Calculate a new penalty based on historical channel liquidity range
authorMatt Corallo <git@bluematt.me>
Thu, 21 Jul 2022 02:08:28 +0000 (02:08 +0000)
committerMatt Corallo <git@bluematt.me>
Thu, 6 Oct 2022 21:10:23 +0000 (21:10 +0000)
Our current `ProbabilisticScorer` attempts to build a model of the
current liquidity across the payment channel network. This works
fine to ignore channels we *just* tried to pay through, but it
fails to remember patterns over longer time horizons.

Specifically, there are *many* channels within the network that are
very often either fully saturated in one direction, or are
regularly rebalanced and rarely saturated. While our model may
discover that, when it decays its offsets or if there is a
temporary change in liquidity, it effectively forgets the "normal"
state of the channel.

This causes substantially suboptimal routing in practice, and
avoiding discarding older knowledge when new datapoints come in is
a potential solution to this.

Here, we implement one such design, using the decaying buckets
added in the previous commit to calculate a probability of payment
success based on a weighted average of recent liquidity estimates
for a channel.

For each min/max liquidity bucket pair (where the min liquidity is
less than the max liquidity), we can calculate the probability that
a payment succeeds using our traditional `amount / capacity`
formula. From there, we weigh the probability by the number of
points in each bucket pair, calculating a total probability for the
payment, and assigning a penalty using the same log-probability
calculation used for the non-historical penalties.

lightning/src/routing/scoring.rs

index 2bae959f4fbe98e9e29de2cf3a510dc961f32ff0..8b475ecc0d31fe43a35a17b9252448a843da753b 100644 (file)
@@ -350,7 +350,8 @@ pub struct ProbabilisticScoringParameters {
        pub base_penalty_amount_multiplier_msat: u64,
 
        /// A multiplier used in conjunction with the negative `log10` of the channel's success
-       /// probability for a payment to determine the liquidity penalty.
+       /// probability for a payment, as determined by our latest estimates of the channel's
+       /// liquidity, to determine the liquidity penalty.
        ///
        /// The penalty is based in part on the knowledge learned from prior successful and unsuccessful
        /// payments. This knowledge is decayed over time based on [`liquidity_offset_half_life`]. The
@@ -359,7 +360,7 @@ pub struct ProbabilisticScoringParameters {
        /// uncertainty bounds of the channel liquidity balance. Amounts above the upper bound will
        /// result in a `u64::max_value` penalty, however.
        ///
-       /// Default value: 40,000 msat
+       /// Default value: 30,000 msat
        ///
        /// [`liquidity_offset_half_life`]: Self::liquidity_offset_half_life
        pub liquidity_penalty_multiplier_msat: u64,
@@ -380,7 +381,8 @@ pub struct ProbabilisticScoringParameters {
        pub liquidity_offset_half_life: Duration,
 
        /// A multiplier used in conjunction with a payment amount and the negative `log10` of the
-       /// channel's success probability for the payment to determine the amount penalty.
+       /// channel's success probability for the payment, as determined by our latest estimates of the
+       /// channel's liquidity, to determine the amount penalty.
        ///
        /// The purpose of the amount penalty is to avoid having fees dominate the channel cost (i.e.,
        /// fees plus penalty) for large payments. The penalty is computed as the product of this
@@ -395,9 +397,45 @@ pub struct ProbabilisticScoringParameters {
        /// probabilities, the multiplier will have a decreasing effect as the negative `log10` will
        /// fall below `1`.
        ///
-       /// Default value: 256 msat
+       /// Default value: 192 msat
        pub liquidity_penalty_amount_multiplier_msat: u64,
 
