mirror of
https://github.com/lightningnetwork/lnd.git
synced 2026-08-17 13:06:14 +02:00
4fe4032b70
exp-012 could not price a warm cache. Every arm it could build bought its knowledge with payments, and payments drain the corridors they teach about, so the drain arm paid in depletion and the restore arm paid in staleness. A served cache costs its consumer nothing because it arrives over an API, and nothing in the simulator could express that. --import-weights can: beliefs from a file, no payments sent. The design decision is what gets served. Neither side's internal representation is servable -- mission control keeps a decaying penalty history keyed by the observer, the evolved routers keep an interval with an evidence count -- but both are derivable from a stream of (from, to, chan_id, amount, success, time). So the API serves OBSERVATIONS. Serving weights would force every consumer into the server's probability model; serving observations lets each consumer build its own, which is the difference between an API only lnd can use and one a competing design can. Both consumers are fed from that one stream. lnd's half goes through MissionControl.ImportHistory, which already ships, so the serving proposal needs no new consumer machinery on lnd's side. The candidate half is a new optional interface, and its being optional is itself a finding: no champion evolved so far implements it, because nothing ever asked them to. import_router_accepts is reported alongside the counts so that an ineffective import stays distinguishable from an undelivered one. Local channels are excluded by default, which is exp-012 part 4's measured rule rather than a guess: observations about remote pairs transfer between vantages, but every payment a node sends crosses one of its own channels, so importing stale claims about those poisons the first hop of everything and tripled lnd's attempt count. --import-local reproduces that deliberately. Two things surfaced while wiring it up. ImportHistory must NOT be forced: importSnapshot applies a success entry for every pair, so a pair we only observed failing carries a zero-amount success, and under force setLastPairResult rewrites the failure amount to successAmt+1 -- a 750k msat failure lands as a 1 msat failure. Unforced also gives the right semantics for a served cache, which should never overwrite fresher local knowledge. And route.Vertex marshals as a 33-element byte array, which is not a reasonable wire shape for something standing in for an API payload, so observations carry hex pubkeys. |
||
|---|---|---|
| .. | ||
| commands | ||
| lncli | ||
| lnd | ||
| routesim | ||