Introduce the Mdfs Iterator that explore the proximity graph using a mana DFS

This commit is contained in:
Kerollmops 2020-10-01 16:28:49 +02:00 committed by Clément Renault
parent d4e80407e5
commit 007e647462
No known key found for this signature in database
GPG Key ID: 92ADA4E935E71FA4
3 changed files with 172 additions and 123 deletions

View File

@ -1,4 +1,5 @@
mod criterion;
mod mdfs;
mod query_tokens;
mod search;
pub mod heed_codec;

158
src/mdfs.rs Normal file
View File

@ -0,0 +1,158 @@
use std::collections::hash_map::Entry::{Occupied, Vacant};
use std::collections::HashMap;
use std::mem;
use roaring::RoaringBitmap;
use crate::Index;
/// A mana depth first search implementation.
pub struct Mdfs<'a> {
index: &'a Index,
rtxn: &'a heed::RoTxn,
words: &'a [(HashMap<String, (u8, RoaringBitmap)>, RoaringBitmap)],
union_cache: HashMap<(usize, u8), RoaringBitmap>,
candidates: RoaringBitmap,
mana: u32,
max_mana: u32,
}
impl<'a> Mdfs<'a> {
pub fn new(
index: &'a Index,
rtxn: &'a heed::RoTxn,
words: &'a [(HashMap<String, (u8, RoaringBitmap)>, RoaringBitmap)],
candidates: RoaringBitmap,
) -> Mdfs<'a>
{
// Compute the number of pairs (windows) we have for this list of words.
let mana = words.len().checked_sub(1).unwrap_or(0) as u32;
let max_mana = mana * 8;
Mdfs { index, rtxn, words, union_cache: HashMap::new(), candidates, mana, max_mana }
}
}
impl<'a> Iterator for Mdfs<'a> {
type Item = anyhow::Result<RoaringBitmap>;
fn next(&mut self) -> Option<Self::Item> {
// If there is less or only one word therefore the only
// possible documents that we can return are the candidates.
if self.words.len() <= 1 {
if self.candidates.is_empty() { return None }
return Some(Ok(mem::take(&mut self.candidates)));
}
let mut answer = RoaringBitmap::new();
while self.mana <= self.max_mana {
let result = mdfs_step(
&self.index,
&self.rtxn,
self.mana,
self.words,
&self.candidates,
&self.candidates,
&mut self.union_cache,
);
match result {
Ok(Some(a)) => {
// We remove the answered documents from the list of
// candidates to be sure we don't search for them again.
self.candidates.difference_with(&a);
answer.union_with(&a);
},
Ok(None) => {
// We found the last iteration for this amount of mana that gives nothing,
// we can now store that the next mana to use for the loop is incremented.
self.mana = self.mana + 1;
// If the answer is empty it means that we found nothing for this amount
// of mana therefore we continue with a bigger mana.
if !answer.is_empty() {
// Otherwise we return the answer.
return Some(Ok(answer));
}
},
Err(e) => return Some(Err(e)),
}
}
None
}
}
fn mdfs_step(
index: &Index,
rtxn: &heed::RoTxn,
mana: u32,
words: &[(HashMap<String, (u8, RoaringBitmap)>, RoaringBitmap)],
candidates: &RoaringBitmap,
parent_docids: &RoaringBitmap,
union_cache: &mut HashMap<(usize, u8), RoaringBitmap>,
) -> anyhow::Result<Option<RoaringBitmap>>
{
use std::cmp::{min, max};
let (words1, words2) = (&words[0].0, &words[1].0);
let pairs = words_pair_combinations(words1, words2);
let tail = &words[1..];
let nb_children = tail.len() as u32 - 1;
// The minimum amount of mana that you must consume is at least 1 and the
// amount of mana that your children can consume. Because the last child must
// consume the remaining mana, it is mandatory that there not too much at the end.
let min_proximity = max(1, mana.saturating_sub(nb_children * 8)) as u8;
// The maximum amount of mana that you can use is 8 or the remaining amount of
// mana minus your children, as you can't just consume all the mana,
// your children must have at least 1 mana.
let max_proximity = min(8, mana - nb_children) as u8;
for proximity in min_proximity..=max_proximity {
let mut docids = match union_cache.entry((words.len(), proximity)) {
Occupied(entry) => entry.get().clone(),
Vacant(entry) => {
let mut docids = RoaringBitmap::new();
if proximity == 8 {
docids = candidates.clone();
} else {
for (w1, w2) in pairs.iter().cloned() {
let key = (w1, w2, proximity);
if let Some(di) = index.word_pair_proximity_docids.get(rtxn, &key)? {
docids.union_with(&di);
}
}
}
entry.insert(docids).clone()
}
};
docids.intersect_with(parent_docids);
if !docids.is_empty() {
let mana = mana.checked_sub(proximity as u32).unwrap();
// We are the last pair, we return without recursing as we don't have any child.
if tail.len() < 2 { return Ok(Some(docids)) }
if let Some(di) = mdfs_step(index, rtxn, mana, tail, candidates, &docids, union_cache)? {
return Ok(Some(di))
}
}
}
Ok(None)
}
fn words_pair_combinations<'h>(
w1: &'h HashMap<String, (u8, RoaringBitmap)>,
w2: &'h HashMap<String, (u8, RoaringBitmap)>,
) -> Vec<(&'h str, &'h str)>
{
let mut pairs = Vec::new();
for (w1, (_typos, docids1)) in w1 {
for (w2, (_typos, docids2)) in w2 {
if !docids1.is_disjoint(&docids2) {
pairs.push((w1.as_str(), w2.as_str()));
}
}
}
pairs
}

