This section describes a version of the Tendermint consensus protocol that uses proposer-based timestamps.
Tendermint provides a deterministic, Byzantine fault-tolerant, source of time,
defined by the Time
field present in the headers of committed blocks.
In the current consensus implementation, the timestamp of a block is
computed by the BFTTime
algorithm:
-
Validators include a timestamp in the
Precommit
messages they broadcast. Timestamps are retrieved from the validators' local clocks, with the only restriction that they must be monotonic:- The timestamp of a
Precommit
message voting for a block cannot be earlier than theTime
field of that block;
- The timestamp of a
-
The timestamp of a block is deterministically computed from the timestamps of a set of
Precommit
messages that certify the commit of the previous block. This certificate, a set ofPrecommit
messages from a round of the previous height, is selected by the block's proposer and stored in theCommit
field of the block:- The block timestamp is the median of the timestamps of the
Precommit
messages included in theCommit
field, weighted by their voting power. Block timestamps are monotonic because timestamps of validPrecommit
messages are monotonic;
- The block timestamp is the median of the timestamps of the
Assuming that the voting power controlled by Byzantine validators is bounded by f
,
the cumulative voting power of any valid Commit
set must be at least 2f+1
.
As a result, the timestamp computed by BFTTime
is not influenced by Byzantine validators,
as the weighted median of Commit
timestamps comes from the clock of a non-faulty validator.
Tendermint does not make any assumptions regarding the clocks of (correct) validators,
as block timestamps have no impact in the consensus protocol.
However, the Time
field of committed blocks is used by other components of Tendermint,
such as IBC, the evidence, staking, and slashing modules.
And it is used based on the common belief that block timestamps
should bear some resemblance to real time, which is not guaranteed.
A more comprehensive discussion of the limitations of BFTTime
can be found in the first draft of this proposal.
Of particular interest is to possibility of having validators equipped with "faulty" clocks,
not fairly accurate with real time, that control more than f
voting power,
plus the proposer's flexibility when selecting a Commit
set,
and thus determining the timestamp for a block.
In the proposed solution, the timestamp of a block is assigned by its proposer, according with its local clock. In other words, the proposer of a block also proposes a timestamp for the block. Validators can accept or reject a proposed block. A block is only accepted if its timestamp is acceptable. A proposed timestamp is acceptable if it is received within a certain time window, determined by synchronous parameters.
PBTS therefore augments the system model considered by Tendermint with synchronous assumptions:
-
Synchronized clocks: simultaneous clock reads at any two correct validators differ by at most
PRECISION
; -
Bounded message delays: the end-to-end delay for delivering a message to all correct validators is bounded by
MSGDELAY
. This assumption is restricted toProposal
messages, broadcast by proposers.
PRECISION
and MSGDELAY
are consensus parameters, shared by all validators,
that define whether the timestamp of a block is acceptable.
Let t
be the time, read from its local clock, at which a validator
receives, for the first time, a proposal with timestamp ts
:
- [Time-Validity] The proposed timestamp
ts
received at local timet
is accepted if it satisfies the timely predicate:ts - PRECISION <= t <= ts + MSGDELAY + PRECISION
The left inequality of the timely predicate establishes that proposed timestamps
should be in the past, when adjusted by the clocks PRECISION
.
The right inequality of the timely predicate establishes that proposed timestamps
should not be too much in the past, more precisely, not more than MSGDELAY
in the past,
when adjusted by the clocks PRECISION
.
A more detailed and formalized description is available in the System Model and Properties document
The implementation of PBTS requires some changes in Tendermint consensus algorithm, summarized below:
-
A proposer timestamps a block with the current time, read from its local clock. The block's timestamp represents the time at which it was assembled (after the
getValue()
call in line 18 of the arXiv algorithm):-
Block timestamps are definitive, meaning that the original timestamp is retained when a block is re-proposed (line 16);
-
To preserve monotonicity, a proposer might need to wait until its clock reads a time greater than the timestamp of the previous block;
-
-
A validator only prevotes for timely blocks, that is, blocks whose timestamps are considered timely (compared to the original Tendermint consensus, a check is added to line 23). If the block proposed in a round is considered untimely, the validator prevotes
nil
(line 26):-
Validators register the time at which they received
Proposal
messages, in order to evaluate the timely predicate; -
Blocks that are re-proposed because they received
2f+1 Prevotes
in a previous round (line 28) are not subject to the timely predicate, as they have already been evaluated as timely at a previous round.
-
The more complex change proposed regards blocks that can be re-proposed in multiple rounds. The current solution improves the first version of the specification (that never had been implemented) by simplifying the way this situation is handled, from a recursive reasoning regarding valid blocks that are re-proposed.
The full solution is detailed and formalized in the Protocol Specification document.
- System Model and Properties
- Protocol Specification
- TLA+ Specification (first draft, not updated)
- PBTS: evidence #355: not really clear the context, probably not going to be solved.
- PBTS: should synchrony parameters be adaptive? #371
- PBTS: Treat proposal and block parts explicitly in the spec #372
- PBTS: margins for proposal times assigned by Byzantine proposers #377