persistor

CHANGELOG.md

@@ -1,9 +1,3 @@
-Tantivy 0.13.0
-======================
-- Bugfix in `FuzzyTermQuery` not matching terms by prefix when it should (@Peachball)
-- Relaxed constraints on the custom/tweak score functions. At the segment level, they can be mut, and they are not required to be Sync + Send.
-- `MMapDirectory::open` does not return a `Result` anymore.
-
 Tantivy 0.12.0
 ======================
 - Removing static dispatch in tokenizers for simplicity. (#762)

Cargo.toml

@@ -18,7 +18,7 @@ byteorder = "1.0"
 crc32fast = "1.2.0"
 once_cell = "1.0"
 regex ={version = "1.3.0", default-features = false, features = ["std"]}
-tantivy-fst = "0.3"
+tantivy-fst = "0.2.1"
 memmap = {version = "0.7", optional=true}
 lz4 = {version="1.20", optional=true}
 snap = "1"
@@ -29,7 +29,7 @@ serde = {version="1.0", features=["derive"]}
 serde_json = "1.0"
 num_cpus = "1.2"
 fs2={version="0.4", optional=true}
-levenshtein_automata = "0.2"
+levenshtein_automata = "0.1"
 notify = {version="4", optional=true}
 uuid = { version = "0.8", features = ["v4", "serde"] }
 crossbeam = "0.7"
@@ -38,19 +38,18 @@ owning_ref = "0.4"
 stable_deref_trait = "1.0.0"
 rust-stemmers = "1.2"
 downcast-rs = { version="1.0" }
-tantivy-query-grammar = { version="0.13", path="./query-grammar" }
+tantivy-query-grammar = { version="0.12", path="./query-grammar" }
 bitpacking = {version="0.8", default-features = false, features=["bitpacker4x"]}
-census = "0.4"
+census =  {path="../census"}
 fnv = "1.0.6"
 owned-read = "0.4"
 failure = "0.1"
 htmlescape = "0.3.1"
-fail = "0.4"
+fail = "0.3"
 murmurhash32 = "0.2"
 chrono = "0.4"
 smallvec = "1.0"
 rayon = "1"
-# ordered-float = "1"
 
 [target.'cfg(windows)'.dependencies]
 winapi = "0.3"
@@ -59,11 +58,9 @@ winapi = "0.3"
 rand = "0.7"
 maplit = "1"
 matches = "0.1.8"
-proptest = "0.9"
-float-cmp = "0.6"
 
 [dev-dependencies.fail]
-version = "0.4"
+version = "0.3"
 features = ["failpoints"]
 
 [profile.release]

README.md

@@ -31,16 +31,12 @@ Tantivy is, in fact, strongly inspired by Lucene's design.
 
 # Benchmark
 
+Tantivy is typically faster than Lucene, but the results depend on 
+the nature of the queries in your workload.
+
 The following [benchmark](https://tantivy-search.github.io/bench/) break downs 
 performance for different type of queries / collection.
 
-
-In general, Tantivy tends to be 
-- slower than Lucene on union with a Top-K due to Block-WAND optimization.
-- faster than Lucene on intersection and phrase queries. 
-
-Your mileage WILL vary depending on the nature of queries and their load.
-
 # Features
 
 - Full-text search

TODO.md

@@ -0,0 +1,9 @@
+- segment writer serializes if already flush
+- flush if exceed max doc. 
+- unit tests
+- reader from `IndexWriter`
+- configurable limit to flush 
+- bundle directory
+- add index worker when exceeds some doc limit
+- flush before prepare commit.
+- segment_writer should not receive a segment 

appveyor.yml

@@ -18,5 +18,5 @@ install:
 build: false
 
 test_script:
-  - REM SET RUST_LOG=tantivy,test & cargo test --all --verbose --no-default-features --features mmap
+  - REM SET RUST_LOG=tantivy,test & cargo test --verbose --no-default-features --features mmap
   - REM SET RUST_BACKTRACE=1 & cargo build --examples

examples/faceted_search_with_tweaked_score.rs

@@ -1,98 +0,0 @@
-use std::collections::HashSet;
-use tantivy::collector::TopDocs;
-use tantivy::doc;
-use tantivy::query::BooleanQuery;
-use tantivy::schema::*;
-use tantivy::{DocId, Index, Score, SegmentReader};
-
-fn main() -> tantivy::Result<()> {
-    let mut schema_builder = Schema::builder();
-
-    let title = schema_builder.add_text_field("title", STORED);
-    let ingredient = schema_builder.add_facet_field("ingredient");
-
-    let schema = schema_builder.build();
-    let index = Index::create_in_ram(schema.clone());
-
-    let mut index_writer = index.writer(30_000_000)?;
-
-    index_writer.add_document(doc!(
-        title => "Fried egg",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/oil"),
-    ));
-    index_writer.add_document(doc!(
-        title => "Scrambled egg",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/butter"),
-        ingredient => Facet::from("/ingredient/milk"),
-        ingredient => Facet::from("/ingredient/salt"),
-    ));
-    index_writer.add_document(doc!(
-        title => "Egg rolls",
-        ingredient => Facet::from("/ingredient/egg"),
-        ingredient => Facet::from("/ingredient/garlic"),
-        ingredient => Facet::from("/ingredient/salt"),
-        ingredient => Facet::from("/ingredient/oil"),
-        ingredient => Facet::from("/ingredient/tortilla-wrap"),
-        ingredient => Facet::from("/ingredient/mushroom"),
-    ));
-    index_writer.commit()?;
-
-    let reader = index.reader()?;
-    let searcher = reader.searcher();
-    {
-        let facets = vec![
-            Facet::from("/ingredient/egg"),
-            Facet::from("/ingredient/oil"),
-            Facet::from("/ingredient/garlic"),
-            Facet::from("/ingredient/mushroom"),
-        ];
-        let query = BooleanQuery::new_multiterms_query(
-            facets
-                .iter()
-                .map(|key| Term::from_facet(ingredient, &key))
-                .collect(),
-        );
-        let top_docs_by_custom_score =
-            TopDocs::with_limit(2).tweak_score(move |segment_reader: &SegmentReader| {
-                let mut ingredient_reader = segment_reader.facet_reader(ingredient).unwrap();
-                let facet_dict = ingredient_reader.facet_dict();
-
-                let query_ords: HashSet<u64> = facets
-                    .iter()
-                    .filter_map(|key| facet_dict.term_ord(key.encoded_str()))
-                    .collect();
-
-                let mut facet_ords_buffer: Vec<u64> = Vec::with_capacity(20);
-
-                move |doc: DocId, original_score: Score| {
-                    ingredient_reader.facet_ords(doc, &mut facet_ords_buffer);
-                    let missing_ingredients = facet_ords_buffer
-                        .iter()
-                        .filter(|ord| !query_ords.contains(ord))
-                        .count();
-                    let tweak = 1.0 / 4_f32.powi(missing_ingredients as i32);
-
-                    original_score * tweak
-                }
-            });
-        let top_docs = searcher.search(&query, &top_docs_by_custom_score)?;
-
-        let titles: Vec<String> = top_docs
-            .iter()
-            .map(|(_, doc_id)| {
-                searcher
-                    .doc(*doc_id)
-                    .unwrap()
-                    .get_first(title)
-                    .unwrap()
-                    .text()
-                    .unwrap()
-                    .to_owned()
-            })
-            .collect();
-        assert_eq!(titles, vec!["Fried egg", "Egg rolls"]);
-    }
-    Ok(())
-}

query-grammar/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "tantivy-query-grammar"
-version = "0.13.0"
+version = "0.12.0"
 authors = ["Paul Masurel <paul.masurel@gmail.com>"]
 license = "MIT"
 categories = ["database-implementations", "data-structures"]

query-grammar/src/query_grammar.rs

@@ -154,11 +154,17 @@ fn negate(expr: UserInputAST) -> UserInputAST {
     expr.unary(Occur::MustNot)
 }
 
+fn must(expr: UserInputAST) -> UserInputAST {
+    expr.unary(Occur::Must)
+}
+
 fn leaf<'a>() -> impl Parser<&'a str, Output = UserInputAST> {
     parser(|input| {
-        char('(')
-            .with(ast())
-            .skip(char(')'))
+        char('-')
+            .with(leaf())
+            .map(negate)
+            .or(char('+').with(leaf()).map(must))
+            .or(char('(').with(ast()).skip(char(')')))
             .or(char('*').map(|_| UserInputAST::from(UserInputLeaf::All)))
             .or(attempt(
                 string("NOT").skip(spaces1()).with(leaf()).map(negate),
@@ -170,16 +176,6 @@ fn leaf<'a>() -> impl Parser<&'a str, Output = UserInputAST> {
     })
 }
 
-fn occur_symbol<'a>() -> impl Parser<&'a str, Output = Occur> {
-    char('-')
-        .map(|_| Occur::MustNot)
-        .or(char('+').map(|_| Occur::Must))
-}
-
-fn occur_leaf<'a>() -> impl Parser<&'a str, Output = (Option<Occur>, UserInputAST)> {
-    (optional(occur_symbol()), boosted_leaf())
-}
-
 fn positive_float_number<'a>() -> impl Parser<&'a str, Output = f32> {
     (many1(digit()), optional((char('.'), many1(digit())))).map(
         |(int_part, decimal_part_opt): (String, Option<(char, String)>)| {
@@ -243,29 +239,21 @@ fn aggregate_binary_expressions(
     }
 }
 
