Introduce the Postings type to decorrelate the DocumentIds

meilisearch-core/src/bucket_sort.rs

@@ -635,12 +635,12 @@ fn fetch_matches<'txn, 'tag>(
                 let is_exact = *is_exact && distance == 0 && input.len() == query.len();
 
                 let before_postings_lists_fetching = Instant::now();
-                if let Some(postings_list) = postings_lists_store.postings_list(reader, input)? {
-                    postings_lists_original_length += postings_list.len();
+                if let Some(Postings { docids, matches }) = postings_lists_store.postings_list(reader, input)? {
+                    postings_lists_original_length += matches.len();
 
                     let input = Rc::from(input);
-                    let postings_list = Rc::new(postings_list);
-                    let postings_list_view = PostingsListView::original(input, postings_list);
+                    let matches = Rc::new(matches);
+                    let postings_list_view = PostingsListView::original(input, matches);
 
                     let mut offset = 0;
                     for group in postings_list_view.linear_group_by_key(|di| di.document_id) {

meilisearch-core/src/store/mod.rs

@@ -22,10 +22,15 @@ pub use self::synonyms::Synonyms;
 pub use self::updates::Updates;
 pub use self::updates_results::UpdatesResults;
 
+use std::borrow::Cow;
 use std::collections::HashSet;
+use std::convert::TryInto;
+use std::{mem, ptr};
 
 use heed::Result as ZResult;
+use heed::{BytesEncode, BytesDecode};
 use meilisearch_schema::{Schema, SchemaAttr};
+use sdset::{Set, SetBuf};
 use serde::de::{self, Deserialize};
 use zerocopy::{AsBytes, FromBytes};
 
@@ -33,7 +38,7 @@ use crate::criterion::Criteria;
 use crate::database::{UpdateEvent, UpdateEventsEmitter};
 use crate::database::{MainT, UpdateT};
 use crate::serde::Deserializer;
-use crate::{query_builder::QueryBuilder, update, DocumentId, Error, MResult};
+use crate::{query_builder::QueryBuilder, update, DocIndex, DocumentId, Error, MResult};
 
 type BEU64 = zerocopy::U64<byteorder::BigEndian>;
 type BEU16 = zerocopy::U16<byteorder::BigEndian>;
@@ -54,6 +59,88 @@ impl DocumentAttrKey {
     }
 }
 
+#[derive(Debug)]
+pub struct Postings<'a> {
+    pub docids: Cow<'a, Set<DocumentId>>,
+    pub matches: Cow<'a, Set<DocIndex>>,
+}
+
+struct PostingsCodec;
+
+impl<'a> BytesEncode<'a> for PostingsCodec {
+    type EItem = Postings<'a>;
+
+    fn bytes_encode(item: &'a Self::EItem) -> Option<Cow<'a, [u8]>> {
+        let u64_size = mem::size_of::<u64>();
+        let docids_size = item.docids.len() * mem::size_of::<DocumentId>();
+        let matches_size = item.matches.len() * mem::size_of::<DocIndex>();
+
+        let mut buffer = Vec::with_capacity(u64_size + docids_size + matches_size);
+
+        let docids_len = item.docids.len();
+        buffer.extend_from_slice(&docids_len.to_be_bytes());
+        buffer.extend_from_slice(item.docids.as_bytes());
+        buffer.extend_from_slice(item.matches.as_bytes());
+
+        Some(Cow::Owned(buffer))
+    }
+}
+
+fn aligned_to(bytes: &[u8], align: usize) -> bool {
+    (bytes as *const _ as *const () as usize) % align == 0
+}
+
+fn from_bytes_to_set<'a, T: 'a>(bytes: &'a [u8]) -> Option<Cow<'a, Set<T>>>
+where T: Clone + FromBytes
+{
+    match zerocopy::LayoutVerified::<_, [T]>::new_slice(bytes) {
+        Some(layout) => Some(Cow::Borrowed(Set::new_unchecked(layout.into_slice()))),
+        None => {
+            let len = bytes.len();
+            let elem_size = mem::size_of::<T>();
+
+            // ensure that it is the alignment that is wrong
+            // and the length is valid
+            if len % elem_size == 0 && !aligned_to(bytes, mem::align_of::<T>()) {
+                let elems = len / elem_size;
+                let mut vec = Vec::<T>::with_capacity(elems);
+
+                unsafe {
+                    let dst = vec.as_mut_ptr() as *mut u8;
+                    ptr::copy_nonoverlapping(bytes.as_ptr(), dst, len);
+                    vec.set_len(elems);
+                }
+
+                return Some(Cow::Owned(SetBuf::new_unchecked(vec)));
+            }
+
+            None
+        }
+    }
+}
+
+impl<'a> BytesDecode<'a> for PostingsCodec {
+    type DItem = Postings<'a>;
+
+    fn bytes_decode(bytes: &'a [u8]) -> Option<Self::DItem> {
+        let u64_size = mem::size_of::<u64>();
+        let docid_size = mem::size_of::<DocumentId>();
+        let docindex_size = mem::size_of::<DocIndex>();
+
+        let (len_bytes, bytes) = bytes.split_at(u64_size);
+        let docids_len = len_bytes.try_into().ok().map(u64::from_be_bytes)? as usize;
+        let docids_size = docids_len * docid_size;
+
+        let docids_bytes = &bytes[..docids_size];
+        let matches_bytes = &bytes[docids_size..];
+
+        let docids = from_bytes_to_set(docids_bytes)?;
+        let matches = from_bytes_to_set(matches_bytes)?;
+
+        Some(Postings { docids, matches })
+    }
+}
+
 fn main_name(name: &str) -> String {
     format!("store-{}", name)
 }

