MeiliSearch/milli/src/search/new/sort.rs

223 lines
8.6 KiB
Rust
Raw Normal View History

2023-06-15 17:36:40 +02:00
use heed::BytesDecode;
2023-03-08 09:55:53 +01:00
use roaring::RoaringBitmap;
2023-02-22 15:34:37 +01:00
use super::logger::SearchLogger;
2023-03-09 11:12:31 +01:00
use super::{RankingRule, RankingRuleOutput, RankingRuleQueryTrait, SearchContext};
2023-06-15 17:36:40 +02:00
use crate::heed_codec::facet::{FacetGroupKeyCodec, OrderedF64Codec};
2023-11-27 11:52:22 +01:00
use crate::heed_codec::{BytesRefCodec, StrRefCodec};
2023-06-15 17:36:40 +02:00
use crate::score_details::{self, ScoreDetails};
2023-03-20 09:30:10 +01:00
use crate::search::facet::{ascending_facet_sort, descending_facet_sort};
use crate::{FieldId, Index, Result};
2023-03-09 11:12:31 +01:00
2023-03-13 14:03:48 +01:00
pub trait RankingRuleOutputIter<'ctx, Query> {
2023-03-09 11:12:31 +01:00
fn next_bucket(&mut self) -> Result<Option<RankingRuleOutput<Query>>>;
}
2023-03-13 14:03:48 +01:00
pub struct RankingRuleOutputIterWrapper<'ctx, Query> {
iter: Box<dyn Iterator<Item = Result<RankingRuleOutput<Query>>> + 'ctx>,
2023-03-09 11:12:31 +01:00
}
2023-03-13 14:03:48 +01:00
impl<'ctx, Query> RankingRuleOutputIterWrapper<'ctx, Query> {
pub fn new(iter: Box<dyn Iterator<Item = Result<RankingRuleOutput<Query>>> + 'ctx>) -> Self {
2023-03-09 11:12:31 +01:00
Self { iter }
}
}
2023-03-13 14:03:48 +01:00
impl<'ctx, Query> RankingRuleOutputIter<'ctx, Query> for RankingRuleOutputIterWrapper<'ctx, Query> {
2023-03-09 11:12:31 +01:00
fn next_bucket(&mut self) -> Result<Option<RankingRuleOutput<Query>>> {
match self.iter.next() {
Some(x) => x.map(Some),
None => Ok(None),
}
}
}
2023-03-27 11:04:04 +02:00
// `Query` type parameter: the same as the type parameter to bucket_sort
// implements RankingRuleQuery trait, either querygraph or placeholdersearch
// The sort ranking rule doesn't need the query parameter, it is doing the same thing
// whether we're doing a querygraph or placeholder search.
//
// Query Stored anyway because every ranking rule must return a query from next_bucket
// ---
// "Mismatch" between new/old impl.:
// - old impl: roaring bitmap as input, ranking rule iterates other all the buckets
// - new impl: still works like that but it shouldn't, because the universe may change for every call to next_bucket, itself due to:
// 1. elements that were already returned by the ranking rule are subtracted from the universe, also done in the old impl (subtracted from the candidates)
// 2. NEW in the new impl.: distinct rule might have been applied btwn calls to next_bucket
// new impl ignores docs removed in (2), which is a missed perf opt issue, see `next_bucket`
// this perf problem is P2
// mostly happens when many documents map to the same distinct attribute value.
2023-03-13 14:03:48 +01:00
pub struct Sort<'ctx, Query> {
2023-02-22 15:34:37 +01:00
field_name: String,
field_id: Option<FieldId>,
is_ascending: bool,
original_query: Option<Query>,
2023-03-13 14:03:48 +01:00
iter: Option<RankingRuleOutputIterWrapper<'ctx, Query>>,
2023-06-15 17:36:40 +02:00
must_redact: bool,
}
2023-03-13 14:03:48 +01:00
impl<'ctx, Query> Sort<'ctx, Query> {
pub fn new(
index: &Index,
2023-03-13 14:03:48 +01:00
rtxn: &'ctx heed::RoTxn,
field_name: String,
is_ascending: bool,
) -> Result<Self> {
let fields_ids_map = index.fields_ids_map(rtxn)?;
let field_id = fields_ids_map.id(&field_name);
2023-06-15 17:36:40 +02:00
let must_redact = Self::must_redact(index, rtxn, &field_name)?;
2023-06-15 17:36:40 +02:00
Ok(Self {
field_name,
field_id,
is_ascending,
original_query: None,
iter: None,
must_redact,
})
}
fn must_redact(index: &Index, rtxn: &'ctx heed::RoTxn, field_name: &str) -> Result<bool> {
2023-07-03 10:20:28 +02:00
let Some(displayed_fields) = index.displayed_fields(rtxn)? else {
return Ok(false);