View File

@ -1,5 +1,4 @@
use std::collections::{HashMap, HashSet};
use std::collections::hash_map::Entry::{Occupied, Vacant};
use fst::{IntoStreamer, Streamer};
use levenshtein_automata::DFA;
@ -9,6 +8,7 @@ use once_cell::sync::Lazy;
use roaring::bitmap::RoaringBitmap;
use crate::query_tokens::{QueryTokens, QueryToken};
use crate::mdfs::Mdfs;
use crate::{Index, DocumentId};
// Building these factories is not free.
@ -132,111 +132,6 @@ impl<'a> Search<'a> {
candidates
}
// TODO Move this elsewhere!
fn mana_depth_first_search(
&self,
words: &[(HashMap<String, (u8, RoaringBitmap)>, RoaringBitmap)],
candidates: &RoaringBitmap,
union_cache: &mut HashMap<(usize, u8), RoaringBitmap>,
) -> anyhow::Result<Option<RoaringBitmap>>
{
fn words_pair_combinations<'h>(
w1: &'h HashMap<String, (u8, RoaringBitmap)>,
w2: &'h HashMap<String, (u8, RoaringBitmap)>,
) -> Vec<(&'h str, &'h str)>
{
let mut pairs = Vec::new();
for (w1, (_typos, docids1)) in w1 {
for (w2, (_typos, docids2)) in w2 {
if !docids1.is_disjoint(&docids2) {
pairs.push((w1.as_str(), w2.as_str()));
}
}
}
pairs
}
fn mdfs(
index: &Index,
rtxn: &heed::RoTxn,
mana: u32,
words: &[(HashMap<String, (u8, RoaringBitmap)>, RoaringBitmap)],
candidates: &RoaringBitmap,
parent_docids: &RoaringBitmap,
union_cache: &mut HashMap<(usize, u8), RoaringBitmap>,
) -> anyhow::Result<Option<RoaringBitmap>>
{
use std::cmp::{min, max};
let (words1, words2) = (&words[0].0, &words[1].0);
let pairs = words_pair_combinations(words1, words2);
let tail = &words[1..];
let nb_children = tail.len() as u32 - 1;
// The minimum amount of mana that you must consume is at least 1 and the
// amount of mana that your children can consume. Because the last child must
// consume the remaining mana, it is mandatory that there not too much at the end.
let min_proximity = max(1, mana.saturating_sub(nb_children * 8)) as u8;
// The maximum amount of mana that you can use is 8 or the remaining amount of
// mana minus your children, as you can't just consume all the mana,
// your children must have at least 1 mana.
let max_proximity = min(8, mana - nb_children) as u8;
for proximity in min_proximity..=max_proximity {
let mut docids = match union_cache.entry((words.len(), proximity)) {
Occupied(entry) => entry.get().clone(),
Vacant(entry) => {
let mut docids = RoaringBitmap::new();
if proximity == 8 {
docids = candidates.clone();
} else {
for (w1, w2) in pairs.iter().cloned() {
let key = (w1, w2, proximity);
if let Some(di) = index.word_pair_proximity_docids.get(rtxn, &key)? {
docids.union_with(&di);
}
}
}
entry.insert(docids).clone()
}
};
docids.intersect_with(parent_docids);
if !docids.is_empty() {
let mana = mana.checked_sub(proximity as u32).unwrap();
// We are the last pair, we return without recursing as we don't have any child.
if tail.len() < 2 { return Ok(Some(docids)) }
if let Some(di) = mdfs(index, rtxn, mana, tail, candidates, &docids, union_cache)? {
return Ok(Some(di))
}
}
}
Ok(None)
}
// Compute the number of pairs (windows) we have for this list of words.
// If there only is one word therefore the only possible documents are the candidates.
let initial_mana = match words.len().checked_sub(1) {
Some(nb_windows) if nb_windows != 0 => nb_windows as u32,
_ => return Ok(Some(candidates.clone())),
};
// TODO We must keep track of where we are in terms of mana and that should either be
// handled by an Iterator or by the caller. Keeping track of the amount of mana
// is an optimization, it makes this mdfs to only be called with the next valid
// mana and not called with all of the previous mana values.
for mana in initial_mana..=initial_mana * 8 {
if let Some(answer) = mdfs(&self.index, &self.rtxn, mana, words, candidates, candidates, union_cache)? {
return Ok(Some(answer));
}
}
Ok(None)
}
pub fn execute(&self) -> anyhow::Result<SearchResult> {
let limit = self.limit;
@ -257,29 +152,24 @@ impl<'a> Search<'a> {
}
let derived_words = self.fetch_words_docids(&fst, dfas)?;
let mut candidates = Self::compute_candidates(&derived_words);
let candidates = Self::compute_candidates(&derived_words);
debug!("candidates: {:?}", candidates);
// The mana depth first search is a revised DFS that explore
// solutions in the order of their proximities.
let mut mdfs = Mdfs::new(self.index, self.rtxn, &derived_words, candidates);
let mut documents = Vec::new();
let mut union_cache = HashMap::new();
// We execute the DFS until we find enough documents, we run it with the
// candidates list and remove the found documents from this list at each iteration.
// We execute the Mdfs iterator until we find enough documents.
while documents.iter().map(RoaringBitmap::len).sum::<u64>() < limit as u64 {
let answer = self.mana_depth_first_search(&derived_words, &candidates, &mut union_cache)?;
let answer = match answer {
Some(answer) if !answer.is_empty() => answer,
_ => break,
};
match mdfs.next().transpose()? {
Some(answer) => {
debug!("answer: {:?}", answer);
// We remove the answered documents from the list of
// candidates to be sure we don't search for them again.
candidates.difference_with(&answer);
documents.push(answer);
},
None => break,
}
}
let found_words = derived_words.into_iter().flat_map(|(w, _)| w).map(|(w, _)| w).collect();