Skip to content

perf(fts): remove OR hot path bookkeeping #7416

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
BubbleCal merged 1 commit into from
Jun 24, 2026
Merged

perf(fts): remove OR hot path bookkeeping #7416

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
136 changes: 85 additions & 51 deletions rust/lance-index/src/scalar/inverted/index.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -769,6 +769,33 @@ impl InvertedIndex {
if limit == 0 {
return Ok((Vec::new(), Vec::new()));
}

fn push_scored_candidate(
candidates: &mut BinaryHeap<Reverse<ScoredDoc>>,
limit: usize,
addr: CandidateAddr,
score: f32,
) -> Result<()> {
// resolve_deferred_candidates ran upstream, so every candidate
// carries a real row_id at this point.
let row_id = match addr {
CandidateAddr::RowId(r) => r,
CandidateAddr::Pending(_) => {
return Err(Error::internal(
"bm25_search post-condition: deferred candidate left unresolved",
));
}
};

if candidates.len() < limit {
candidates.push(Reverse(ScoredDoc::new(row_id, score)));
} else if candidates.peek().unwrap().0.score.0 < score {
candidates.pop();
candidates.push(Reverse(ScoredDoc::new(row_id, score)));
}
Ok(())
}

let mask = prefilter.mask();

let mut candidates = BinaryHeap::new();
Expand Down Expand Up @@ -870,10 +897,6 @@ impl InvertedIndex {
grouped_expansions,
candidates: part_candidates,
} = res;
let grouped_positions = grouped_expansions
.iter()
.map(|group| group.position)
.collect::<HashSet<_>>();
let mut idf_by_position = Vec::with_capacity(tokens_by_position.len());
for token in &tokens_by_position {
let idf_weight = match idf_cache.get(token) {
Expand All @@ -886,53 +909,61 @@ impl InvertedIndex {
};
idf_by_position.push(idf_weight);
}
for DocCandidate {
addr,
posting_doc_id,
freqs,
doc_length,
} in part_candidates
{
// resolve_deferred_candidates ran upstream, so every
// candidate carries a real row_id at this point.
let row_id = match addr {
CandidateAddr::RowId(r) => r,
CandidateAddr::Pending(_) => {
return Err(Error::internal(
"bm25_search post-condition: deferred candidate left unresolved",
));
}
};
let mut score = 0.0;
for (term_index, freq) in freqs.into_iter() {
if grouped_positions.contains(&term_index) {
continue;