};
2023-06-15 17:36:40 +02:00
Ok(!displayed_fields.iter().any(|&field| field == field_name))
}
}
2023-03-13 14:03:48 +01:00
impl<'ctx, Query: RankingRuleQueryTrait> RankingRule<'ctx, Query> for Sort<'ctx, Query> {
2023-02-22 15:34:37 +01:00
fn id(&self) -> String {
let Self { field_name, is_ascending, .. } = self;
2023-06-06 18:21:31 +02:00
format!("{field_name}:{}", if *is_ascending { "asc" } else { "desc" })
2023-02-22 15:34:37 +01:00
}
fn start_iteration(
&mut self,
2023-03-13 14:03:48 +01:00
ctx: &mut SearchContext<'ctx>,
2023-02-22 15:34:37 +01:00
_logger: &mut dyn SearchLogger<Query>,
parent_candidates: &RoaringBitmap,
2023-03-20 09:30:10 +01:00
parent_query: &Query,
) -> Result<()> {
let iter: RankingRuleOutputIterWrapper<Query> = match self.field_id {
Some(field_id) => {
let number_db = ctx
.index
.facet_id_f64_docids
2023-11-27 11:52:22 +01:00
.remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>();
let string_db = ctx
.index
.facet_id_string_docids
2023-11-27 11:52:22 +01:00
.remap_key_type::<FacetGroupKeyCodec<BytesRefCodec>>();
let (number_iter, string_iter) = if self.is_ascending {
let number_iter = ascending_facet_sort(
ctx.txn,
number_db,
field_id,
parent_candidates.clone(),
)?;
let string_iter = ascending_facet_sort(
ctx.txn,
string_db,
field_id,
parent_candidates.clone(),
)?;
(itertools::Either::Left(number_iter), itertools::Either::Left(string_iter))
} else {
let number_iter = descending_facet_sort(
ctx.txn,
number_db,
field_id,
parent_candidates.clone(),
)?;
let string_iter = descending_facet_sort(
ctx.txn,
string_db,
field_id,
parent_candidates.clone(),
)?;
(itertools::Either::Right(number_iter), itertools::Either::Right(string_iter))
};
2023-06-15 17:36:40 +02:00
let number_iter = number_iter.map(|r| -> Result<_> {
let (docids, bytes) = r?;
Ok((
docids,
serde_json::Value::Number(
serde_json::Number::from_f64(
OrderedF64Codec::bytes_decode(bytes).expect("some number"),
)
.expect("too big float"),
),
))
});
let string_iter = string_iter.map(|r| -> Result<_> {
let (docids, bytes) = r?;
Ok((
docids,
serde_json::Value::String(
StrRefCodec::bytes_decode(bytes).expect("some string").to_owned(),
),
))
});
2023-03-20 09:30:10 +01:00
let query_graph = parent_query.clone();
2023-06-15 17:36:40 +02:00
let ascending = self.is_ascending;
let field_name = self.field_name.clone();
let must_redact = self.must_redact;
RankingRuleOutputIterWrapper::new(Box::new(number_iter.chain(string_iter).map(
move |r| {
2023-06-15 17:36:40 +02:00
let (docids, value) = r?;
Ok(RankingRuleOutput {
query: query_graph.clone(),
candidates: docids,
score: ScoreDetails::Sort(score_details::Sort {
field_name: field_name.clone(),
ascending,
redacted: must_redact,
value,
}),
})
},
)))
}
None => RankingRuleOutputIterWrapper::new(Box::new(std::iter::empty())),
};
2023-03-20 09:30:10 +01:00
self.original_query = Some(parent_query.clone());
self.iter = Some(iter);
Ok(())
}
fn next_bucket(
&mut self,
2023-03-13 14:03:48 +01:00
_ctx: &mut SearchContext<'ctx>,
2023-02-22 15:34:37 +01:00
_logger: &mut dyn SearchLogger<Query>,
universe: &RoaringBitmap,
) -> Result<Option<RankingRuleOutput<Query>>> {
let iter = self.iter.as_mut().unwrap();
if let Some(mut bucket) = iter.next_bucket()? {
bucket.candidates &= universe;
Ok(Some(bucket))
} else {
let query = self.original_query.as_ref().unwrap().clone();
2023-06-15 17:36:40 +02:00
Ok(Some(RankingRuleOutput {
query,
candidates: universe.clone(),
score: ScoreDetails::Sort(score_details::Sort {
field_name: self.field_name.clone(),
ascending: self.is_ascending,
redacted: self.must_redact,
value: serde_json::Value::Null,
}),
}))
}
}
fn end_iteration(
&mut self,
2023-03-13 14:03:48 +01:00
_ctx: &mut SearchContext<'ctx>,
2023-02-22 15:34:37 +01:00
_logger: &mut dyn SearchLogger<Query>,
) {
self.original_query = None;
self.iter = None;
}
}