This commit defines a new heuristic WeightedCombAttachment that takes a
set of sub-heuristics, and produces a final node score by querying the
sub-heuristics and combining the scores from them according to their
weights.
This way it will look like a regular, single heuristic to the autopilot
agemnt, but can be a more complex combination of several.
To prepare for combinning scores from multiple heuristics, we require the
scores returned from the NodeSores API to be in the range [0.0, 1.0].
The prefAttach heuristic is altered to scale the returned scores such
that the most connected node in the grpah is given a score of 1.0.
Since NodeScores no longer returns fully populated AttachmentDirectives,
we make this explicit by defining a new type NodeScore that includes a
subset of what the AttachmentDirective does.
We create a new type NodeScore which is a tuple (NodeID, score). The
weightedChoice and chooseN algorithms are altered to expect this type.
This is done in order to simplify the types we are using, since we were
only using a subset of the fields in AttachmentDirective.
Since we want to combine scores from multiple heuristics, things get
complicated if the heuristics report their own individual channel sizes.
Therefore we change the NodeScores interface slightly, letting the agent
specify the wanted channel size, and let the heuristic score the nodes
accordingly.
We let the agent call ChannelBudget on its constraints directly, and
not go through the heuristic. This is needed since when we want to have
multiple active heuristics concurrently, it won't make sense anymore to
ask each of the heuristics.
The mockConstraints are also updated to act as the mockHeuristic did
before, by making it possible to control the responses it gives by
sending them on the contained channels.
To decouple the autopilot heuristic from the constraints, we start by
abstracting them behind an interface to make them easier to mock. We
also rename them HeuristicConstraints->AgentConstraints to make it clear
that they are now constraints the agent must adhere to.
This commit fixes a subtle bug within the autopilot manager, that would
cause the active pilot to not be reset in case it wasn't started
successfully.
We also make sure the associated goroutines close over the started
pilot, and not the active pilot.
This commit makes the weightedChoice algorithm take a slice of weights
instead of a map of node scores. This let us avoid costly map allocation
and iteration.
In addition we make the chooseN algorithm keep track of the remaining
nodes by keeping a slice of weights through its entire run, similarly
avoiding costly map allocation and iteration.
In total this brings the runtime of the TestChooseNSample testcase down
from ~73s to ~3.6s.
This addition to the unit tests makes sure nodes that have no channels
in the graph are left out od the scored nodes, implicitly giving them a
score of 0.
This commit makes the autopilot agent use the new NodeScores heuristic
API to select channel candiates, instead of the Select API. The result
will be similar, but instead of selecting a set of nodes to open
channels to, we get a score based results which can later be used
together with other heuristics to choose nodes to open channels to.
This commit also makes the existing autopilot agent tests compatible
with the new NodeScores API.
This commit adds a new method NodeScores to the AttachementHeuristic
interface. Its intended use is to score a set of nodes according to
their preference as channel counterparties.
The PrefAttach heuristic gets a NodeScores method that will score the
ndoes according to their number of already existing channels, similar to
what is done already in Select.
This commit defines a new struct HeuristicConstraints that will be used
to keep track of the initial constraints the autopilot agent needs to
adhere to. This is currently done in the ConstrainedPrefAttachement
heuristic itself, but this lets us share these constraints and common
method netween several heuristics.
This commit ensures that the mock attachment
directives use unique keys, ensuring that they
aren't skipped due to already having pending
connection requests. The tests fail when
they're all the same since they collide
in the pendingConns map.
This commit modifies the autopilot agent to track
all pending connection requests, and forgo further
attempts if a connection is already present.
Previously, the agent would try and establish
hundreds of requests to a node, especially if the
connections were timing out and not returning.
This resulted in an OOM OMM when cranking up
maxchannels to 200, since there would be close
to 10k pending connections before the program was
terminated. The issue was compounded by periodic
batch timeouts, causing autopilot to try and
process thousands of triggers for failing
connections to the same peer.
With these fixes, autopilot will skip nodes that we
are trying to connect to during heuristic selection.
The CPU and memory utilization have been significantly
reduced as a result.