-fn operand_leaf<'a>() -> impl Parser<&'a str, Output = (BinaryOperand, UserInputAST)> {
-    (
+pub fn ast<'a>() -> impl Parser<&'a str, Output = UserInputAST> {
+    let operand_leaf = (
         binary_operand().skip(spaces()),
         boosted_leaf().skip(spaces()),
-    )
-}
-
-pub fn ast<'a>() -> impl Parser<&'a str, Output = UserInputAST> {
-    let boolean_expr = (boosted_leaf().skip(spaces()), many1(operand_leaf()))
+    );
+    let boolean_expr = (boosted_leaf().skip(spaces().silent()), many1(operand_leaf))
         .map(|(left, right)| aggregate_binary_expressions(left, right));
-    let whitespace_separated_leaves = many1(occur_leaf().skip(spaces().silent())).map(
-        |subqueries: Vec<(Option<Occur>, UserInputAST)>| {
+    let whitespace_separated_leaves =
+        many1(boosted_leaf().skip(spaces().silent())).map(|subqueries: Vec<UserInputAST>| {
             if subqueries.len() == 1 {
-                let (occur_opt, ast) = subqueries.into_iter().next().unwrap();
-                match occur_opt.unwrap_or(Occur::Should) {
-                    Occur::Must | Occur::Should => ast,
-                    Occur::MustNot => UserInputAST::Clause(vec![(Some(Occur::MustNot), ast)]),
-                }
+                subqueries.into_iter().next().unwrap()
             } else {
                 UserInputAST::Clause(subqueries.into_iter().collect())
             }
-        },
-    );
+        });
     let expr = attempt(boolean_expr).or(whitespace_separated_leaves);
     spaces().with(expr).skip(spaces())
 }
@@ -295,12 +283,6 @@ mod test {
         );
     }
 
-    #[test]
-    fn test_occur_symbol() {
-        assert_eq!(super::occur_symbol().parse("-"), Ok((Occur::MustNot, "")));
-        assert_eq!(super::occur_symbol().parse("+"), Ok((Occur::Must, "")));
-    }
-
     #[test]
     fn test_positive_float_number() {
         fn valid_parse(float_str: &str, expected_val: f32, expected_remaining: &str) {
@@ -348,7 +330,7 @@ mod test {
             "Err(UnexpectedParse)"
         );
         test_parse_query_to_ast_helper("NOTa", "\"NOTa\"");
-        test_parse_query_to_ast_helper("NOT a", "(-\"a\")");
+        test_parse_query_to_ast_helper("NOT a", "-(\"a\")");
     }
 
     #[test]
@@ -356,16 +338,16 @@ mod test {
         assert!(parse_to_ast().parse("a^2^3").is_err());
         assert!(parse_to_ast().parse("a^2^").is_err());
         test_parse_query_to_ast_helper("a^3", "(\"a\")^3");
-        test_parse_query_to_ast_helper("a^3 b^2", "(*(\"a\")^3 *(\"b\")^2)");
+        test_parse_query_to_ast_helper("a^3 b^2", "((\"a\")^3 (\"b\")^2)");
         test_parse_query_to_ast_helper("a^1", "\"a\"");
     }
 
     #[test]
     fn test_parse_query_to_ast_binary_op() {
-        test_parse_query_to_ast_helper("a AND b", "(+\"a\" +\"b\")");
-        test_parse_query_to_ast_helper("a OR b", "(?\"a\" ?\"b\")");
-        test_parse_query_to_ast_helper("a OR b AND c", "(?\"a\" ?(+\"b\" +\"c\"))");
-        test_parse_query_to_ast_helper("a AND b         AND c", "(+\"a\" +\"b\" +\"c\")");
+        test_parse_query_to_ast_helper("a AND b", "(+(\"a\") +(\"b\"))");
+        test_parse_query_to_ast_helper("a OR b", "(?(\"a\") ?(\"b\"))");
+        test_parse_query_to_ast_helper("a OR b AND c", "(?(\"a\") ?((+(\"b\") +(\"c\"))))");
+        test_parse_query_to_ast_helper("a AND b         AND c", "(+(\"a\") +(\"b\") +(\"c\"))");
         assert_eq!(
             format!("{:?}", parse_to_ast().parse("a OR b aaa")),
             "Err(UnexpectedParse)"
@@ -403,13 +385,6 @@ mod test {
         test_parse_query_to_ast_helper("weight: <= 70.5", "weight:{\"*\" TO \"70.5\"]");
     }
 
-    #[test]
-    fn test_occur_leaf() {
-        let ((occur, ast), _) = super::occur_leaf().parse("+abc").unwrap();
-        assert_eq!(occur, Some(Occur::Must));
-        assert_eq!(format!("{:?}", ast), "\"abc\"");
-    }
-
     #[test]
     fn test_range_parser() {
         // testing the range() parser separately
@@ -438,67 +413,32 @@ mod test {
     fn test_parse_query_to_triming_spaces() {
         test_parse_query_to_ast_helper("   abc", "\"abc\"");
         test_parse_query_to_ast_helper("abc ", "\"abc\"");
-        test_parse_query_to_ast_helper("(  a OR abc)", "(?\"a\" ?\"abc\")");
-        test_parse_query_to_ast_helper("(a  OR abc)", "(?\"a\" ?\"abc\")");
-        test_parse_query_to_ast_helper("(a OR  abc)", "(?\"a\" ?\"abc\")");
-        test_parse_query_to_ast_helper("a OR abc ", "(?\"a\" ?\"abc\")");
-        test_parse_query_to_ast_helper("(a OR abc )", "(?\"a\" ?\"abc\")");
-        test_parse_query_to_ast_helper("(a OR  abc) ", "(?\"a\" ?\"abc\")");
+        test_parse_query_to_ast_helper("(  a OR abc)", "(?(\"a\") ?(\"abc\"))");
+        test_parse_query_to_ast_helper("(a  OR abc)", "(?(\"a\") ?(\"abc\"))");
+        test_parse_query_to_ast_helper("(a OR  abc)", "(?(\"a\") ?(\"abc\"))");
+        test_parse_query_to_ast_helper("a OR abc ", "(?(\"a\") ?(\"abc\"))");
+        test_parse_query_to_ast_helper("(a OR abc )", "(?(\"a\") ?(\"abc\"))");
+        test_parse_query_to_ast_helper("(a OR  abc) ", "(?(\"a\") ?(\"abc\"))");
     }
 
     #[test]
-    fn test_parse_query_single_term() {
+    fn test_parse_query_to_ast() {
         test_parse_query_to_ast_helper("abc", "\"abc\"");
-    }
-
-    #[test]
-    fn test_parse_query_default_clause() {
-        test_parse_query_to_ast_helper("a b", "(*\"a\" *\"b\")");
-    }
-
-    #[test]
-    fn test_parse_query_must_default_clause() {
-        test_parse_query_to_ast_helper("+(a b)", "(*\"a\" *\"b\")");
-    }
-
-    #[test]
-    fn test_parse_query_must_single_term() {
-        test_parse_query_to_ast_helper("+d", "\"d\"");
-    }
-
-    #[test]
-    fn test_single_term_with_field() {
+        test_parse_query_to_ast_helper("a b", "(\"a\" \"b\")");
+        test_parse_query_to_ast_helper("+(a b)", "+((\"a\" \"b\"))");
+        test_parse_query_to_ast_helper("+d", "+(\"d\")");
+        test_parse_query_to_ast_helper("+(a b) +d", "(+((\"a\" \"b\")) +(\"d\"))");
+        test_parse_query_to_ast_helper("(+a +b) d", "((+(\"a\") +(\"b\")) \"d\")");
+        test_parse_query_to_ast_helper("(+a)", "+(\"a\")");
+        test_parse_query_to_ast_helper("(+a +b)", "(+(\"a\") +(\"b\"))");
         test_parse_query_to_ast_helper("abc:toto", "abc:\"toto\"");
-    }
-
-    #[test]
-    fn test_single_term_with_float() {
         test_parse_query_to_ast_helper("abc:1.1", "abc:\"1.1\"");
-    }
-
-    #[test]
-    fn test_must_clause() {
-        test_parse_query_to_ast_helper("(+a +b)", "(+\"a\" +\"b\")");
-    }
-
-    #[test]
-    fn test_parse_test_query_plus_a_b_plus_d() {
-        test_parse_query_to_ast_helper("+(a b) +d", "(+(*\"a\" *\"b\") +\"d\")");
-    }
-
-    #[test]
-    fn test_parse_test_query_other() {
-        test_parse_query_to_ast_helper("(+a +b) d", "(*(+\"a\" +\"b\") *\"d\")");
-        test_parse_query_to_ast_helper("+abc:toto", "abc:\"toto\"");
-        test_parse_query_to_ast_helper("(+abc:toto -titi)", "(+abc:\"toto\" -\"titi\")");
-        test_parse_query_to_ast_helper("-abc:toto", "(-abc:\"toto\")");
-        test_parse_query_to_ast_helper("abc:a b", "(*abc:\"a\" *\"b\")");
+        test_parse_query_to_ast_helper("+abc:toto", "+(abc:\"toto\")");
+        test_parse_query_to_ast_helper("(+abc:toto -titi)", "(+(abc:\"toto\") -(\"titi\"))");
+        test_parse_query_to_ast_helper("-abc:toto", "-(abc:\"toto\")");
+        test_parse_query_to_ast_helper("abc:a b", "(abc:\"a\" \"b\")");
         test_parse_query_to_ast_helper("abc:\"a b\"", "abc:\"a b\"");
         test_parse_query_to_ast_helper("foo:[1 TO 5]", "foo:[\"1\" TO \"5\"]");
-    }
-
-    #[test]
-    fn test_parse_query_with_range() {
         test_parse_query_to_ast_helper("[1 TO 5]", "[\"1\" TO \"5\"]");
         test_parse_query_to_ast_helper("foo:{a TO z}", "foo:{\"a\" TO \"z\"}");
         test_parse_query_to_ast_helper("foo:[1 TO toto}", "foo:[\"1\" TO \"toto\"}");

query-grammar/src/user_input_ast.rs

@@ -85,14 +85,15 @@ impl UserInputBound {
 }
 
 pub enum UserInputAST {
-    Clause(Vec<(Option<Occur>, UserInputAST)>),
+    Clause(Vec<UserInputAST>),
+    Unary(Occur, Box<UserInputAST>),
     Leaf(Box<UserInputLeaf>),
     Boost(Box<UserInputAST>, f32),
 }
 
 impl UserInputAST {
     pub fn unary(self, occur: Occur) -> UserInputAST {
-        UserInputAST::Clause(vec![(Some(occur), self)])
+        UserInputAST::Unary(occur, Box::new(self))
     }
 
     fn compose(occur: Occur, asts: Vec<UserInputAST>) -> UserInputAST {
@@ -103,7 +104,7 @@ impl UserInputAST {
         } else {
             UserInputAST::Clause(
                 asts.into_iter()
-                    .map(|ast: UserInputAST| (Some(occur), ast))
+                    .map(|ast: UserInputAST| ast.unary(occur))
                     .collect::<Vec<_>>(),
             )
         }
@@ -134,36 +135,25 @@ impl From<UserInputLeaf> for UserInputAST {
     }
 }
 
