use std::collections::HashMap; use std::mem::take; use anyhow::Context; use itertools::Itertools; use log::debug; use ordered_float::OrderedFloat; use roaring::RoaringBitmap; use crate::facet::FacetType; use crate::search::criteria::{resolve_query_tree, CriteriaBuilder}; use crate::search::facet::FacetIter; use crate::search::query_tree::Operation; use crate::{FieldsIdsMap, FieldId, Index}; use super::{Criterion, CriterionParameters, CriterionResult}; /// Threshold on the number of candidates that will make /// the system to choose between one algorithm or another. const CANDIDATES_THRESHOLD: u64 = 1000; pub struct AscDesc<'t> { index: &'t Index, rtxn: &'t heed::RoTxn<'t>, field_name: String, field_id: FieldId, ascending: bool, query_tree: Option, candidates: Box> + 't>, bucket_candidates: RoaringBitmap, faceted_candidates: RoaringBitmap, parent: Box, } impl<'t> AscDesc<'t> { pub fn asc( index: &'t Index, rtxn: &'t heed::RoTxn, parent: Box, field_name: String, ) -> anyhow::Result { Self::new(index, rtxn, parent, field_name, true) } pub fn desc( index: &'t Index, rtxn: &'t heed::RoTxn, parent: Box, field_name: String, ) -> anyhow::Result { Self::new(index, rtxn, parent, field_name, false) } fn new( index: &'t Index, rtxn: &'t heed::RoTxn, parent: Box, field_name: String, ascending: bool, ) -> anyhow::Result { let fields_ids_map = index.fields_ids_map(rtxn)?; let field_id = fields_ids_map .id(&field_name) .with_context(|| format!("field {:?} isn't registered", field_name))?; Ok(AscDesc { index, rtxn, field_name, field_id, ascending, query_tree: None, candidates: Box::new(std::iter::empty()), faceted_candidates: index.number_faceted_documents_ids(rtxn, field_id)?, bucket_candidates: RoaringBitmap::new(), parent, }) } } impl<'t> Criterion for AscDesc<'t> { #[logging_timer::time("AscDesc::{}")] fn next(&mut self, params: &mut CriterionParameters) -> anyhow::Result> { loop { debug!( "Facet {}({}) iteration", if self.ascending { "Asc" } else { "Desc" }, self.field_name ); match self.candidates.next().transpose()? { None => { match self.parent.next(params)? { Some(CriterionResult { query_tree, candidates, filtered_candidates, bucket_candidates }) => { self.query_tree = query_tree; let mut candidates = match (&self.query_tree, candidates) { (_, Some(candidates)) => candidates & &self.faceted_candidates, (Some(qt), None) => { let context = CriteriaBuilder::new(&self.rtxn, &self.index)?; let candidates = resolve_query_tree(&context, qt, params.wdcache)?; candidates & &self.faceted_candidates }, (None, None) => take(&mut self.faceted_candidates), }; if let Some(filtered_candidates) = filtered_candidates { candidates &= filtered_candidates; } match bucket_candidates { Some(bucket_candidates) => self.bucket_candidates |= bucket_candidates, None => self.bucket_candidates |= &candidates, } if candidates.is_empty() { continue; } self.candidates = facet_ordered( self.index, self.rtxn, self.field_id, self.facet_type, self.ascending, candidates, )?; }, None => return Ok(None), } }, Some(mut candidates) => { candidates -= params.excluded_candidates; return Ok(Some(CriterionResult { query_tree: self.query_tree.clone(), candidates: Some(candidates), filtered_candidates: None, bucket_candidates: Some(take(&mut self.bucket_candidates)), })); } } } } } fn field_id_facet_type( fields_ids_map: &FieldsIdsMap, faceted_fields: &HashMap, field: &str, ) -> anyhow::Result<(FieldId, FacetType)> { let id = fields_ids_map .id(field) .with_context(|| format!("field {:?} isn't registered", field))?; let facet_type = faceted_fields .get(field) .with_context(|| format!("field {:?} isn't faceted", field))?; Ok((id, *facet_type)) } /// Returns an iterator over groups of the given candidates in ascending or descending order. /// /// It will either use an iterative or a recursive method on the whole facet database depending /// on the number of candidates to rank. fn facet_ordered<'t>( index: &'t Index, rtxn: &'t heed::RoTxn, field_id: FieldId, ascending: bool, candidates: RoaringBitmap, ) -> anyhow::Result> + 't>> { if candidates.len() <= CANDIDATES_THRESHOLD { let iter = iterative_facet_ordered_iter(index, rtxn, field_id, ascending, candidates)?; Ok(Box::new(iter.map(Ok)) as Box>) } else { let facet_fn = if ascending { FacetIter::new_reducing } else { FacetIter::new_reverse_reducing }; let iter = facet_fn(rtxn, index, field_id, candidates)?; Ok(Box::new(iter.map(|res| res.map(|(_, docids)| docids)))) } } /// Fetch the whole list of candidates facet values one by one and order them by it. /// /// This function is fast when the amount of candidates to rank is small. fn iterative_facet_ordered_iter<'t>( index: &'t Index, rtxn: &'t heed::RoTxn, field_id: FieldId, ascending: bool, candidates: RoaringBitmap, ) -> anyhow::Result + 't> { let mut docids_values = Vec::with_capacity(candidates.len() as usize); for docid in candidates.iter() { let left = (field_id, docid, f64::MIN); let right = (field_id, docid, f64::MAX); let mut iter = index .field_id_docid_facet_f64s .range(rtxn, &(left..=right))?; let entry = if ascending { iter.next() } else { iter.last() }; if let Some(((_, _, value), ())) = entry.transpose()? { docids_values.push((docid, OrderedFloat(value))); } } docids_values.sort_unstable_by_key(|(_, v)| *v); let iter = docids_values.into_iter(); let iter = if ascending { Box::new(iter) as Box> } else { Box::new(iter.rev()) }; // The itertools GroupBy iterator doesn't provide an owned version, we are therefore // required to collect the result into an owned collection (a Vec). // https://github.com/rust-itertools/itertools/issues/499 let vec: Vec<_> = iter .group_by(|(_, v)| v.clone()) .into_iter() .map(|(_, ids)| ids.map(|(id, _)| id).collect()) .collect(); Ok(vec.into_iter()) }