meilisearch-core/src/store/postings_lists.rs

@@ -1,13 +1,19 @@
-use crate::DocIndex;
-use crate::database::MainT;
-use heed::types::{ByteSlice, CowSlice};
-use heed::Result as ZResult;
-use sdset::{Set, SetBuf};
 use std::borrow::Cow;
+use std::convert::TryInto;
+use std::{mem, ptr};
+
+use heed::Result as ZResult;
+use heed::types::{ByteSlice, CowSlice};
+use sdset::{Set, SetBuf};
+use slice_group_by::GroupBy;
+
+use crate::database::MainT;
+use crate::{DocIndex, DocumentId};
+use crate::store::{Postings, PostingsCodec};
 
 #[derive(Copy, Clone)]
 pub struct PostingsLists {
-    pub(crate) postings_lists: heed::Database<ByteSlice, CowSlice<DocIndex>>,
+    pub(crate) postings_lists: heed::Database<ByteSlice, PostingsCodec>,
 }
 
 impl PostingsLists {
@@ -15,9 +21,14 @@ impl PostingsLists {
         self,
         writer: &mut heed::RwTxn<MainT>,
         word: &[u8],
-        words_indexes: &Set<DocIndex>,
+        matches: &Set<DocIndex>,
     ) -> ZResult<()> {
-        self.postings_lists.put(writer, word, words_indexes)
+        let docids = matches.linear_group_by_key(|m| m.document_id).map(|g| g[0].document_id).collect();
+        let docids = Cow::Owned(SetBuf::new_unchecked(docids));
+        let matches = Cow::Borrowed(matches);
+        let postings = Postings { docids, matches };
+
+        self.postings_lists.put(writer, word, &postings)
     }
 
     pub fn del_postings_list(self, writer: &mut heed::RwTxn<MainT>, word: &[u8]) -> ZResult<bool> {
@@ -32,11 +43,7 @@ impl PostingsLists {
         self,
         reader: &'txn heed::RoTxn<MainT>,
         word: &[u8],
-    ) -> ZResult<Option<Cow<'txn, Set<DocIndex>>>> {
-        match self.postings_lists.get(reader, word)? {
-            Some(Cow::Borrowed(slice)) => Ok(Some(Cow::Borrowed(Set::new_unchecked(slice)))),
-            Some(Cow::Owned(vec)) => Ok(Some(Cow::Owned(SetBuf::new_unchecked(vec)))),
-            None => Ok(None),
-        }
+    ) -> ZResult<Option<Postings<'txn>>> {
+        self.postings_lists.get(reader, word)
     }
 }