Here we outline the design of the AddressBook and how it used by the Peer Exchange Reactor (PEX) to ensure we are connected to good peers and to gossip peers to others.
Certain peers are special in that they are specified by the user as persistent
,
which means we auto-redial them if the connection fails, or if we fail to dial
them.
Some peers can be marked as private
, which means
we will not put them in the address book or gossip them to others.
All peers except private peers are tracked using the address book.
Peer discovery begins with a list of seeds. When we have no peers, or have been unable to find enough peers from existing ones, we dial a randomly selected seed to get a list of peers to dial.
On startup, we will also immediately dial the given list of persistent_peers
,
and will attempt to maintain persistent connections with them. If the connections die, or we fail to dial,
we will redial every 5s for a few minutes, then switch to an exponential backoff schedule,
and after about a day of trying, stop dialing the peer.
So long as we have less than MinNumOutboundPeers
, we periodically request additional peers
from each of our own. If sufficient time goes by and we still can't find enough peers,
we try the seeds again.
Peers listen on a configurable ListenAddr that they self-report in their NodeInfo during handshakes with other peers. Peers accept up to (MaxNumPeers - MinNumOutboundPeers) incoming peers.
Peers are tracked via their ID (their PubKey.Address()). Peers are added to the address book from the PEX when they first connect to us or when we hear about them from other peers.
The address book is arranged in sets of buckets, and distinguishes between vetted (old) and unvetted (new) peers. It keeps different sets of buckets for vetted and unvetted peers. Buckets provide randomization over peer selection. Peers are put in buckets according to their IP groups.
A vetted peer can only be in one bucket. An unvetted peer can be in multiple buckets, and each instance of the peer can have a different IP:PORT.
If we're trying to add a new peer but there's no space in its bucket, we'll remove the worst peer from that bucket to make room.
When a peer is first added, it is unvetted.
Marking a peer as vetted is outside the scope of the p2p
package.
For Tendermint, a Peer becomes vetted once it has contributed sufficiently
at the consensus layer; ie. once it has sent us valid and not-yet-known
votes and/or block parts for NumBlocksForVetted
blocks.
Other users of the p2p package can determine their own conditions for when a peer is marked vetted.
If a peer becomes vetted but there are already too many vetted peers, a randomly selected one of the vetted peers becomes unvetted.
If a peer becomes unvetted (either a new peer, or one that was previously vetted), a randomly selected one of the unvetted peers is removed from the address book.
More fine-grained tracking of peer behaviour can be done using a trust metric (see below), but it's best to start with something simple.
When we need more peers, we pick them randomly from the addrbook with some configurable bias for unvetted peers. The bias should be lower when we have fewer peers and can increase as we obtain more, ensuring that our first peers are more trustworthy, but always giving us the chance to discover new good peers.
We track the last time we dialed a peer and the number of unsuccessful attempts we've made. If too many attempts are made, we mark the peer as bad.
Connection attempts are made with exponential backoff (plus jitter). Because
the selection process happens every ensurePeersPeriod
, we might not end up
dialing a peer for much longer than the backoff duration.
If we fail to connect to the peer after 16 tries (with exponential backoff), we remove from address book completely.
When we’re asked for peers, we select them as follows:
maxGetSelection
peersminGetSelection
peers - if we have less than that, select them all.getSelectionPercent
of the peersSend the selected peers. Note we select peers for sending without bias for vetted/unvetted.
There are various cases where we decide a peer has misbehaved and we disconnect from them. When this happens, the peer is removed from the address book and black listed for some amount of time. We call this "Disconnect and Mark". Note that the bad behaviour may be detected outside the PEX reactor itself (for instance, in the mconnection, or another reactor), but it must be communicated to the PEX reactor so it can remove and mark the peer.
In the PEX, if a peer sends us an unsolicited list of peers, or if the peer sends a request too soon after another one, we Disconnect and MarkBad.
The quality of peers can be tracked in more fine-grained detail using a Proportional-Integral-Derivative (PID) controller that incorporates current, past, and rate-of-change data to inform peer quality.
While a PID trust metric has been implemented, it remains for future work to use it in the PEX.
See the trustmetric and trustmetric useage architecture docs for more details.