In this commit, we implement an optimization to the autopilot agent to
ensure that we don't spin and waste CPU when we either have a large
graph, or a high max channel target for the agent. Before this commit,
each time we went to read the state of a channel from disk, we would
decompress the EC Point each time. However, for the case of the instal
ChannlEdge struct to feed to the agent, we only actually need to obtain
the pubkey, and can save the potentially expensive point decompression
for each directional channel in the graph.
We do this to avoid a huge amount of goroutines piling up when autopilot
is trying to open many channels, as they will all block trying to send
the update on the stateUpdates channel. Now we instead send them on a
buffered channel, similar to what is done with the nodeUpdates.
We do this to avoid a huge amount of goroutines piling up on initial
graph sync, as they will all block trying to send the node update on the
stateUpdates channel. Now we instead make a new buffered channel
nodeUpdates, and just return immediately if there is already a signal in
the channel waiting to be processed.
In this commit, we modify the Node interface to return a set of raw
bytes, rather than the full pubkey struct. We do this as within the
package, commonly we only require the pubkey bytes for fingerprinting
purposes. Before this commit, we were forced to _always_ decompress the
pubkey which can be expensive done thousands of times a second.
In this commit, we modify the balanceUpdate autopilot signal to update
the balance according to what's returned to the WalletBalance callback
rather than explicitly tracking the balance. This gives the agent a
better sense of what the wallet's balance actually is.
Adds a new external signal alerting autopilot that
new nodes have been added to the channel graph or
an existing node has modified its channel
announcment. This allows autopilot to examine its
current state, and attempt to open channels if our
target state is not yet met.
In this commit, we alter our mock heuristic to also take in a quit chan.
It's possible that at the end of a test the agent is blocked on a
NeedMoreChans/Select call as their mock implementations use channels. To
prevent this, we use the agent's quit chan so that the heuristic can
safely exit once the agent does.
In this commit, we refactor the existing connection logic outside of the
ChanController's OpenChannel method. We do this as previously it was
possible for peers to stall us while attempting to connect to them. In
order to remedy this, we now attempt to connect the peer before tracking
them in our set of pending opens.
The commit ensures that for every channel, there will always
be two entries in the edges bucket. If the policy from one or
both ends of the channel is unknown, it is marked as such.
This allows efficient lookup of incoming edges. This is
required for backwards payment path finding.
In this commit, we fix an existing bug that would at times cause us to
spiral out of control and potentially created thousands of outbound
connections. Our fix is simple: limit the total number of outstanding
channel establishment attempts. Without this limit, we have no way to
bound the number of active goroutines.
Fixes#772.
In this commit, we fix a regression introduced by a recent change which
would allow the agent to detect a channel as failed, and blacklist the
node, promising faster convergence with the ideal state of the
heuristic.
The source of this bug is that we would use the set of blacklisted
nodes in order to compute how many additional channels we should open.
If 10 failures happened, then we would think that we had opened up 10
channels, and stop much earlier than we actually should.
To fix this, while ensuring we don’t retry to failed peers, the
NeedMoreChans method will now return *how* anymore channels to open,
and the Select method will take in how many channels it should try to
open *exactly*.
In this commit we modify the ConstrainedPrefAttachment.Select method to
first shuffle the set of potential candidates before selecting them.
This serves to remove the existing grouping between candidates which
may have influenced the selection.
This commit modifies the Select method for the
ConstrainedPrefAttachment attachment heuristic slightly. Previously, it
was possible for an autopilot.Agent to go over the allotted number of
channels as it would unconditionally attempt to establish channel with
all returned Attachment Directives. To remedy this, we now assume that
we already have active, or pending channels to each of the nodes in the
set of skipNodes. Therefore, we now use the size of the skipNodes map
as an upper limit within the primary selection loop.
In this commit, we ensure that we grab the mutex for the pending open
channel state when we attempt to merge the pending state with the
committed state.
This commit adds tracking of the pending channels state within the
autopilot.Agent. This fixes a class of bugs which was discovered during
the latest test net block storm wherein the Agent would attempt to
repeatedly attach to the same node due to rapid closure of other
channels.
In this commit we fix this issue by ensuring that we always factor in
the pending channel state when querying the heuristic w.r.t if we need
more channels, and if so to which nodes should be attached to.
This commit fixes a prior occasional test flake caused by the collision
of the randomly selected 64-bit integers. In order to get around this,
we now instead have a atomic monotonically increasing counter for each
channel ID used within the tests.