diff --git a/indexer-selection/src/performance.rs b/indexer-selection/src/performance.rs
index 7943f0d..045886d 100644
--- a/indexer-selection/src/performance.rs
+++ b/indexer-selection/src/performance.rs
@@ -2,8 +2,8 @@ use candidate_selection::Normalized;
 
 #[derive(Clone, Debug, Default)]
 pub struct Performance {
-    fast: Frame,
-    slow: Frame,
+    fast: ShortTerm,
+    slow: LongTerm,
 }
 
 const FAST_BIAS: f64 = 0.8;
@@ -25,14 +25,8 @@ impl Performance {
     }
 
     pub fn feedback(&mut self, success: bool, latency_ms: u16) {
-        for frame in [&mut self.fast, &mut self.slow] {
-            frame.total_latency_ms += latency_ms as f64;
-            if success {
-                frame.success_count += 1.0;
-            } else {
-                frame.failure_count += 1.0;
-            }
-        }
+        self.fast.feedback(success, latency_ms);
+        self.slow.feedback(success, latency_ms);
     }
 
     pub fn decay(&mut self) {
@@ -41,26 +35,28 @@ impl Performance {
     }
 
     fn success_rate(&self) -> Normalized {
-        let fast = self.fast.success_rate().as_f64();
-        let slow = self.slow.success_rate().as_f64();
-        Normalized::new((fast * FAST_BIAS) + (slow * (1.0 - FAST_BIAS))).unwrap_or(Normalized::ONE)
+        let fast = self.fast.success_rate();
+        let slow = self.slow.success_rate();
+        let success_rate = (fast * FAST_BIAS) + (slow * (1.0 - FAST_BIAS));
+        // limit an individual indexer's success rate to 99%
+        Normalized::new(success_rate.min(0.99)).unwrap()
     }
 
     fn latency_ms(&self) -> u16 {
         let fast = self.fast.latency_ms() as f64;
-        let slow = self.slow.latency_ms() as f64;
+        let slow = self.slow.latency_percentile(99) as f64;
         ((fast * FAST_BIAS) + (slow * (1.0 - FAST_BIAS))) as u16
     }
 }
 
 #[derive(Clone, Debug, Default)]
-struct Frame {
+struct ShortTerm {
     total_latency_ms: f64,
     success_count: f64,
     failure_count: f64,
 }
 
-impl Frame {
+impl ShortTerm {
     fn decay(&mut self, rate_hz: f64) {
         debug_assert!((0.0 < rate_hz) && (rate_hz < 1.0));
         let retain = 1.0 - rate_hz;
@@ -69,13 +65,20 @@ impl Frame {
         self.failure_count *= retain;
     }
 
-    fn success_rate(&self) -> Normalized {
-        // add 1 to pull success rate upward
+    fn feedback(&mut self, success: bool, latency_ms: u16) {
+        self.total_latency_ms += latency_ms as f64;
+        if success {
+            self.success_count += 1.0;
+        } else {
+            self.failure_count += 1.0;
+        }
+    }
+
+    fn success_rate(&self) -> f64 {
+        // add 1 to pull success rate upward, and avoid divide by zero
         let s = self.success_count + 1.0;
         let f = self.failure_count;
-        let p = s / (s + f);
-        // limit an individual indexer's success rate to 99%
-        Normalized::new(p.min(0.99)).unwrap()
+        s / (s + f)
     }
 
     fn latency_ms(&self) -> u16 {
@@ -84,3 +87,71 @@ impl Frame {
         avg_latency_ms as u16
     }
 }
+
+#[derive(Clone, Debug, Default)]
+struct LongTerm {
+    latency_hist: [f32; 29],
+    failure_count: f64,
+}
+
+const LATENCY_BINS: [u16; 29] = [
+    32, 64, 96, 128, // + 2^5
+    192, 256, 320, 384, // + 2^6
+    512, 640, 768, 896, // + 2^7
+    1152, 1408, 1664, 1920, // + 2^8
+    2432, 2944, 3456, 3968, // + 2^9
+    4992, 6016, 7040, 8064, // + 2^10
+    10112, 12160, 14208, 16256, // + 2^11
+    20352, // + 2^12
+];
+
+impl LongTerm {
+    fn decay(&mut self, rate_hz: f64) {
+        debug_assert!((0.0 < rate_hz) && (rate_hz < 1.0));
+        let retain = 1.0 - rate_hz;
+        self.failure_count *= retain;
+        for count in &mut self.latency_hist {
+            *count *= retain as f32;
+        }
+    }
+
+    fn feedback(&mut self, success: bool, latency_ms: u16) {
+        if !success {
+            self.failure_count += 1.0;
+        }
+
+        for (count, bin_value) in self
+            .latency_hist
+            .iter_mut()
+            .zip(&LATENCY_BINS)
+            .take(LATENCY_BINS.len() - 1)
+        {
+            if latency_ms <= *bin_value {
+                *count += 1.0;
+                return;
+            }
+        }
+        *self.latency_hist.last_mut().unwrap() += 1.0;
+    }
+
+    fn success_rate(&self) -> f64 {
+        // add 1 to pull success rate upward, and avoid divide by zero
+        let total = self.latency_hist.iter().map(|c| *c as f64).sum::<f64>() + 1.0;
+        let s = total - self.failure_count;
+        let f = self.failure_count;
+        s / (s + f)
+    }
+
+    pub fn latency_percentile(&self, p: u8) -> u16 {
+        debug_assert!((1..=99).contains(&p));
+        let target = self.latency_hist.iter().map(|c| *c as f64).sum::<f64>() * (p as f64 / 100.0);
+        let mut sum = 0.0;
+        for (count, bin_value) in self.latency_hist.iter().zip(&LATENCY_BINS) {
+            sum += *count as f64;
+            if sum >= target {
+                return *bin_value;
+            }
+        }
+        panic!("failed to calculate latency percentile");
+    }
+}