+       /// A multiplier used in conjunction with the negative `log10` of the channel's success
+       /// probability for the payment, as determined based on the history of our estimates of the
+       /// channel's available liquidity, to determine a penalty.
+       ///
+       /// This penalty is similar to [`liquidity_penalty_multiplier_msat`], however, instead of using
+       /// only our latest estimate for the current liquidity available in the channel, it estimates
+       /// success probability based on the estimated liquidity available in the channel through
+       /// history. Specifically, every time we update our liquidity bounds on a given channel, we
+       /// track which of several buckets those bounds fall into, exponentially decaying the
+       /// probability of each bucket as new samples are added.
+       ///
+       /// Default value: 10,000 msat
+       ///
+       /// [`liquidity_penalty_multiplier_msat`]: Self::liquidity_penalty_multiplier_msat
+       pub historical_liquidity_penalty_multiplier_msat: u64,
+
+       /// A multiplier used in conjunction with the payment amount and the negative `log10` of the
+       /// channel's success probability for the payment, as determined based on the history of our
+       /// estimates of the channel's available liquidity, to determine a penalty.
+       ///
+       /// The purpose of the amount penalty is to avoid having fees dominate the channel cost for
+       /// large payments. The penalty is computed as the product of this multiplier and the `2^20`ths
+       /// of the payment amount, weighted by the negative `log10` of the success probability.
+       ///
+       /// This penalty is similar to [`liquidity_penalty_amount_multiplier_msat`], however, instead
+       /// of using only our latest estimate for the current liquidity available in the channel, it
+       /// estimates success probability based on the estimated liquidity available in the channel
+       /// through history. Specifically, every time we update our liquidity bounds on a given
+       /// channel, we track which of several buckets those bounds fall into, exponentially decaying
+       /// the probability of each bucket as new samples are added.
+       ///
+       /// Default value: 64 msat
+       ///
+       /// [`liquidity_penalty_amount_multiplier_msat`]: Self::liquidity_penalty_amount_multiplier_msat
+       pub historical_liquidity_penalty_amount_multiplier_msat: u64,
+
        /// Manual penalties used for the given nodes. Allows to set a particular penalty for a given
        /// node. Note that a manual penalty of `u64::max_value()` means the node would not ever be
        /// considered during path finding.
@@ -605,6 +643,8 @@ impl ProbabilisticScoringParameters {
                        liquidity_penalty_multiplier_msat: 0,
                        liquidity_offset_half_life: Duration::from_secs(3600),
                        liquidity_penalty_amount_multiplier_msat: 0,
+                       historical_liquidity_penalty_multiplier_msat: 0,
+                       historical_liquidity_penalty_amount_multiplier_msat: 0,
                        manual_node_penalties: HashMap::new(),
                        anti_probing_penalty_msat: 0,
                        considered_impossible_penalty_msat: 0,
@@ -625,9 +665,11 @@ impl Default for ProbabilisticScoringParameters {
                Self {
                        base_penalty_msat: 500,
                        base_penalty_amount_multiplier_msat: 8192,
-                       liquidity_penalty_multiplier_msat: 40_000,
+                       liquidity_penalty_multiplier_msat: 30_000,
                        liquidity_offset_half_life: Duration::from_secs(3600),
-                       liquidity_penalty_amount_multiplier_msat: 256,
+                       liquidity_penalty_amount_multiplier_msat: 192,
+                       historical_liquidity_penalty_multiplier_msat: 10_000,
+                       historical_liquidity_penalty_amount_multiplier_msat: 64,
                        manual_node_penalties: HashMap::new(),
                        anti_probing_penalty_msat: 250,
                        considered_impossible_penalty_msat: 1_0000_0000_000,
@@ -718,14 +760,17 @@ impl<L: Deref<Target = u64>, BRT: Deref<Target = HistoricalBucketRangeTracker>,
        fn penalty_msat(&self, amount_msat: u64, params: &ProbabilisticScoringParameters) -> u64 {
                let max_liquidity_msat = self.max_liquidity_msat();
                let min_liquidity_msat = core::cmp::min(self.min_liquidity_msat(), max_liquidity_msat);
-               if amount_msat <= min_liquidity_msat {
+
+               let mut res = if amount_msat <= min_liquidity_msat {
                        0
                } else if amount_msat >= max_liquidity_msat {
                        // Equivalent to hitting the else clause below with the amount equal to the effective
                        // capacity and without any certainty on the liquidity upper bound, plus the
                        // impossibility penalty.
                        let negative_log10_times_2048 = NEGATIVE_LOG10_UPPER_BOUND * 2048;
-                       self.combined_penalty_msat(amount_msat, negative_log10_times_2048, params)
+                       Self::combined_penalty_msat(amount_msat, negative_log10_times_2048,
+                                       params.liquidity_penalty_multiplier_msat,
+                                       params.liquidity_penalty_amount_multiplier_msat)
                                .saturating_add(params.considered_impossible_penalty_msat)
                } else {
                        let numerator = (max_liquidity_msat - amount_msat).saturating_add(1);
@@ -738,25 +783,96 @@ impl<L: Deref<Target = u64>, BRT: Deref<Target = HistoricalBucketRangeTracker>,
                        } else {
                                let negative_log10_times_2048 =
                                        approx::negative_log10_times_2048(numerator, denominator);
-                               self.combined_penalty_msat(amount_msat, negative_log10_times_2048, params)
+                               Self::combined_penalty_msat(amount_msat, negative_log10_times_2048,
+                                       params.liquidity_penalty_multiplier_msat,
+                                       params.liquidity_penalty_amount_multiplier_msat)
+                       }
+               };
+
+               if params.historical_liquidity_penalty_multiplier_msat != 0 ||
+                  params.historical_liquidity_penalty_amount_multiplier_msat != 0 {
+                       // If historical penalties are enabled, calculate the penalty by walking the set of
+                       // historical liquidity bucket (min, max) combinations (where min_idx < max_idx)
+                       // and, for each, calculate the probability of success given our payment amount, then
+                       // total the weighted average probability of success.
+                       //
+                       // We use a sliding scale to decide which point within a given bucket will be compared
+                       // to the amount being sent - for lower-bounds, the amount being sent is compared to
+                       // the lower edge of the first bucket (i.e. zero), but compared to the upper 7/8ths of
+                       // the last bucket (i.e. 9 times the index, or 63), with each bucket in between
+                       // increasing the comparison point by 1/64th. For upper-bounds, the same applies,
+                       // however with an offset of 1/64th (i.e. starting at one and ending at 64). This
+                       // avoids failing to assign penalties to channels at the edges.
+                       //
+                       // If we used the bottom edge of buckets, we'd end up never assigning any penalty at
+                       // all to such a channel when sending less than ~0.19% of the channel's capacity (e.g.
+                       // ~200k sats for a 1 BTC channel!).
+                       //
+                       // If we used the middle of each bucket we'd never assign any penalty at all when
+                       // sending less than 1/16th of a channel's capacity, or 1/8th if we used the top of the
+                       // bucket.
+                       let mut total_valid_points_tracked = 0;
+                       for (min_idx, min_bucket) in self.min_liquidity_offset_history.buckets.iter().enumerate() {
+                               for max_bucket in self.max_liquidity_offset_history.buckets.iter().take(8 - min_idx) {
+                                       total_valid_points_tracked += (*min_bucket as u64) * (*max_bucket as u64);
+                               }
+                       }
+                       if total_valid_points_tracked == 0 {
+                               // If we don't have any valid points, redo the non-historical calculation with no
+                               // liquidity bounds tracked and the historical penalty multipliers.
+                               let max_capacity = self.capacity_msat.saturating_sub(amount_msat).saturating_add(1);
+                               let negative_log10_times_2048 =
+                                       approx::negative_log10_times_2048(max_capacity, self.capacity_msat.saturating_add(1));
+                               res = res.saturating_add(Self::combined_penalty_msat(amount_msat, negative_log10_times_2048,
+                                       params.historical_liquidity_penalty_multiplier_msat,
+                                       params.historical_liquidity_penalty_amount_multiplier_msat));
+                               return res;
                        }
+
+                       let payment_amt_64th_bucket = amount_msat * 64 / self.capacity_msat;
+                       debug_assert!(payment_amt_64th_bucket <= 64);
+                       if payment_amt_64th_bucket > 64 { return res; }
+
+                       let mut cumulative_success_prob_times_billion = 0;
+                       for (min_idx, min_bucket) in self.min_liquidity_offset_history.buckets.iter().enumerate() {
+                               for (max_idx, max_bucket) in self.max_liquidity_offset_history.buckets.iter().enumerate().take(8 - min_idx) {
+                                       let bucket_prob_times_million = (*min_bucket as u64) * (*max_bucket as u64)
+                                               * 1024 * 1024 / total_valid_points_tracked;
+                                       let min_64th_bucket = min_idx as u64 * 9;
+                                       let max_64th_bucket = (7 - max_idx as u64) * 9 + 1;
+                                       if payment_amt_64th_bucket > max_64th_bucket {
+                                               // Success probability 0, the payment amount is above the max liquidity
+                                       } else if payment_amt_64th_bucket <= min_64th_bucket {
+                                               cumulative_success_prob_times_billion += bucket_prob_times_million * 1024;
+                                       } else {
+                                               cumulative_success_prob_times_billion += bucket_prob_times_million *
+                                                       (max_64th_bucket - payment_amt_64th_bucket) * 1024 /
+                                                       (max_64th_bucket - min_64th_bucket);
+                                       }
+                               }
+                       }
+                       let historical_negative_log10_times_2048 = approx::negative_log10_times_2048(cumulative_success_prob_times_billion + 1, 1024 * 1024 * 1024);
+                       res = res.saturating_add(Self::combined_penalty_msat(amount_msat,
+                               historical_negative_log10_times_2048, params.historical_liquidity_penalty_multiplier_msat,
+                               params.historical_liquidity_penalty_amount_multiplier_msat));
                }
+
+               res
        }
 