-fn print_occur_ast(
-    occur_opt: Option<Occur>,
-    ast: &UserInputAST,
-    formatter: &mut fmt::Formatter,
-) -> fmt::Result {
-    if let Some(occur) = occur_opt {
-        write!(formatter, "{}{:?}", occur, ast)?;
-    } else {
-        write!(formatter, "*{:?}", ast)?;
-    }
-    Ok(())
-}
-
 impl fmt::Debug for UserInputAST {
-    fn fmt(&self, formatter: &mut fmt::Formatter) -> fmt::Result {
+    fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
         match *self {
             UserInputAST::Clause(ref subqueries) => {
                 if subqueries.is_empty() {
                     write!(formatter, "<emptyclause>")?;
                 } else {
                     write!(formatter, "(")?;
-                    print_occur_ast(subqueries[0].0, &subqueries[0].1, formatter)?;
+                    write!(formatter, "{:?}", &subqueries[0])?;
                     for subquery in &subqueries[1..] {
-                        write!(formatter, " ")?;
-                        print_occur_ast(subquery.0, &subquery.1, formatter)?;
+                        write!(formatter, " {:?}", subquery)?;
                     }
                     write!(formatter, ")")?;
                 }
                 Ok(())
             }
+            UserInputAST::Unary(ref occur, ref subquery) => {
+                write!(formatter, "{}({:?})", occur, subquery)
+            }
             UserInputAST::Leaf(ref subquery) => write!(formatter, "{:?}", subquery),
             UserInputAST::Boost(ref leaf, boost) => write!(formatter, "({:?})^{}", leaf, boost),
         }

src/collector/custom_score_top_collector.rs

@@ -28,7 +28,7 @@ where
 /// It is the segment local version of the [`CustomScorer`](./trait.CustomScorer.html).
 pub trait CustomSegmentScorer<TScore>: 'static {
     /// Computes the score of a specific `doc`.
-    fn score(&mut self, doc: DocId) -> TScore;
+    fn score(&self, doc: DocId) -> TScore;
 }
 
 /// `CustomScorer` makes it possible to define any kind of score.
@@ -117,9 +117,9 @@ where
 
 impl<F, TScore> CustomSegmentScorer<TScore> for F
 where
-    F: 'static + FnMut(DocId) -> TScore,
+    F: 'static + Sync + Send + Fn(DocId) -> TScore,
 {
-    fn score(&mut self, doc: DocId) -> TScore {
+    fn score(&self, doc: DocId) -> TScore {
         (self)(doc)
     }
 }

src/collector/mod.rs

@@ -84,7 +84,7 @@ See the `custom_collector` example.
 
 */
 
-use crate::{DocId, Searcher, Executor};
+use crate::DocId;
 use crate::Score;
 use crate::SegmentLocalId;
 use crate::SegmentReader;
@@ -100,9 +100,6 @@ mod top_collector;
 
 mod top_score_collector;
 pub use self::top_score_collector::TopDocs;
-#[cfg(test)]
-pub(crate) use self::top_score_collector::TopScoreSegmentCollector;
-
 
 mod custom_score_top_collector;
 pub use self::custom_score_top_collector::{CustomScorer, CustomSegmentScorer};
@@ -112,9 +109,6 @@ pub use self::tweak_score_top_collector::{ScoreSegmentTweaker, ScoreTweaker};
 
 mod facet_collector;
 pub use self::facet_collector::FacetCollector;
-use crate::fastfield::DeleteBitSet;
-use crate::query::{Scorer, Weight};
-use std::borrow::BorrowMut;
 
 /// `Fruit` is the type for the result of our collection.
 /// e.g. `usize` for the `Count` collector.
@@ -122,8 +116,6 @@ pub trait Fruit: Send + downcast_rs::Downcast {}
 
 impl<T> Fruit for T where T: Send + downcast_rs::Downcast {}
 
-
-
 /// Collectors are in charge of collecting and retaining relevant
 /// information from the document found and scored by the query.
 ///
@@ -162,20 +154,6 @@ pub trait Collector: Sync {
     /// Combines the fruit associated to the collection of each segments
     /// into one fruit.
     fn merge_fruits(&self, segment_fruits: Vec<Self::Fruit>) -> crate::Result<Self::Fruit>;
-
-    fn collect_weight(&self, searcher: &Searcher, weight: &dyn Weight, executor: &Executor) -> crate::Result<Self::Fruit> {
-        let segment_readers = searcher.segment_readers();
-        let fruits = executor.map(
-            |(segment_ord, segment_reader)| {
-                let mut scorer = weight.scorer(segment_reader, 1.0f32)?;
-                let segment_collector =
-                    self.for_segment(segment_ord as u32, segment_reader)?;
-                Ok(segment_collector.collect_scorer(scorer.borrow_mut(), segment_reader.delete_bitset()))
-            },
-            segment_readers.iter().enumerate(),
-        )?;
-        self.merge_fruits(fruits)
-    }
 }
 
 /// The `SegmentCollector` is the trait in charge of defining the
@@ -183,7 +161,7 @@ pub trait Collector: Sync {
 ///
 /// `.collect(doc, score)` will be called for every documents
 /// matching the query.
-pub trait SegmentCollector: 'static + Sized {
+pub trait SegmentCollector: 'static {
     /// `Fruit` is the type for the result of our collection.
     /// e.g. `usize` for the `Count` collector.
     type Fruit: Fruit;
@@ -193,19 +171,6 @@ pub trait SegmentCollector: 'static + Sized {
 
     /// Extract the fruit of the collection from the `SegmentCollector`.
     fn harvest(self) -> Self::Fruit;
-
-    fn collect_scorer(mut self, scorer: &mut dyn Scorer, delete_bitset: Option<&DeleteBitSet>) -> Self::Fruit {
-        if let Some(delete_bitset) = delete_bitset {
-            scorer.for_each(&mut |doc, score| {
-                if delete_bitset.is_alive(doc) {
-                    self.collect(doc, score);
-                }
-            });
-        } else {
-            scorer.for_each(&mut |doc, score| self.collect(doc, score));
-        }
-        self.harvest()
-    }
 }
 
 // -----------------------------------------------

src/collector/top_collector.rs

@@ -56,7 +56,7 @@ impl<T: PartialOrd, D: PartialOrd> PartialEq for ComparableDoc<T, D> {
 impl<T: PartialOrd, D: PartialOrd> Eq for ComparableDoc<T, D> {}
 
 pub(crate) struct TopCollector<T> {
-    pub limit: usize,
+    limit: usize,
     _marker: PhantomData<T>,
 }
 
@@ -69,7 +69,9 @@ where
     /// # Panics
     /// The method panics if limit is 0
     pub fn with_limit(limit: usize) -> TopCollector<T> {
-        assert!(limit > 0, "Limit must be strictly greater than 0.");
+        if limit < 1 {
+            panic!("Limit must be strictly greater than 0.");
+        }
         TopCollector {
             limit,
             _marker: PhantomData,
@@ -122,13 +124,13 @@ where
 /// The theorical complexity for collecting the top `K` out of `n` documents
 /// is `O(n log K)`.
 pub(crate) struct TopSegmentCollector<T> {
-    pub limit: usize,
+    limit: usize,
     heap: BinaryHeap<ComparableDoc<T, DocId>>,
     segment_id: u32,
 }
 
 impl<T: PartialOrd> TopSegmentCollector<T> {
-    pub fn new(segment_id: SegmentLocalId, limit: usize) -> TopSegmentCollector<T> {
+    fn new(segment_id: SegmentLocalId, limit: usize) -> TopSegmentCollector<T> {
         TopSegmentCollector {
             limit,
             heap: BinaryHeap::with_capacity(limit),
@@ -159,10 +161,6 @@ impl<T: PartialOrd + Clone> TopSegmentCollector<T> {
         self.heap.len() >= self.limit
     }
 
-    pub fn pruning_score(&self) -> Option<T> {
-        self.heap.peek().map(|head| head.feature.clone())
-    }
-
     /// Collects a document scored by the given feature
     ///
     /// It collects documents until it has reached the max capacity. Once it reaches capacity, it

src/collector/top_score_collector.rs

@@ -8,13 +8,12 @@ use crate::collector::{
 };
 use crate::fastfield::FastFieldReader;
 use crate::schema::Field;
-use crate::{DocAddress, Executor, Searcher};
+use crate::DocAddress;
 use crate::DocId;
 use crate::Score;
 use crate::SegmentLocalId;
 use crate::SegmentReader;
 use std::fmt;
-use crate::query::{Weight, PruningScorerIfPossible};
 
 /// The `TopDocs` collector keeps track of the top `K` documents
 /// sorted by their score.
@@ -67,7 +66,7 @@ struct ScorerByFastFieldReader {
 }
 
 impl CustomSegmentScorer<u64> for ScorerByFastFieldReader {
-    fn score(&mut self, doc: DocId) -> u64 {
+    fn score(&self, doc: DocId) -> u64 {
         self.ff_reader.get_u64(u64::from(doc))
     }
 }
@@ -418,42 +417,6 @@ impl Collector for TopDocs {
         true
     }
 
