Synchronisation heuristic¶
When a new node joins the network, it must bootstrap: fetch and validate the chain before starting to bake or attest new blocks. A bootstrapping node cannot bake or attest new blocks, so for efficiency it should not bother to track a mempool: a pool of active operations.
Knowing whether a node is bootstrapped is challenging because a node cannot trust (in general) its neighbours. We tackle this problem using a synchronisation heuristic.
Synchronisation heuristic status¶
A synchronisation heuristic is used to determine whether the node is synchronised with respect to its peers.
The current synchronisation heuristic uses a synchronisation status as follows:
Unsynced
: not synchronisedSynced
: synchronised and the chain is not stuckStuck
: synchronised and the chain is stuck
Bootstrapped¶
We consider a node bootstrapped if the heuristic’s status has been
Synced
or Stuck
at least once.
Once the node is bootstrapped, the synchronisation status is still
updated (at least after each change of head) and is accessible via the RPC
/chains/<chain>/is_bootstrapped
. This RPC returns a pair where the
first component indicates if the node is bootstrapped and the second
component is the current status.
Basic description of the heuristic¶
The synchronisation heuristic relies on a notion of candidate: a pair of a block’s timestamp and a peer. The block from which the timestamp is taken must have been validated locally and it must be the most-recent such block advertised by the peer.
The heuristic is parameterised by two values (see node configuration):
a
threshold
(configuration parametersynchronisation_threshold
): the number of candidates kept by the heuristic;a
latency
(configuration parameterlatency
): a delay in seconds to control possible forks and the latency of the network (see Acceptable values for parameters).
The heuristic maintains the most recent threshold
candidates it is
aware of, with the constraint that there is at most one candidate per
peer.
The heuristic status is Synced
if the threshold
kept
candidates are all recent enough: their age is less than latency
seconds (the age is counted with respect to the current local
time). Otherwise, the heuristic status is Stuck
if the
threshold
kept candidates all have the same timestamp. Otherwise,
the status is Unsynced
.
Acceptable values for parameters¶
The heuristic accepts any value for the threshold
, but values
of 1
or less are mainly used for testing and debugging:
If
threshold
is negative, then the status is alwaysUnsynced
.If
threshold
is0
, then the status is alwaysSynced
.If
threshold
is1
, then the status cannot beStuck
.
Other values are acceptable for threshold
, but a small
threshold
(between 2
and 10
; the default being 4
) is
probably best: performances and accuracy may degrade for values much
higher than 10
.
The default value for latency
is 50
seconds. Whatever you
change it to, it should be at least long enough to include a few
blocks’ worth of the protocol’s baking rate plus some network latency,
but not so long that the node considers itself bootstrapped even
though it is many blocks away from the chain’s head.
A good value for latency
is 2
to 5
times the time between
blocks, plus a small delta for network delays. At the time of
writing, the time between two consecutive blocks is 10
seconds
when the chain is healthy (see the consensus algorithm).
A shorter latency
might give false negatives: delays from a few
neighbours might result in the node considering itself not
synchronised yet, whilst a longer latency
might give false
positives: the node considers itself synchronised whilst it still has
several blocks to catch-up on.
Formal description of the heuristic¶
In the normal case of the synchronisation heuristic, when the
threshold
parameter is greater than 1
, the status is computed
as follows:
The status is
Synced
if there are more thanthreshold
candidates that are more recent thanlatency
.The status is
Stuck
if thethreshold
most recent candidates have the same timestamp that is older thanlatency
.The status is
Unsynced
otherwise.
If the heuristic fails¶
The heuristic may fail and declare a node not synchronised, thus not
bootstrapped, when actually it should be. The administrator of the
node can use the RPC patch /chains/main {"bootstrapped": true}
to
force the node bootstrapped state, but this should be used carefully.
If you see an issue with the current heuristic, please report it.