        /// Computes the liquidity penalty from the penalty multipliers.
        #[inline(always)]
-       fn combined_penalty_msat(
-               &self, amount_msat: u64, negative_log10_times_2048: u64,
-               params: &ProbabilisticScoringParameters
+       fn combined_penalty_msat(amount_msat: u64, negative_log10_times_2048: u64,
+               liquidity_penalty_multiplier_msat: u64, liquidity_penalty_amount_multiplier_msat: u64,
        ) -> u64 {
                let liquidity_penalty_msat = {
                        // Upper bound the liquidity penalty to ensure some channel is selected.
-                       let multiplier_msat = params.liquidity_penalty_multiplier_msat;
+                       let multiplier_msat = liquidity_penalty_multiplier_msat;
                        let max_penalty_msat = multiplier_msat.saturating_mul(NEGATIVE_LOG10_UPPER_BOUND);
                        (negative_log10_times_2048.saturating_mul(multiplier_msat) / 2048).min(max_penalty_msat)
                };
                let amount_penalty_msat = negative_log10_times_2048
-                       .saturating_mul(params.liquidity_penalty_amount_multiplier_msat)
+                       .saturating_mul(liquidity_penalty_amount_multiplier_msat)
                        .saturating_mul(amount_msat) / 2048 / AMOUNT_PENALTY_DIVISOR;
 