-    fn collect_weight(&self, searcher: &Searcher, weight: &dyn Weight, executor: &Executor) -> crate::Result<Self::Fruit> {
-        let segment_readers = searcher.segment_readers();
-        let fruits = executor.map(
-            |(segment_ord, segment_reader)| {
-                match weight.pruning_scorer(segment_reader, 1.0f32)? {
-                    PruningScorerIfPossible::NonPruning(mut scorer) => {
-                        let segment_collector =
-                            self.for_segment(segment_ord as u32, segment_reader)?;
-                        let fruit =
-                            segment_collector.collect_scorer(scorer.as_mut(), segment_reader.delete_bitset());
-                        Ok(fruit)
-                    }
-                    PruningScorerIfPossible::Pruning(mut pruning_scorer) => {
-                        let limit = self.0.limit;
-                        let mut segment_collector =
-                            self.for_segment(segment_ord as u32, segment_reader)?;
-                        for _ in 0..limit {
-                            if !pruning_scorer.advance() {
-                                return Ok(segment_collector.harvest());
-                            }
-                            segment_collector.collect(pruning_scorer.doc(), pruning_scorer.score());
-                        }
-                        let mut pruning_score = segment_collector.0.pruning_score().unwrap_or(0.0f32);
-                        while pruning_scorer.advance_with_pruning(pruning_score) {
-                            segment_collector.0.collect(pruning_scorer.doc(), pruning_scorer.score());
-                            pruning_score = segment_collector.0.pruning_score().unwrap_or(0.0f32);
-                        }
-                        Ok(segment_collector.harvest())
-                    }
-                }
-            },
-            segment_readers.iter().enumerate(),
-        )?;
-        self.merge_fruits(fruits)
-    }
-
     fn merge_fruits(
         &self,
         child_fruits: Vec<Vec<(Score, DocAddress)>>,
@@ -465,12 +428,6 @@ impl Collector for TopDocs {
 /// Segment Collector associated to `TopDocs`.
 pub struct TopScoreSegmentCollector(TopSegmentCollector<Score>);
 
-impl TopScoreSegmentCollector {
-    pub fn new(segment_id: SegmentLocalId, limit: usize) -> Self {
-        TopScoreSegmentCollector(TopSegmentCollector::new(segment_id, limit))
-    }
-}
-
 impl SegmentCollector for TopScoreSegmentCollector {
     type Fruit = Vec<(Score, DocAddress)>;
 

src/collector/tweak_score_top_collector.rs

@@ -29,7 +29,7 @@ where
 /// It is the segment local version of the [`ScoreTweaker`](./trait.ScoreTweaker.html).
 pub trait ScoreSegmentTweaker<TScore>: 'static {
     /// Tweak the given `score` for the document `doc`.
-    fn score(&mut self, doc: DocId, score: Score) -> TScore;
+    fn score(&self, doc: DocId, score: Score) -> TScore;
 }
 
 /// `ScoreTweaker` makes it possible to tweak the score
@@ -121,9 +121,9 @@ where
 
 impl<F, TScore> ScoreSegmentTweaker<TScore> for F
 where
-    F: 'static + FnMut(DocId, Score) -> TScore,
+    F: 'static + Sync + Send + Fn(DocId, Score) -> TScore,
 {
-    fn score(&mut self, doc: DocId, score: Score) -> TScore {
+    fn score(&self, doc: DocId, score: Score) -> TScore {
         (self)(doc, score)
     }
 }

src/common/composite_file.rs

@@ -190,11 +190,12 @@ mod test {
     use crate::schema::Field;
     use std::io::Write;
     use std::path::Path;
+    
 
     #[test]
     fn test_composite_file() {
         let path = Path::new("test_path");
-        let mut directory = RAMDirectory::create();
+        let mut directory = RAMDirectory::default();
         {
             let w = directory.open_write(path).unwrap();
             let mut composite_write = CompositeWrite::wrap(w);

src/common/mod.rs

@@ -2,6 +2,7 @@ pub mod bitpacker;
 mod bitset;
 mod composite_file;
 mod counting_writer;
+mod mutable_enum;
 mod serialize;
 mod vint;
 
@@ -9,6 +10,7 @@ pub use self::bitset::BitSet;
 pub(crate) use self::bitset::TinySet;
 pub(crate) use self::composite_file::{CompositeFile, CompositeWrite};
 pub use self::counting_writer::CountingWriter;
+pub(crate) use self::mutable_enum::MutableEnum;
 pub use self::serialize::{BinarySerializable, FixedSize};
 pub use self::vint::{read_u32_vint, serialize_vint_u32, write_u32_vint, VInt};
 pub use byteorder::LittleEndian as Endianness;

src/common/mutable_enum.rs

@@ -0,0 +1,37 @@
+use std::ops::{Deref, DerefMut};
+
+pub(crate) struct MutableEnum<T>(Option<T>);
+
+impl<T> MutableEnum<T> {
+    pub fn wrap(val: T) -> Self {
+        MutableEnum(Some(val))
+    }
+
+    pub fn into(self) -> T {
+        self.0.unwrap()
+    }
+}
+
+impl<T> MutableEnum<T> {
+    pub fn map_mutate<E, F>(&mut self, transformation: F) -> Result<(), E>
+    where
+        F: FnOnce(T) -> Result<T, E>,
+    {
+        self.0 = self.0.take().map(transformation).transpose()?;
+        Ok(())
+    }
+}
+
+impl<T> Deref for MutableEnum<T> {
+    type Target = T;
+
+    fn deref(&self) -> &Self::Target {
+        self.0.as_ref().unwrap()
+    }
+}
+
+impl<T> DerefMut for MutableEnum<T> {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        self.0.as_mut().unwrap()
+    }
+}

src/core/index.rs

@@ -12,8 +12,8 @@ use crate::directory::INDEX_WRITER_LOCK;
 use crate::directory::{Directory, RAMDirectory};
 use crate::error::DataCorruption;
 use crate::error::TantivyError;
-use crate::indexer::index_writer::HEAP_SIZE_MIN;
 use crate::indexer::segment_updater::save_new_metas;
+use crate::indexer::IndexWriterConfig;
 use crate::reader::IndexReader;
 use crate::reader::IndexReaderBuilder;
 use crate::schema::Field;
@@ -21,6 +21,7 @@ use crate::schema::FieldType;
 use crate::schema::Schema;
 use crate::tokenizer::{TextAnalyzer, TokenizerManager};
 use crate::IndexWriter;
+use num_cpus;
 use std::borrow::BorrowMut;
 use std::collections::HashSet;
 use std::fmt;
@@ -272,30 +273,14 @@ impl Index {
     pub fn writer_with_num_threads(
         &self,
         num_threads: usize,
-        overall_heap_size_in_bytes: usize,
+        overall_heap_size_in_bytes: u64,
     ) -> crate::Result<IndexWriter> {
-        let directory_lock = self
-            .directory
-            .acquire_lock(&INDEX_WRITER_LOCK)
-            .map_err(|err| {
-                TantivyError::LockFailure(
-                    err,
-                    Some(
-                        "Failed to acquire index lock. If you are using\
-                         a regular directory, this means there is already an \
-                         `IndexWriter` working on this `Directory`, in this process \
-                         or in a different process."
-                            .to_string(),
-                    ),
-                )
-            })?;
-        let heap_size_in_bytes_per_thread = overall_heap_size_in_bytes / num_threads;
-        IndexWriter::new(
-            self,
-            num_threads,
-            heap_size_in_bytes_per_thread,
-            directory_lock,
-        )
+        let config = IndexWriterConfig {
+            max_indexing_threads: num_threads,
+            memory_budget: overall_heap_size_in_bytes,
+            ..Default::default()
+        };
+        self.writer_from_config(config)
     }
 
     /// Creates a multithreaded writer
@@ -308,13 +293,36 @@ impl Index {
     /// If the lockfile already exists, returns `Error::FileAlreadyExists`.
     /// # Panics
     /// If the heap size per thread is too small, panics.
-    pub fn writer(&self, overall_heap_size_in_bytes: usize) -> crate::Result<IndexWriter> {
-        let mut num_threads = num_cpus::get();
-        let heap_size_in_bytes_per_thread = overall_heap_size_in_bytes / num_threads;
-        if heap_size_in_bytes_per_thread < HEAP_SIZE_MIN {
-            num_threads = (overall_heap_size_in_bytes / HEAP_SIZE_MIN).max(1);
-        }
-        self.writer_with_num_threads(num_threads, overall_heap_size_in_bytes)
+    pub fn writer(&self, overall_heap_size_in_bytes: u64) -> crate::Result<IndexWriter> {
+        let config = IndexWriterConfig {
+            max_indexing_threads: num_cpus::get(),
+            memory_budget: overall_heap_size_in_bytes,
+            ..Default::default()
+        };
+        self.writer_from_config(config)
+    }
+
+    /// Creates a new writer with a given configuration.
+    ///
+    /// See [`IndexWriterConfig`](./struct.IndexWriterConfig.html) for more information.
+    pub fn writer_from_config(&self, mut config: IndexWriterConfig) -> crate::Result<IndexWriter> {
+        config.validate()?;
+        let directory_lock = self
+            .directory
+            .acquire_lock(&INDEX_WRITER_LOCK)
+            .map_err(|err| {
+                TantivyError::LockFailure(
+                    err,
+                    Some(
+                        "Failed to acquire index lock. If you are using \
+                         a regular directory, this means there is already an \
+                         `IndexWriter` working on this `Directory`, in this process \
+                         or in a different process."
+                            .to_string(),
+                    ),
+                )
+            })?;
+        IndexWriter::new(self, config, directory_lock)
     }
 
     /// Accessor to the index schema
@@ -468,7 +476,7 @@ mod tests {
             .try_into()
             .unwrap();
         assert_eq!(reader.searcher().num_docs(), 0);
-        test_index_on_commit_reload_policy_aux(field, &index, &reader);
+        test_index_on_commit_reload_policy_aux(field, index.clone(), &index, &reader);
     }
 
     #[cfg(feature = "mmap")]
@@ -492,7 +500,7 @@ mod tests {
                 .try_into()
                 .unwrap();
             assert_eq!(reader.searcher().num_docs(), 0);
-            test_index_on_commit_reload_policy_aux(field, &index, &reader);
+            test_index_on_commit_reload_policy_aux(field, index.clone(), &index, &reader);
         }
 
         #[test]
@@ -534,12 +542,16 @@ mod tests {
                 .try_into()
                 .unwrap();
             assert_eq!(reader.searcher().num_docs(), 0);
-            test_index_on_commit_reload_policy_aux(field, &write_index, &reader);
+            test_index_on_commit_reload_policy_aux(field, read_index, &write_index, &reader);
         }
     }
 
