With this commit we make our routing more robust by looking for the
k-shortest paths rather than a single shortest path and using that
unconditionally. In a nut shell Yen’s algorithm does the following:
* Find the shortest path from the source to the destination
* From k=1…K, walk the kth shortest path and find possible
divergence from each node in the path
Our version of Yen’s implemented is slightly modified, rather than
actually increasing the edge weights or remove vertexes from the graph,
we instead use two black-lists: one for edges and the other for
vertexes. Our modified version of Djikstra’s algorithm is then made
aware of these black lists in order to more efficiently implement the
path iteration via spur and root node.
This commit modifies the findRoute method by first calling it findPath,
but also making the following modifications.
First, two new black-listing maps are now passed in. These two maps
contain vertexes but also edges to ignore while performing path
finding. These maps will be used in order to ensure that we don’t
duplicate paths or back-track when executing our KSP algorithm.
Next, we now ensure that the path returned from the findPath function
is ordered properly in the direction of source to target. Such a change
is required for our KSP algorithm to function properly.
This commit modifies the findRoute function to decouple the
validation+creation of a route, from the path finding algorithm itself.
When we say “route”, we mean the full payment route complete with
time-lock and fee information. When we say “path” we simple mean an
ordered set of channel edges from one node to another target node.
With this commit we can now perform path finding independent of route
creation which will be needed in the up coming refactor to implement a
new modified k-shortest paths algorithm.
This commit slightly modified findRoute to accept the node which should
be used as the starting point in our path finding algorithm. With this
change, as we move to a k-shortest paths algorithm this modification
will be needed as all of our path finding attempts won’t always
originate from a the same starting point.
In this commit we now utilize the node distance heap that was added in
a prior commit into our core path finding logic. With this new data
structure, we no longer linearly scan the distance of all vertexes from
the source node when deciding which one to greedily explore.
Instead, we now start with the source added to our distance heap, then
new vertexes are progressively added to our heap as their edges are
explored. With this change, we move the computational complexity of our
path finding algorithm closer to the theoretical limit.
This commit modifies our modified version of Dijkstra's to include
sufficient link capacity within the algorithm’s relaxation condition.
With this change, we’ll now avoid exploring the graph in the direction
of a link that has insufficient capacity, allowing us to terminate path
finding more quickly.
This commit introduces a new heap struct that will be used to keep
track of the next closest node to the source during path finding within
our modified Dijkstra's algorithm.
This commit modifies the path finding routines to properly use the new
channel edge related API exposed by the database. Additionally, a new
type `ChannelHop` has been introduced which couples an edges routing
policy with the capacity and origin chain of the channel.
This fixes the bug reproduced in #114. The prevHop map in newRoute may
include many more edges than what is used to produce the final route, and
thus the check prior to building the route could result with incorrect
errors being reported. We move this check to after the number of edges
to be used for the route is deduced.
This commit seems to fix a sporadic error within the integration tests
which would at times cause a panic when a payment as initiated.
This issue was with the way were deleting from the middle of the slice
of unvisited nodes within the graph. Assigning the last element to the
middle would at times cause a panic the last element may be nil. To fix
this, we now manually copy every item over by one, preserving the order
of the slice, and possibly fixing the panic once and for all.
This commit overhauls the routing package significantly to simplify the
code, conform to the rest of the coding style within the package, and
observe the new authenticated gossiping scheme outlined in BOLT07.
As a major step towards a more realistic path finding algorithm, fees
are properly calculated and observed during path finding. If a path has
sufficient capacity _before_ fees are applied, but afterwards the
finalized route would exceed the capacity of a single link, the route
is marked as invalid.
Currently a naive weighting algorithm is used which only factors in the
time-lock delta at each hop, thereby optimizing for the lowest time
lock. Fee calculation also isn’t finalized since we aren’t yet using
milli-satoshi throughout the daemon. The final TODO item within the PR
is to properly perform a multi-path search and rank the results based
on a summation heuristic rather than just return the first (out of
many) route found.
On the server side, once nodes are initially connected to the daemon,
our routing table will be synced with the peer’s using a naive “just
send everything scheme” to hold us over until I spec out some a
efficient graph reconciliation protocol. Additionally, the routing
table is now pruned by the channel router itself once new blocks arrive
rather than depending on peers to tell us when a channel flaps or is
closed.
Finally, the validation of peer announcements aren’t yet fully
implemented as they’ll be implemented within the pending discovery
package that was blocking on the completion of this package. Most off
the routing message processing will be moved out of this package and
into the discovery package where full validation will be carried out.