                liquidity_penalty_msat.saturating_add(amount_penalty_msat)
@@ -2199,35 +2315,35 @@ mod tests {
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 3_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1985);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1983);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 4_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1639);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1637);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 5_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1607);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1606);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 6_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1262);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1331);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 7_450_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1262);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1387);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 7_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1262);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1379);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 8_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1262);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1363);
                let usage = ChannelUsage {
                        effective_capacity: EffectiveCapacity::Total { capacity_msat: 9_950_000_000, htlc_maximum_msat: Some(1_000) }, ..usage
                };
-               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1262);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 1355);
        }
 
        #[test]
@@ -2251,7 +2367,7 @@ mod tests {
 
                let params = ProbabilisticScoringParameters {
                        base_penalty_msat: 500, liquidity_penalty_multiplier_msat: 1_000,
-                       anti_probing_penalty_msat: 0, ..Default::default()
+                       anti_probing_penalty_msat: 0, ..ProbabilisticScoringParameters::zero_penalty()
                };
                let scorer = ProbabilisticScorer::new(params, &network_graph, &logger);
                assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 558);
@@ -2259,7 +2375,7 @@ mod tests {
                let params = ProbabilisticScoringParameters {
                        base_penalty_msat: 500, liquidity_penalty_multiplier_msat: 1_000,
                        base_penalty_amount_multiplier_msat: (1 << 30),
-                       anti_probing_penalty_msat: 0, ..Default::default()
+                       anti_probing_penalty_msat: 0, ..ProbabilisticScoringParameters::zero_penalty()
                };
 
                let scorer = ProbabilisticScorer::new(params, &network_graph, &logger);
@@ -2362,6 +2478,36 @@ mod tests {
                assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), u64::max_value());
        }
 
+       #[test]
+       fn remembers_historical_failures() {
+               let logger = TestLogger::new();
+               let network_graph = network_graph(&logger);
+               let params = ProbabilisticScoringParameters {
+                       historical_liquidity_penalty_multiplier_msat: 1024,
+                       historical_liquidity_penalty_amount_multiplier_msat: 1024,
+                       ..ProbabilisticScoringParameters::zero_penalty()
+               };
+               let mut scorer = ProbabilisticScorer::new(params, &network_graph, &logger);
+               let source = source_node_id();
+               let target = target_node_id();
+
+               let usage = ChannelUsage {
+                       amount_msat: 100,
+                       inflight_htlc_msat: 0,
+                       effective_capacity: EffectiveCapacity::Total { capacity_msat: 1_024, htlc_maximum_msat: Some(1_024) },
+               };
+               // With no historical data the normal liquidity penalty calculation is used.
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 47);
+
+               scorer.payment_path_failed(&payment_path_for_amount(1).iter().collect::<Vec<_>>(), 42);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 2048);
+
+               // Even after we tell the scorer we definitely have enough available liquidity, it will
+               // still remember that there was some failure in the past, and assign a non-0 penalty.
+               scorer.payment_path_failed(&payment_path_for_amount(1000).iter().collect::<Vec<_>>(), 43);
+               assert_eq!(scorer.channel_penalty_msat(42, &source, &target, usage), 198);
+       }
+
        #[test]
        fn adds_anti_probing_penalty() {
                let logger = TestLogger::new();