-    fn test_index_on_commit_reload_policy_aux(field: Field, index: &Index, reader: &IndexReader) {
-        let mut reader_index = reader.index();
+    fn test_index_on_commit_reload_policy_aux(
+        field: Field,
+        mut reader_index: Index,
+        index: &Index,
+        reader: &IndexReader,
+    ) {
         let (sender, receiver) = crossbeam::channel::unbounded();
         let _watch_handle = reader_index.directory_mut().watch(Box::new(move || {
             let _ = sender.send(());

src/core/index_meta.rs

@@ -3,7 +3,9 @@ use crate::core::SegmentId;
 use crate::schema::Schema;
 use crate::Opstamp;
 use census::{Inventory, TrackedObject};
+use serde;
 use serde::{Deserialize, Serialize};
+use serde_json;
 use std::collections::HashSet;
 use std::fmt;
 use std::path::PathBuf;

src/core/searcher.rs

@@ -4,6 +4,7 @@ use crate::core::Executor;
 use crate::core::InvertedIndexReader;
 use crate::core::SegmentReader;
 use crate::query::Query;
+use crate::query::Scorer;
 use crate::query::Weight;
 use crate::schema::Document;
 use crate::schema::Schema;
@@ -23,9 +24,17 @@ fn collect_segment<C: Collector>(
     segment_reader: &SegmentReader,
 ) -> crate::Result<C::Fruit> {
     let mut scorer = weight.scorer(segment_reader, 1.0f32)?;
-    let segment_collector =
-        collector.for_segment(segment_ord as u32, segment_reader)?;
-    Ok(segment_collector.collect_scorer(&mut scorer, segment_reader.delete_bitset()))
+    let mut segment_collector = collector.for_segment(segment_ord as u32, segment_reader)?;
+    if let Some(delete_bitset) = segment_reader.delete_bitset() {
+        scorer.for_each(&mut |doc, score| {
+            if delete_bitset.is_alive(doc) {
+                segment_collector.collect(doc, score);
+            }
+        });
+    } else {
+        scorer.for_each(&mut |doc, score| segment_collector.collect(doc, score));
+    }
+    Ok(segment_collector.harvest())
 }
 
 /// Holds a list of `SegmentReader`s ready for search.

src/core/segment.rs

@@ -3,41 +3,140 @@ use crate::core::Index;
 use crate::core::SegmentId;
 use crate::core::SegmentMeta;
 use crate::directory::error::{OpenReadError, OpenWriteError};
-use crate::directory::Directory;
+use crate::directory::{Directory, ManagedDirectory, RAMDirectory};
 use crate::directory::{ReadOnlySource, WritePtr};
 use crate::indexer::segment_serializer::SegmentSerializer;
 use crate::schema::Schema;
 use crate::Opstamp;
 use std::fmt;
+use std::ops::{Deref, DerefMut};
 use std::path::PathBuf;
+use crate::indexer::{ResourceManager};
+
+#[derive(Clone)]
+pub(crate) enum SegmentDirectory {
+    Persisted(ManagedDirectory),
+    Volatile {
+        directory: RAMDirectory,
+    },
+}
+
+impl SegmentDirectory {
+    fn new_volatile(memory_manager: ResourceManager) -> SegmentDirectory {
+        SegmentDirectory::Volatile {
+            directory: RAMDirectory::create_with_memory_manager(memory_manager),
+        }
+    }
+}
+
+impl fmt::Debug for SegmentDirectory {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        match self {
+            SegmentDirectory::Volatile { .. } => write!(f, "volatile")?,
+            SegmentDirectory::Persisted(dir) => write!(f, "Persisted({:?})", dir)?,
+        }
+        Ok(())
+    }
+}
+
+impl From<ManagedDirectory> for SegmentDirectory {
+    fn from(directory: ManagedDirectory) -> Self {
+        SegmentDirectory::Persisted(directory)
+    }
+}
+
+impl Deref for SegmentDirectory {
+    type Target = dyn Directory;
+
+    fn deref(&self) -> &Self::Target {
+        match self {
+            SegmentDirectory::Volatile {
+                directory, ..
+            } => directory,
+            SegmentDirectory::Persisted(dir) => dir,
+        }
+    }
+}
+
+impl DerefMut for SegmentDirectory {
+    fn deref_mut(&mut self) -> &mut Self::Target {
+        match self {
+            SegmentDirectory::Volatile { directory, .. } => directory,
+            SegmentDirectory::Persisted(dir) => dir,
+        }
+    }
+}
 
 /// A segment is a piece of the index.
 #[derive(Clone)]
 pub struct Segment {
-    index: Index,
+    schema: Schema,
     meta: SegmentMeta,
+    directory: SegmentDirectory,
 }
 
 impl fmt::Debug for Segment {
     fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        write!(f, "Segment({:?})", self.id().uuid_string())
+        write!(
+            f,
+            "Segment(id={:?}, directory={:?})",
+            self.id().uuid_string(),
+            self.directory
+        )
     }
 }
 
 impl Segment {
-    /// Creates a new segment given an `Index` and a `SegmentId`
-    pub(crate) fn for_index(index: Index, meta: SegmentMeta) -> Segment {
-        Segment { index, meta }
+    pub(crate) fn new_persisted(
+        meta: SegmentMeta,
+        directory: ManagedDirectory,
+        schema: Schema,
+    ) -> Segment {
+        Segment {
+            meta,
+            schema,
+            directory: SegmentDirectory::from(directory),
+        }
     }
 
-    /// Returns the index the segment belongs to.
-    pub fn index(&self) -> &Index {
-        &self.index
+    /// Creates a new segment that embeds its own `RAMDirectory`.
+    ///
+    /// That segment is entirely dissociated from the index directory.
+    /// It will be persisted by a background thread in charge of IO.
+    pub fn new_volatile(meta: SegmentMeta, schema: Schema, resource_manager: ResourceManager) -> Segment {
+        Segment {
+            schema,
+            meta,
+            directory: SegmentDirectory::new_volatile(resource_manager),
+        }
+    }
+
+    /// Creates a new segment given an `Index` and a `SegmentId`
+    pub(crate) fn for_index(index: Index, meta: SegmentMeta) -> Segment {
+        let segment_directory = index.directory().clone();
+        Segment::new_persisted(meta, segment_directory, index.schema())
+    }
+
+    /// Persists a given `Segment` to a directory.
+    pub fn persist(&mut self, mut dest_directory: ManagedDirectory) -> crate::Result<()> {
+        if let SegmentDirectory::Persisted(_) = self.directory {
+            // this segment is already persisted.
+            return Ok(());
+        }
+        if let SegmentDirectory::Volatile { directory, ..} = &self.directory {
+            directory.persist(&mut dest_directory)?;
+        }
+        self.directory = SegmentDirectory::Persisted(dest_directory);
+        Ok(())
+    }
+
+    pub fn into_volatile(&self, memory_manager: ResourceManager) -> Segment {
+        Segment::new_volatile(self.meta.clone(), self.schema.clone(), memory_manager)
     }
 
     /// Returns our index's schema.
     pub fn schema(&self) -> Schema {
-        self.index.schema()
+        self.schema.clone()
     }
 
     /// Returns the segment meta-information
@@ -51,16 +150,18 @@ impl Segment {
     /// as we finalize a fresh new segment.
     pub(crate) fn with_max_doc(self, max_doc: u32) -> Segment {
         Segment {
-            index: self.index,
+            schema: self.schema,
             meta: self.meta.with_max_doc(max_doc),
+            directory: self.directory,
         }
     }
 
     #[doc(hidden)]
     pub fn with_delete_meta(self, num_deleted_docs: u32, opstamp: Opstamp) -> Segment {
         Segment {
-            index: self.index,
+            schema: self.schema,
             meta: self.meta.with_delete_meta(num_deleted_docs, opstamp),
+            directory: self.directory,
         }
     }
 
@@ -80,15 +181,15 @@ impl Segment {
     /// Open one of the component file for a *regular* read.
     pub fn open_read(&self, component: SegmentComponent) -> Result<ReadOnlySource, OpenReadError> {
         let path = self.relative_path(component);
-        let source = self.index.directory().open_read(&path)?;
+        let source = self.directory.open_read(&path)?;
         Ok(source)
     }
 
     /// Open one of the component file for *regular* write.
     pub fn open_write(&mut self, component: SegmentComponent) -> Result<WritePtr, OpenWriteError> {
         let path = self.relative_path(component);
-        let write = self.index.directory_mut().open_write(&path)?;
-        Ok(write)
+        let wrt = self.directory.open_write(&path)?;
+        Ok(wrt)
     }
 }
 

src/directory/managed_directory.rs

@@ -11,6 +11,7 @@ use crate::error::DataCorruption;
 use crate::Directory;
 
 use crc32fast::Hasher;
+use serde_json;
 use std::collections::HashSet;
 use std::io;
 use std::io::Write;

src/directory/mmap_directory.rs

@@ -1,3 +1,10 @@
+use fs2;
+use notify;
+
+use self::fs2::FileExt;
+use self::notify::RawEvent;
+use self::notify::RecursiveMode;
+use self::notify::Watcher;
 use crate::core::META_FILEPATH;
 use crate::directory::error::LockError;
 use crate::directory::error::{
@@ -13,11 +20,8 @@ use crate::directory::WatchCallback;
 use crate::directory::WatchCallbackList;
 use crate::directory::WatchHandle;
 use crate::directory::{TerminatingWrite, WritePtr};
-use fs2::FileExt;
+use atomicwrites;
 use memmap::Mmap;
-use notify::RawEvent;
-use notify::RecursiveMode;
-use notify::Watcher;
 use serde::{Deserialize, Serialize};
 use std::collections::HashMap;
 use std::convert::From;
@@ -220,13 +224,17 @@ struct MmapDirectoryInner {
 }
 
 impl MmapDirectoryInner {
-    fn new(root_path: PathBuf, temp_directory: Option<TempDir>) -> MmapDirectoryInner {
-        MmapDirectoryInner {
+    fn new(
+        root_path: PathBuf,
+        temp_directory: Option<TempDir>,
+    ) -> Result<MmapDirectoryInner, OpenDirectoryError> {
+        let mmap_directory_inner = MmapDirectoryInner {
             root_path,
             mmap_cache: Default::default(),
             _temp_directory: temp_directory,
             watcher: RwLock::new(None),
-        }
+        };
+        Ok(mmap_directory_inner)
     }
 
     fn watch(&self, watch_callback: WatchCallback) -> crate::Result<WatchHandle> {
@@ -260,11 +268,14 @@ impl fmt::Debug for MmapDirectory {
 }
 
 impl MmapDirectory {
-    fn new(root_path: PathBuf, temp_directory: Option<TempDir>) -> MmapDirectory {
-        let inner = MmapDirectoryInner::new(root_path, temp_directory);
-        MmapDirectory {
+    fn new(
+        root_path: PathBuf,
+        temp_directory: Option<TempDir>,
+    ) -> Result<MmapDirectory, OpenDirectoryError> {
+        let inner = MmapDirectoryInner::new(root_path, temp_directory)?;
+        Ok(MmapDirectory {
             inner: Arc::new(inner),
-        }
+        })
     }
 
