From b31b46da74b415fb25ed854fbfdf7ed5ae165d11 Mon Sep 17 00:00:00 2001
From: David Cromberge <davecromberge@gmail.com>
Date: Tue, 10 Nov 2026 11:08:57 +0000
Subject: [PATCH] perf(theta,tuple): hash a-not-b keys with a seeded folded
 multiply

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 CHANGELOG.md                                  |  1 +
 .../src/thetafamily/common/a_not_b.rs         | 44 ++++++++++++++++++-
 tests-integration/tests/theta_test/a_not_b.rs | 25 +++++++++++
 3 files changed, 69 insertions(+), 1 deletion(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 36602987..8a3fc882 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -22,6 +22,7 @@ All significant changes to this project will be documented in this file.
 * `BloomFilter::insert` is faster for streams of new items when multiple hashes are configured.
 * Compressed `CompactThetaSketch` serialization and deserialization are faster.
 * Tuple sketch serialization and deserialization are faster.
+* `ThetaANotB` and `TupleANotB` are faster when either input is unordered.
 
 ## v0.6.0 (2026-10-29)
 
diff --git a/datasketches/src/thetafamily/common/a_not_b.rs b/datasketches/src/thetafamily/common/a_not_b.rs
index d1caeea5..8bfb41d5 100644
--- a/datasketches/src/thetafamily/common/a_not_b.rs
+++ b/datasketches/src/thetafamily/common/a_not_b.rs
@@ -16,6 +16,8 @@
 // under the License.
 
 use std::collections::HashSet;
+use std::hash::BuildHasher;
+use std::hash::Hasher;
 
 use crate::error::Error;
 use crate::error::ErrorKind;
@@ -106,7 +108,8 @@ where
         }
         entries
     } else {
-        let mut b_keys: HashSet<u64> = HashSet::with_capacity(b_num_retained);
+        let mut b_keys: HashSet<u64, RetainedHashState> =
+            HashSet::with_capacity_and_hasher(b_num_retained, RetainedHashState::new());
         for hash in b.hashes() {
             if hash < theta {
                 b_keys.insert(hash);
@@ -170,3 +173,42 @@ where
     let out_ordered = ordered || input_ordered || (entries.len() == 1 && theta == MAX_THETA);
     CompactSketchState::non_empty(entries, theta, seed_hash, out_ordered)
 }
+
+struct RetainedHashState(u64);
+
+impl RetainedHashState {
+    fn new() -> Self {
+        Self(std::collections::hash_map::RandomState::new().hash_one(0u64))
+    }
+}
+
+impl BuildHasher for RetainedHashState {
+    type Hasher = RetainedHashHasher;
+
+    fn build_hasher(&self) -> RetainedHashHasher {
+        RetainedHashHasher {
+            seed: self.0,
+            hash: 0,
+        }
+    }
+}
+
+struct RetainedHashHasher {
+    seed: u64,
+    hash: u64,
+}
+
+impl Hasher for RetainedHashHasher {
+    fn finish(&self) -> u64 {
+        self.hash
+    }
+
+    fn write(&mut self, _bytes: &[u8]) {
+        unreachable!("retained hashes are hashed as u64");
+    }
+
+    fn write_u64(&mut self, hash: u64) {
+        let product = u128::from(hash ^ self.seed) * 0x9E37_79B9_7F4A_7C15;
+        self.hash = (product as u64) ^ ((product >> 64) as u64);
+    }
+}
diff --git a/tests-integration/tests/theta_test/a_not_b.rs b/tests-integration/tests/theta_test/a_not_b.rs
index ff8e64c0..e22d4caf 100644
--- a/tests-integration/tests/theta_test/a_not_b.rs
+++ b/tests-integration/tests/theta_test/a_not_b.rs
@@ -263,3 +263,28 @@ fn test_estimation_disjoint_returns_a() {
     assert!(r.is_estimation_mode());
     assert_that!(r.estimate(), near(10000.0, 10000.0 * 0.02));
 }
+
+fn unordered_sketch_with_hashes(hashes: &[u64]) -> CompactThetaSketch {
+    let mut bytes = sketch_with_range(0, hashes.len() as u64)
+        .compact(false)
+        .serialize();
+    let offset = bytes.len() - hashes.len() * 8;
+    for (slot, hash) in bytes[offset..].chunks_exact_mut(8).zip(hashes) {
+        slot.copy_from_slice(&hash.to_le_bytes());
+    }
+    CompactThetaSketch::deserialize(&bytes).unwrap()
+}
+
+#[test]
+fn test_unordered_hashes_sharing_low_bits() {
+    let hash = |k: u64| (k << 32) | 0x5A5A_5A5A;
+    let a: Vec<u64> = (1..=3000).map(hash).collect();
+    let b: Vec<u64> = (1001..=4000).map(hash).collect();
+    let a = unordered_sketch_with_hashes(&a);
+    let b = unordered_sketch_with_hashes(&b);
+    assert!(!a.is_ordered() && !b.is_ordered());
+
+    let r = ThetaANotB::default().compute(&a, &b, true).unwrap();
+    let retained: Vec<u64> = r.iter().map(|entry| entry.hash()).collect();
+    assert_eq!(retained, (1..=1000).map(hash).collect::<Vec<_>>());
+}