if grouped_expansions.is_empty() {
for DocCandidate {
addr,
freqs,
doc_length,
..
} in part_candidates
{
let mut score = 0.0;
for (term_index, freq) in freqs.into_iter() {
debug_assert!((term_index as usize) < idf_by_position.len());
score += idf_by_position[term_index as usize]
* scorer.doc_weight(freq, doc_length);
}
debug_assert!((term_index as usize) < idf_by_position.len());
score +=
idf_by_position[term_index as usize] * scorer.doc_weight(freq, doc_length);
push_scored_candidate(&mut candidates, limit, addr, score)?;
}
for group in &grouped_expansions {
for term in &group.terms {
let Some(freq) = term.frequency(posting_doc_id) else {
} else {
let grouped_positions = grouped_expansions
.iter()
.map(|group| group.position)
.collect::<HashSet<_>>();
for DocCandidate {
addr,
posting_doc_id,
freqs,
doc_length,
} in part_candidates
{
let mut score = 0.0;
for (term_index, freq) in freqs.into_iter() {
if grouped_positions.contains(&term_index) {
continue;
};
let idf_weight = match idf_cache.get(&term.token) {
Some(weight) => *weight,
None => {
let weight = scorer.query_weight(&term.token);
idf_cache.insert(term.token.clone(), weight);
weight
}
};
score += idf_weight * scorer.doc_weight(freq, doc_length);
}
debug_assert!((term_index as usize) < idf_by_position.len());
score += idf_by_position[term_index as usize]
* scorer.doc_weight(freq, doc_length);
}
}
if candidates.len() < limit {
candidates.push(Reverse(ScoredDoc::new(row_id, score)));
} else if candidates.peek().unwrap().0.score.0 < score {
candidates.pop();
candidates.push(Reverse(ScoredDoc::new(row_id, score)));
for group in &grouped_expansions {
for term in &group.terms {
let Some(freq) = term.frequency(posting_doc_id) else {
continue;
};
let idf_weight = match idf_cache.get(&term.token) {
Some(weight) => *weight,
None => {
let weight = scorer.query_weight(&term.token);
idf_cache.insert(term.token.clone(), weight);
weight
}
};
score += idf_weight * scorer.doc_weight(freq, doc_length);
}
}
push_scored_candidate(&mut candidates, limit, addr, score)?;
}
}
}
Expand Down Expand Up @@ -1587,12 +1618,14 @@ impl InvertedPartition {
.map(|index| tokens.position(index))
.collect::<Vec<_>>();
let mut token_ids = Vec::with_capacity(tokens.len());
let mut matched_positions = HashSet::new();
let mut matched_positions = required_positions.as_ref().map(|_| HashSet::new());
for (index, token) in tokens.into_iter().enumerate() {
let token_id = self.map(&token);
if let Some(token_id) = token_id {
let position = token_positions[index];
matched_positions.insert(position);
if let Some(matched_positions) = matched_positions.as_mut() {
matched_positions.insert(position);
}
token_ids.push((token_id, token, position));
} else if is_phrase_query || is_and_query {
// if the token is not found, we can't do phrase or AND query
Expand All @@ -1602,8 +1635,9 @@ impl InvertedPartition {
if token_ids.is_empty() {
return Ok(LoadedPostings::empty());
}
if let Some(required_positions) = required_positions
&& !required_positions.is_subset(&matched_positions)
if let Some(required_positions) = required_positions.as_ref()
&& let Some(matched_positions) = matched_positions.as_ref()
&& !required_positions.is_subset(matched_positions)
{
return Ok(LoadedPostings::empty());
}
Expand Down
126 changes: 108 additions & 18 deletions rust/lance-index/src/scalar/inverted/wand.rs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -659,6 +659,14 @@ struct AndWindowStats {
candidates_returned: usize,
}

#[derive(Default)]
struct AndSearchStats {
pruned_before_return_start: usize,
candidates_seen: usize,
full_scores: usize,
freqs_collected: usize,
}

impl Eq for TailPosting {}

impl PartialOrd for TailPosting {
Expand Down Expand Up @@ -842,18 +850,18 @@ impl<'a, S: Scorer> Wand<'a, S> {

let mut candidates = BinaryHeap::with_capacity(std::cmp::min(limit, BLOCK_SIZE * 10));
let mut num_comparisons = 0;
let mut and_candidates_seen = 0;
let mut and_full_scores = 0;
let mut freqs_collected = 0;
let pruned_before_return_start = self.and_candidates_pruned_before_return;
let mut and_search_stats = (self.operator == Operator::And).then_some(AndSearchStats {
pruned_before_return_start: self.and_candidates_pruned_before_return,
..Default::default()
});
loop {
self.raise_to_shared_floor(params.wand_factor);
let Some((doc, mut score)) = self.next()? else {
break;
};
num_comparisons += 1;
if self.operator == Operator::And {
and_candidates_seen += 1;
if let Some(and_stats) = and_search_stats.as_mut() {
and_stats.candidates_seen += 1;
}

// Either a real row_id (so we can run the mask check
Expand Down Expand Up @@ -907,14 +915,16 @@ impl<'a, S: Scorer> Wand<'a, S> {
{
continue;
}
and_full_scores += 1;
if let Some(and_stats) = and_search_stats.as_mut() {
and_stats.full_scores += 1;
}
self.score(doc_length)
};