     /// Creates a new MmapDirectory in a temporary directory.
@@ -274,7 +285,7 @@ impl MmapDirectory {
     pub fn create_from_tempdir() -> Result<MmapDirectory, OpenDirectoryError> {
         let tempdir = TempDir::new().map_err(OpenDirectoryError::IoError)?;
         let tempdir_path = PathBuf::from(tempdir.path());
-        Ok(MmapDirectory::new(tempdir_path, Some(tempdir)))
+        MmapDirectory::new(tempdir_path, Some(tempdir))
     }
 
     /// Opens a MmapDirectory in a directory.
@@ -292,7 +303,7 @@ impl MmapDirectory {
                 directory_path,
             )))
         } else {
-            Ok(MmapDirectory::new(PathBuf::from(directory_path), None))
+            Ok(MmapDirectory::new(PathBuf::from(directory_path), None)?)
         }
     }
 

src/directory/mod.rs

@@ -13,7 +13,10 @@ mod footer;
 mod managed_directory;
 mod ram_directory;
 mod read_only_source;
+mod spilling_writer;
 mod watch_event_router;
+mod persistor;
+
 
 /// Errors specific to the directory module.
 pub mod error;
@@ -23,6 +26,7 @@ pub use self::directory::{Directory, DirectoryClone};
 pub use self::directory_lock::{Lock, INDEX_WRITER_LOCK, META_LOCK};
 pub use self::ram_directory::RAMDirectory;
 pub use self::read_only_source::ReadOnlySource;
+pub(crate) use self::spilling_writer::{SpillingResult, SpillingWriter};
 pub use self::watch_event_router::{WatchCallback, WatchCallbackList, WatchHandle};
 use std::io::{self, BufWriter, Write};
 use std::path::PathBuf;
@@ -79,10 +83,16 @@ impl<W: TerminatingWrite> TerminatingWrite for BufWriter<W> {
     }
 }
 
+impl TerminatingWrite for Vec<u8> {
+    fn terminate_ref(&mut self, _a: AntiCallToken) -> io::Result<()> {
+        Ok(())
+    }
+}
+
 #[cfg(test)]
 impl<'a> TerminatingWrite for &'a mut Vec<u8> {
     fn terminate_ref(&mut self, _a: AntiCallToken) -> io::Result<()> {
-        self.flush()
+        Ok(())
     }
 }
 

src/directory/persistor.rs

@@ -0,0 +1,38 @@
+use crate::indexer::{SegmentManager, ResourceManager, MergeOperationInventory};
+use std::thread::JoinHandle;
+use crate::{IndexWriterConfig, SegmentId};
+use std::collections::HashSet;
+
+pub(crate) struct Persistor {
+    memory_manager: ResourceManager,
+    thread_handle: JoinHandle<()>,
+}
+
+impl Persistor {
+    pub(crate) fn create_and_start(segment_manager: SegmentManager,
+                            memory_manager: ResourceManager,
+                            config: IndexWriterConfig) -> crate::Result<Persistor> {
+        let memory_manager_clone = memory_manager.clone();
+        let thread_handle = std::thread::Builder::new()
+            .name("persistor-thread".to_string())
+            .spawn(move || {
+                while let Ok(_) = memory_manager_clone.wait_until_in_range(config.persist_low..) {
+                    segment_manager.largest_segment_not_in_merge();
+                }
+            }).map_err(|_err| crate::TantivyError::ErrorInThread("Failed to start persistor thread.".to_string()))?;
+        Ok(Persistor {
+            memory_manager,
+            thread_handle
+        })
+    }
+
+    /// Stop the persisting thread.
+    ///
+    /// The memory manager will be terminated, which will unlock the thread from any waiting
+    /// position.
+    /// This method blocks for a short amount of tim until the persistor thread has terminated.
+    pub fn stop(self) {
+        self.memory_manager.terminate();
+        let _ = self.thread_handle.join();
+    }
+}

src/directory/ram_directory.rs

@@ -11,6 +11,7 @@ use std::io::{self, BufWriter, Cursor, Seek, SeekFrom, Write};
 use std::path::{Path, PathBuf};
 use std::result;
 use std::sync::{Arc, RwLock};
+use crate::indexer::ResourceManager;
 
 /// Writer associated with the `RAMDirectory`
 ///
@@ -82,11 +83,12 @@ impl TerminatingWrite for VecWriter {
 struct InnerDirectory {
     fs: HashMap<PathBuf, ReadOnlySource>,
     watch_router: WatchCallbackList,
+    memory_manager: ResourceManager,
 }
 
 impl InnerDirectory {
     fn write(&mut self, path: PathBuf, data: &[u8]) -> bool {
-        let data = ReadOnlySource::new(Vec::from(data));
+        let data = ReadOnlySource::new_with_allocation(Vec::from(data), &self.memory_manager);
         self.fs.insert(path, data).is_some()
     }
 
@@ -112,8 +114,8 @@ impl InnerDirectory {
         self.watch_router.subscribe(watch_handle)
     }
 
-    fn total_mem_usage(&self) -> usize {
-        self.fs.values().map(|f| f.len()).sum()
+    fn total_mem_usage(&self) -> u64 {
+        self.fs.values().map(|source| source.len() as u64).sum()
     }
 }
 
@@ -134,14 +136,30 @@ pub struct RAMDirectory {
 }
 
 impl RAMDirectory {
-    /// Constructor
+
+    /// Creates a new RAMDirectory.
+    ///
+    /// Check `.create_with_memory_manager(..)` if you want to associate an external memory
+    /// manager to your RAMDirectory.
     pub fn create() -> RAMDirectory {
-        Self::default()
+        RAMDirectory::default()
+    }
+
+    /// Constructor
+    pub fn create_with_memory_manager(memory_manager: ResourceManager) -> RAMDirectory {
+        let inner_directory = InnerDirectory {
+            fs: Default::default(),
+            watch_router: Default::default(),
+            memory_manager
+        };
+        RAMDirectory {
+            fs: Arc::new(RwLock::new(inner_directory))
+        }
     }
 
     /// Returns the sum of the size of the different files
     /// in the RAMDirectory.
-    pub fn total_mem_usage(&self) -> usize {
+    pub fn total_mem_usage(&self) -> u64 {
         self.fs.read().unwrap().total_mem_usage()
     }
 
@@ -227,6 +245,9 @@ mod tests {
     use crate::Directory;
     use std::io::Write;
     use std::path::Path;
+    use crate::indexer::ResourceManager;
+    use crate::directory::TerminatingWrite;
+    use std::mem;
 
     #[test]
     fn test_persist() {
@@ -244,4 +265,59 @@ mod tests {
         assert_eq!(directory_copy.atomic_read(path_atomic).unwrap(), msg_atomic);
         assert_eq!(directory_copy.atomic_read(path_seq).unwrap(), msg_seq);
     }
+
+    #[test]
+    fn test_memory_manager_several_path() {
+        let memory_manager = ResourceManager::default();
+        let mut ram_directory = RAMDirectory::create_with_memory_manager(memory_manager.clone());
+        assert!(ram_directory.atomic_write(Path::new("/titi"), b"abcd").is_ok());
+        assert_eq!(memory_manager.total_amount(), 4u64);
+        assert!(ram_directory.atomic_write(Path::new("/toto"), b"abcde").is_ok());
+        assert_eq!(memory_manager.total_amount(), 9u64);
+    }
+
+    #[test]
+    fn test_memory_manager_override() {
+        let memory_manager = ResourceManager::default();
+        let mut ram_directory = RAMDirectory::create_with_memory_manager(memory_manager.clone());
+        assert!(ram_directory.atomic_write(Path::new("/titi"), b"abcde").is_ok());
+        assert_eq!(memory_manager.total_amount(), 5u64);
+        assert!(ram_directory.atomic_write(Path::new("/titi"), b"abcdef").is_ok());
+        assert_eq!(memory_manager.total_amount(), 6u64);
+    }
+
+    #[test]
+    fn test_memory_manager_seq_wrt() {
+        let memory_manager = ResourceManager::default();
+        let mut ram_directory = RAMDirectory::create_with_memory_manager(memory_manager.clone());
+        let mut wrt = ram_directory.open_write(Path::new("/titi")).unwrap();
+        assert!(wrt.write_all(b"abcde").is_ok());
+        assert!(wrt.terminate().is_ok());
+        assert_eq!(memory_manager.total_amount(), 5u64);
+        assert!(ram_directory.atomic_write(Path::new("/titi"), b"abcdef").is_ok());
+        assert_eq!(memory_manager.total_amount(), 6u64);
+    }
+
+    #[test]
+    fn test_release_on_drop() {
+        let memory_manager = ResourceManager::default();
+        let mut ram_directory = RAMDirectory::create_with_memory_manager(memory_manager.clone());
+        let mut wrt = ram_directory.open_write(Path::new("/titi")).unwrap();
+        assert!(wrt.write_all(b"abcde").is_ok());
+        assert!(wrt.terminate().is_ok());
+        assert_eq!(memory_manager.total_amount(), 5u64);
+        let mut wrt2 = ram_directory.open_write(Path::new("/toto")).unwrap();
+        assert!(wrt2.write_all(b"abcdefghijkl").is_ok());
+        assert!(wrt2.terminate().is_ok());
+        assert_eq!(memory_manager.total_amount(), 17u64);
+        let source = ram_directory.open_read(Path::new("/titi")).unwrap();
+        let source_clone = source.clone();
+        assert_eq!(memory_manager.total_amount(), 17u64);
+        mem::drop(ram_directory);
+        assert_eq!(memory_manager.total_amount(), 5u64);
+        mem::drop(source);
+        assert_eq!(memory_manager.total_amount(), 5u64);
+        mem::drop(source_clone);
+        assert_eq!(memory_manager.total_amount(), 0u64);
+    }
 }

src/directory/read_only_source.rs

@@ -2,6 +2,7 @@ use crate::common::HasLen;
 use stable_deref_trait::{CloneStableDeref, StableDeref};
 use std::ops::Deref;
 use std::sync::Arc;
+use crate::indexer::{Allocation, ResourceManager};
 
 pub type BoxedData = Box<dyn Deref<Target = [u8]> + Send + Sync + 'static>;
 
