determine if two sequential blocks used the same validator set by querying only one header.
ValidatorsHash is also important to the light client algorithm for performing block
validation. The light client uses this field to ensure that the validator set
it fetched from a full node is correct. It can be sure of the correctness of
the retrieved structure by hashing it and checking the hash against the ValidatorsHash
of the block it is verifying. Because a validator that the light client trusts
signed over the ValidatorsHash, it can be certain of the validity of the
structure. Without this check, phony validator sets could be handed to the light
client and the code tricked into believing a different validator set was present
at a height, opening up a major hole in the light client security model.
This creates a recursive problem. To verify the validator set that signed the
block at height H, what information do we need? We could fetch the
NextValidatorsHash from height H-1, but how do we verify that that hash is correct?
The section below details a change to allow the ProposerAddress to be calculated
from a field added to the block. This would allow the Address to be dropped
from the block. Consumers of the chain could run the proposer selection algorithm
to determine who proposed each block.
I would advocate against this. Any consumer of the chain that wanted to know which validator proposed a block would have to run the proposer selection algorithm. This algorithm is not widely implemented, meaning that consumers in other languages would need to implement the algorithm to determine a piece of basic information about the chain.
The proof of lock round is the round of consensus for a height in which the Tendermint algorithm observed a super majority of voting power on the network for a block.
Including this value in the block will allow validation of currently
un-validated metadata. Specifically, including this value will allow Tendermint
to validate that the ProposerAddress in the block is correct. Without knowing
the locked round number, Tendermint cannot calculate which validator was supposed
to propose a height. Because of this, our validation logic does not check that
the ProposerAddress included in the block corresponds to the validator that
proposed the height. Instead, the validation logic simply checks that the value
is an address of one of the known validators.
Currently, we maintain the committed round in the Commit for height H, which is
written into the block at height H+1. This value corresponds to the round in
which the proposer of height H+1 received the commit for height H. The proof
of lock round would not subsume this value.
Currently we store the every piece of each block in the BlockStore.
I suspect that this has lead to some mistakes in reasoning around the merits of
consolidating fields in the block. We could update the storage scheme we use to
store only some pieces of each block and still achieve a space savings without having
to change the block structure at all.
The main way to achieve this would be by no longer saving data that does not change.
At each height we save a set of data that is unlikely to have changed from the
previous height in the block structure, this includes the ValidatorAddresses,
the ValidatorsHash, the ChainID. These do not need to be saved along with
each block. We could easily save the value and the height at which the value
was updated and construct each block using the data that existed at the time.
This document does not make any specific recommendations around storage since that is likely to change with upcoming improvements to the database infrastructure. However, it's important to note that removing fields from the block for the purposes of 'saving space' may not be that meaningful. We should instead focus our attention of removing fields from the block that are no longer needed for correct functioning of the protocol.
Block propagation suffers from the same issue that plagues block storage, we
propagate all of the contents of each block proto even when these contents are redundant
or unchanged from previous blocks. For example, we propagate the ValidatorAddresses
for each block in the CommitSig structure even when it never changed from a
previous height. We could achieve a speed-up in many cases by communicating the
hashes first and letting peers request additional information when they do not
recognize the communicated hash.
For example, in the case of the ValidatorAddresses, the node would first
communicate the ValidatorsHash of the block to its peers. The peers would
check their storage for a validator set matching the provided hash. If the peer
has a matching set, it would populate its local block structure with the
appropriate values from its store. If peer did not have a matching set, it would
issue a request to its peers, either via P2P or RPC for the data it did not have.
Conceptually, this is very similar to how content addressing works in protocols such as git where pushing a commit does not require pushing the entire contents of the tree referenced by the commit.
As outlined in the section On Tendermint Blocks, there
is a distinction between what data is referenced in the Merkle root hash and the
contents of the proto structure we currently call the Block.
Any changes to the Merkle root hash will necessarily be breaking for legacy light clients. Either a soft-upgrades scheme will need to be implemented or a hard fork will be required for chains and light clients to function with the new hashes. This means that all of the additions and deletions from the Merkle root hash proposed by this document will be light client breaking.
Changes to the block structure alone are not necessarily light client breaking if the
data being hashed are identical and legacy views into the data are provided
for old light clients during transitions. For example, a newer version of the
block structure could move the ValidatorAddress field to a different field
in the block while still including it in the hashed data of the LastCommitHash.
As long as old light clients could still fetch the old data structure, then
this would not be light client breaking.