if candidates.len() < limit {
let freqs = self.iter_term_freqs().collect();
if self.operator == Operator::And {
freqs_collected += 1;
if let Some(and_stats) = and_search_stats.as_mut() {
and_stats.freqs_collected += 1;
}
candidates.push(Reverse((
ScoredDoc::new(row_id, score),
Expand All @@ -928,8 +938,8 @@ impl<'a, S: Scorer> Wand<'a, S> {
}
} else if score > candidates.peek().unwrap().0.0.score.0 {
let freqs = self.iter_term_freqs().collect();
if self.operator == Operator::And {
freqs_collected += 1;
if let Some(and_stats) = and_search_stats.as_mut() {
and_stats.freqs_collected += 1;
}
candidates.pop();
candidates.push(Reverse((
Expand All @@ -956,13 +966,15 @@ impl<'a, S: Scorer> Wand<'a, S> {
);
}
metrics.record_comparisons(num_comparisons);
let and_candidates_pruned_before_return = self
.and_candidates_pruned_before_return
.saturating_sub(pruned_before_return_start);
metrics.record_and_candidates_seen(and_candidates_seen);
metrics.record_and_candidates_pruned_before_return(and_candidates_pruned_before_return);
metrics.record_and_full_scores(and_full_scores);
metrics.record_freqs_collected(freqs_collected);
if let Some(and_stats) = and_search_stats {
let and_candidates_pruned_before_return = self
.and_candidates_pruned_before_return
.saturating_sub(and_stats.pruned_before_return_start);
metrics.record_and_candidates_seen(and_stats.candidates_seen);
metrics.record_and_candidates_pruned_before_return(and_candidates_pruned_before_return);
metrics.record_and_full_scores(and_stats.full_scores);
metrics.record_freqs_collected(and_stats.freqs_collected);
}

// The heap entry's `row_id` slot is either a real row_id
// (DocSet had row_ids) or the doc_id widened to u64
Expand Down Expand Up @@ -2009,6 +2021,44 @@ mod tests {
}
}

struct PanicOnAndMetrics {
comparisons: AtomicUsize,
}

impl PanicOnAndMetrics {
fn new() -> Self {
Self {
comparisons: AtomicUsize::new(0),
}
}
}

impl MetricsCollector for PanicOnAndMetrics {
fn record_parts_loaded(&self, _: usize) {}

fn record_index_loads(&self, _: usize) {}

fn record_comparisons(&self, n: usize) {
self.comparisons.fetch_add(n, Ordering::Relaxed);
}

fn record_and_candidates_seen(&self, _: usize) {
panic!("OR search should not record AND candidate metrics");
}

fn record_and_candidates_pruned_before_return(&self, _: usize) {
panic!("OR search should not record AND prune metrics");
}

fn record_and_full_scores(&self, _: usize) {
panic!("OR search should not record AND scoring metrics");
}

fn record_freqs_collected(&self, _: usize) {
panic!("OR search should not record AND frequency metrics");
}
}

fn generate_posting_list(
doc_ids: Vec<u32>,
max_score: f32,
Expand Down Expand Up @@ -2351,6 +2401,46 @@ mod tests {
);
}

#[rstest]
fn test_or_search_does_not_record_and_metrics(#[values(false, true)] is_compressed: bool) {
let mut docs = DocSet::default();
for row_id in 0..6 {
docs.append(row_id, 1);
}

let postings = vec![
PostingIterator::with_query_weight(
String::from("alpha"),
0,
0,
1.0,
generate_posting_list(vec![0, 1, 4], 1.0, None, is_compressed),
docs.len(),
),
PostingIterator::with_query_weight(
String::from("beta"),
1,
1,
1.0,
generate_posting_list(vec![1, 2, 5], 1.0, None, is_compressed),
docs.len(),
),
];

let mut wand = Wand::new(Operator::Or, postings.into_iter(), &docs, UnitScorer);
let metrics = PanicOnAndMetrics::new();
let candidates = wand
.search(
&FtsSearchParams::default(),
Arc::new(RowAddrMask::default()),
&metrics,
)
.unwrap();

assert_eq!(sorted_candidate_row_ids(candidates), vec![0, 1, 2, 4, 5]);
assert!(metrics.comparisons.load(Ordering::Relaxed) > 0);
}

#[test]
fn test_wand_new_uses_precomputed_query_weight() {
let mut docs = DocSet::default();
Expand Down
Loading