@@ -15,6 +16,7 @@ pub struct ReadOnlySource {
     data: Arc<BoxedData>,
     start: usize,
     stop: usize,
+    allocation: Option<Arc<Allocation>>
 }
 
 unsafe impl StableDeref for ReadOnlySource {}
@@ -35,11 +37,13 @@ impl From<Arc<BoxedData>> for ReadOnlySource {
             data,
             start: 0,
             stop: len,
+            allocation: None
         }
     }
 }
 
 impl ReadOnlySource {
+
     pub(crate) fn new<D>(data: D) -> ReadOnlySource
     where
         D: Deref<Target = [u8]> + Send + Sync + 'static,
@@ -49,9 +53,24 @@ impl ReadOnlySource {
             data: Arc::new(Box::new(data)),
             start: 0,
             stop: len,
+            allocation: None
         }
     }
 
+    pub(crate) fn new_with_allocation<D>(data: D, memory_manager: &ResourceManager) -> ReadOnlySource
+        where
+            D: Deref<Target = [u8]> + Send + Sync + 'static,
+    {
+        let len = data.len();
+        ReadOnlySource {
+            data: Arc::new(Box::new(data)),
+            start: 0,
+            stop: len,
+            allocation: Some(Arc::new(memory_manager.allocate(len as u64)))
+        }
+    }
+
+
     /// Creates an empty ReadOnlySource
     pub fn empty() -> ReadOnlySource {
         ReadOnlySource::new(&[][..])
@@ -98,6 +117,7 @@ impl ReadOnlySource {
             data: self.data.clone(),
             start: self.start + start,
             stop: self.start + stop,
+            allocation: self.allocation.clone()
         }
     }
 

src/directory/spilling_writer.rs

@@ -0,0 +1,186 @@
+use crate::common::MutableEnum;
+use crate::directory::{TerminatingWrite, WritePtr};
+use std::io::{self, Write};
+
+/// Represents the state of the `SpillingWriter`.
+enum SpillingState {
+    Buffer {
+        buffer: Vec<u8>,
+        capacity: usize,
+        write_factory: Box<dyn FnOnce() -> io::Result<WritePtr>>,
+    },
+    Spilled(WritePtr),
+}
+
+impl SpillingState {
+    fn new(
+        limit: usize,
+        write_factory: Box<dyn FnOnce() -> io::Result<WritePtr>>,
+    ) -> SpillingState {
+        SpillingState::Buffer {
+            buffer: Vec::with_capacity(limit),
+            capacity: limit,
+            write_factory,
+        }
+    }
+
+    // Change the state in such a way that it is ready to accept
+    // `extra_capacity` bytes.
+    //
+    fn reserve(self, extra_capacity: usize) -> io::Result<SpillingState> {
+        match self {
+            SpillingState::Buffer {
+                buffer,
+                capacity,
+                write_factory,
+            } => {
+                if capacity >= extra_capacity {
+                    Ok(SpillingState::Buffer {
+                        buffer,
+                        capacity: capacity - extra_capacity,
+                        write_factory,
+                    })
+                } else {
+                    let mut wrt = write_factory()?;
+                    wrt.write_all(&buffer[..])?;
+                    Ok(SpillingState::Spilled(wrt))
+                }
+            }
+            SpillingState::Spilled(wrt) => Ok(SpillingState::Spilled(wrt)),
+        }
+    }
+}
+
+/// The `SpillingWriter` is a writer that start by writing in a
+/// buffer.
+///
+/// Once a memory limit is reached, the spilling writer will
+/// call a given `WritePtr` factory and start spilling into it.
+///
+/// Spilling here includes:
+/// - writing all of the data that were written in the in-memory buffer so far
+/// - writing subsequent data as well.
+///
+/// Once entering "spilling" mode, the `SpillingWriter` stays in this mode.
+pub struct SpillingWriter {
+    state: MutableEnum<SpillingState>,
+}
+
+impl SpillingWriter {
+    //// Creates a new `Spilling Writer`.
+    pub fn new(
+        limit: usize,
+        write_factory: Box<dyn FnOnce() -> io::Result<WritePtr>>,
+    ) -> SpillingWriter {
+        let state = SpillingState::new(limit, write_factory);
+        SpillingWriter {
+            state: MutableEnum::wrap(state),
+        }
+    }
+
+    /// Finalizes the `SpillingWriter`.
+    ///
+    /// The `SpillingResult` object is an enum specific
+    /// to whether the `SpillingWriter` reached the spilling limit
+    /// (In that case, the buffer is returned).
+    ///
+    /// If the writer reached the spilling mode, the underlying `WritePtr`
+    /// is terminated and SpillingResult::Spilled is returned.
+    pub fn finalize(self) -> io::Result<SpillingResult> {
+        match self.state.into() {
+            SpillingState::Spilled(wrt) => {
+                wrt.terminate()?;
+                Ok(SpillingResult::Spilled)
+            }
+            SpillingState::Buffer { buffer, .. } => Ok(SpillingResult::Buffer(buffer)),
+        }
+    }
+}
+
+/// enum used as the result of `.finalize()`.
+pub enum SpillingResult {
+    Spilled,
+    Buffer(Vec<u8>),
+}
+
+impl io::Write for SpillingWriter {
+    fn write(&mut self, payload: &[u8]) -> io::Result<usize> {
+        self.write_all(payload)?;
+        Ok(payload.len())
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        if let SpillingState::Spilled(wrt) = &mut *self.state {
+            wrt.flush()?;
+        }
+        Ok(())
+    }
+
+    fn write_all(&mut self, payload: &[u8]) -> io::Result<()> {
+        self.state.map_mutate(|mut state| {
+            state = state.reserve(payload.len())?;
+            match &mut state {
+                SpillingState::Buffer { buffer, .. } => {
+                    buffer.extend_from_slice(payload);
+                }
+                SpillingState::Spilled(wrt) => {
+                    wrt.write_all(payload)?;
+                }
+            }
+            Ok(state)
+        })
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::SpillingWriter;
+    use crate::directory::spilling_writer::SpillingResult;
+    use crate::directory::RAMDirectory;
+    use crate::Directory;
+    use std::io::{self, Write};
+    use std::path::Path;
+
+    #[test]
+    fn test_no_spilling() {
+        let ram_directory = RAMDirectory::create();
+        let mut ram_directory_clone = ram_directory.clone();
+        let path = Path::new("test");
+        let write_factory = Box::new(move || {
+            ram_directory_clone
+                .open_write(path)
+                .map_err(|err| io::Error::new(io::ErrorKind::Other, err))
+        });
+        let mut spilling_wrt = SpillingWriter::new(10, write_factory);
+        assert!(spilling_wrt.write_all(b"abcd").is_ok());
+        if let SpillingResult::Buffer(buf) = spilling_wrt.finalize().unwrap() {
+            assert_eq!(buf, b"abcd")
+        } else {
+            panic!("spill writer should not have spilled");
+        }
+        assert!(!ram_directory.exists(path));
+    }
+
+    #[test]
+    fn test_spilling() {
+        let ram_directory = RAMDirectory::create();
+        let mut ram_directory_clone = ram_directory.clone();
+        let path = Path::new("test");
+        let write_factory = Box::new(move || {
+            ram_directory_clone
+                .open_write(path)
+                .map_err(|err| io::Error::new(io::ErrorKind::Other, err))
+        });
+        let mut spilling_wrt = SpillingWriter::new(10, write_factory);
+        assert!(spilling_wrt.write_all(b"abcd").is_ok());
+        assert!(spilling_wrt.write_all(b"efghijklmnop").is_ok());
+        if let SpillingResult::Spilled = spilling_wrt.finalize().unwrap() {
+        } else {
+            panic!("spill writer should have spilled");
+        }
+        assert_eq!(
+            ram_directory.atomic_read(path).unwrap(),
+            b"abcdefghijklmnop"
+        );
+    }
+}

src/error.rs

@@ -7,6 +7,7 @@ use crate::directory::error::{Incompatibility, LockError};
 use crate::fastfield::FastFieldNotAvailableError;
 use crate::query;
 use crate::schema;
+use serde_json;
 use std::fmt;
 use std::path::PathBuf;
 use std::sync::PoisonError;

src/fastfield/delete.rs

@@ -89,7 +89,7 @@ mod tests {
 
     fn test_delete_bitset_helper(bitset: &BitSet, max_doc: u32) {
         let test_path = PathBuf::from("test");
-        let mut directory = RAMDirectory::create();
+        let mut directory = RAMDirectory::default();
         {
             let mut writer = directory.open_write(&*test_path).unwrap();
             write_delete_bitset(bitset, max_doc, &mut writer).unwrap();

src/indexer/index_writer.rs

@@ -8,22 +8,26 @@ use crate::core::SegmentComponent;
 use crate::core::SegmentId;
 use crate::core::SegmentMeta;
 use crate::core::SegmentReader;
-use crate::directory::TerminatingWrite;
 use crate::directory::{DirectoryLock, GarbageCollectionResult};
+use crate::directory::{TerminatingWrite, WatchCallbackList};
 use crate::docset::DocSet;
 use crate::error::TantivyError;
 use crate::fastfield::write_delete_bitset;
 use crate::indexer::delete_queue::{DeleteCursor, DeleteQueue};
 use crate::indexer::doc_opstamp_mapping::DocToOpstampMapping;
 use crate::indexer::operation::DeleteOperation;
+use crate::indexer::segment_manager::SegmentRegisters;
+use crate::indexer::segment_register::SegmentRegister;
 use crate::indexer::stamper::Stamper;
-use crate::indexer::MergePolicy;
-use crate::indexer::SegmentEntry;
+use crate::indexer::{SegmentEntry, ResourceManager};
 use crate::indexer::SegmentWriter;
+use crate::indexer::{IndexWriterConfig, MergePolicy};
+use crate::reader::NRTReader;
 use crate::schema::Document;
 use crate::schema::IndexRecordOption;
 use crate::schema::Term;
-use crate::Opstamp;
+use crate::tokenizer::TokenizerManager;
+use crate::{IndexReader, Opstamp};
 use crossbeam::channel;
 use futures::executor::block_on;
 use futures::future::Future;
@@ -31,18 +35,10 @@ use smallvec::smallvec;
 use smallvec::SmallVec;
 use std::mem;
 use std::ops::Range;
-use std::sync::Arc;
+use std::sync::{Arc, RwLock};
 use std::thread;
 use std::thread::JoinHandle;
 
-// Size of the margin for the heap. A segment is closed when the remaining memory
-// in the heap goes below MARGIN_IN_BYTES.
-pub const MARGIN_IN_BYTES: usize = 1_000_000;
-
-// We impose the memory per thread to be at least 3 MB.
-pub const HEAP_SIZE_MIN: usize = ((MARGIN_IN_BYTES as u32) * 3u32) as usize;
-pub const HEAP_SIZE_MAX: usize = u32::max_value() as usize - MARGIN_IN_BYTES;
-
 // Add document will block if the number of docs waiting in the queue to be indexed
 // reaches `PIPELINE_MAX_SIZE_IN_DOCS`
 const PIPELINE_MAX_SIZE_IN_DOCS: usize = 10_000;
@@ -69,8 +65,9 @@ pub struct IndexWriter {
     _directory_lock: Option<DirectoryLock>,
 
     index: Index,
+    config: IndexWriterConfig,
 
-    heap_size_in_bytes_per_thread: usize,
+    segment_registers: Arc<RwLock<SegmentRegisters>>,
 
     workers_join_handle: Vec<JoinHandle<crate::Result<()>>>,
 
@@ -80,13 +77,14 @@ pub struct IndexWriter {
     segment_updater: SegmentUpdater,
 
     worker_id: usize,
-
-    num_threads: usize,
-
     delete_queue: DeleteQueue,
 
     stamper: Stamper,
     committed_opstamp: Opstamp,
+
+    on_commit: WatchCallbackList,
+
+    memory_manager: ResourceManager,
 }
 
 fn compute_deleted_bitset(
@@ -133,7 +131,6 @@ fn compute_deleted_bitset(
 /// For instance, there was no delete operation between the state of the `segment_entry` and
 /// the `target_opstamp`, `segment_entry` is not updated.
 pub(crate) fn advance_deletes(
-    mut segment: Segment,
     segment_entry: &mut SegmentEntry,
     target_opstamp: Opstamp,
 ) -> crate::Result<()> {
@@ -142,25 +139,33 @@ pub(crate) fn advance_deletes(
         return Ok(());
     }
 
-    if segment_entry.delete_bitset().is_none() && segment_entry.delete_cursor().get().is_none() {
+    let delete_bitset_opt = segment_entry.take_delete_bitset();
+
+    // We avoid directly advancing the `SegmentEntry` delete cursor, because
+    // we do not want to end up in an invalid state if the delete bitset
+    // serialization fails.
+    let mut delete_cursor = segment_entry.delete_cursor();
+
+    if delete_bitset_opt.is_none() && segment_entry.delete_cursor().get().is_none() {
         // There has been no `DeleteOperation` between the segment status and `target_opstamp`.
         return Ok(());
     }
 
+    // We open our current serialized segment to compute the new deleted bitset.
+    let segment = segment_entry.segment().clone();
     let segment_reader = SegmentReader::open(&segment)?;
 
     let max_doc = segment_reader.max_doc();
-    let mut delete_bitset: BitSet = match segment_entry.delete_bitset() {
-        Some(previous_delete_bitset) => (*previous_delete_bitset).clone(),
-        None => BitSet::with_max_value(max_doc),
-    };
+
+    let mut delete_bitset: BitSet =
+        delete_bitset_opt.unwrap_or_else(|| BitSet::with_max_value(max_doc));
 
     let num_deleted_docs_before = segment.meta().num_deleted_docs();
 
     compute_deleted_bitset(
         &mut delete_bitset,
         &segment_reader,
-        segment_entry.delete_cursor(),
+        &mut delete_cursor,
         &DocToOpstampMapping::None,
         target_opstamp,
     )?;
@@ -179,32 +184,40 @@ pub(crate) fn advance_deletes(
     let num_deleted_docs: u32 = delete_bitset.len() as u32;
     if num_deleted_docs > num_deleted_docs_before {
         // There are new deletes. We need to write a new delete file.
-        segment = segment.with_delete_meta(num_deleted_docs as u32, target_opstamp);
-        let mut delete_file = segment.open_write(SegmentComponent::DELETE)?;
+        let mut delete_file = segment
+            .with_delete_meta(num_deleted_docs as u32, target_opstamp)
+            .open_write(SegmentComponent::DELETE)?;
         write_delete_bitset(&delete_bitset, max_doc, &mut delete_file)?;
         delete_file.terminate()?;
+        segment_entry.reset_delete_meta(num_deleted_docs as u32, target_opstamp);
     }
 
-    segment_entry.set_meta(segment.meta().clone());
+    // Regardless of whether we did end up having to write a new file or not
+    // we advance the `delete_cursor`. This is an optimisation. We want to ensure we do not
+    // check that a given deleted term does not match any of our docs more than once.
+    segment_entry.set_delete_cursor(delete_cursor);
     Ok(())
 }
 
 fn index_documents(
-    memory_budget: usize,
+    config: IndexWriterConfig,
     segment: Segment,
     grouped_document_iterator: &mut dyn Iterator<Item = OperationGroup>,
     segment_updater: &mut SegmentUpdater,
+    tokenizers: &TokenizerManager,
     mut delete_cursor: DeleteCursor,
+    memory_manager: ResourceManager
 ) -> crate::Result<bool> {
     let schema = segment.schema();
 
-    let mut segment_writer = SegmentWriter::for_segment(memory_budget, segment.clone(), &schema)?;
+    let mut segment_writer = SegmentWriter::for_segment(
+        &config, segment, &schema, tokenizers, memory_manager)?;
     for document_group in grouped_document_iterator {
         for doc in document_group {
             segment_writer.add_document(doc, &schema)?;
         }
         let mem_usage = segment_writer.mem_usage();
-        if mem_usage >= memory_budget - MARGIN_IN_BYTES {
+        if mem_usage >= config.heap_size_before_flushing() {
             info!(
                 "Buffer limit reached, flushing segment with maxdoc={}.",
                 segment_writer.max_doc()
@@ -223,24 +236,14 @@ fn index_documents(
     // the worker thread.
     assert!(max_doc > 0);
 
-    let doc_opstamps: Vec<Opstamp> = segment_writer.finalize()?;
-
-    let segment_with_max_doc = segment.with_max_doc(max_doc);
+    let (segment, doc_opstamps): (Segment, Vec<Opstamp>) = segment_writer.finalize()?;
 
     let last_docstamp: Opstamp = *(doc_opstamps.last().unwrap());
 
-    let delete_bitset_opt = apply_deletes(
-        &segment_with_max_doc,
-        &mut delete_cursor,
-        &doc_opstamps,
-        last_docstamp,
-    )?;
+    let delete_bitset_opt =
+        apply_deletes(&segment, &mut delete_cursor, &doc_opstamps, last_docstamp)?;
 
-    let segment_entry = SegmentEntry::new(
-        segment_with_max_doc.meta().clone(),
-        delete_cursor,
-        delete_bitset_opt,
-    );
+    let segment_entry = SegmentEntry::new(segment, delete_cursor, delete_bitset_opt);
     block_on(segment_updater.schedule_add_segment(segment_entry))?;
     Ok(true)
 }
@@ -292,21 +295,10 @@ impl IndexWriter {
     /// If the heap size per thread is too small, panics.
     pub(crate) fn new(
         index: &Index,
-        num_threads: usize,
-        heap_size_in_bytes_per_thread: usize,
+        mut config: IndexWriterConfig,
         directory_lock: DirectoryLock,
     ) -> crate::Result<IndexWriter> {
-        if heap_size_in_bytes_per_thread < HEAP_SIZE_MIN {
-            let err_msg = format!(
-                "The heap size per thread needs to be at least {}.",
-                HEAP_SIZE_MIN
-            );
-            return Err(TantivyError::InvalidArgument(err_msg));
-        }
-        if heap_size_in_bytes_per_thread >= HEAP_SIZE_MAX {
-            let err_msg = format!("The heap size per thread cannot exceed {}", HEAP_SIZE_MAX);
-            return Err(TantivyError::InvalidArgument(err_msg));
-        }
+        config.validate()?;
         let (document_sender, document_receiver): (OperationSender, OperationReceiver) =
             channel::bounded(PIPELINE_MAX_SIZE_IN_DOCS);
 
@@ -314,16 +306,26 @@ impl IndexWriter {
 
         let current_opstamp = index.load_metas()?.opstamp;
 
+        let meta = index.load_metas()?;
+
         let stamper = Stamper::new(current_opstamp);
 
+        let commited_segments = SegmentRegister::new(
+            index.directory(),
+            &index.schema(),
+            meta.segments,
+            &delete_queue.cursor(),
+        );
+        let segment_registers = Arc::new(RwLock::new(SegmentRegisters::new(commited_segments)));
+
         let segment_updater =
-            SegmentUpdater::create(index.clone(), stamper.clone(), &delete_queue.cursor())?;
+            SegmentUpdater::create(segment_registers.clone(), index.clone(), stamper.clone())?;
 
         let mut index_writer = IndexWriter {
             _directory_lock: Some(directory_lock),
 
-            heap_size_in_bytes_per_thread,
             index: index.clone(),
+            config,
 
             operation_receiver: document_receiver,
             operation_sender: document_sender,
@@ -331,7 +333,6 @@ impl IndexWriter {
             segment_updater,
 
             workers_join_handle: vec![],
-            num_threads,
 
             delete_queue,
 
@@ -339,6 +340,10 @@ impl IndexWriter {
             stamper,
 
             worker_id: 0,
+            segment_registers,
+            on_commit: Default::default(),
+
+            memory_manager: Default::default()
         };
         index_writer.start_workers()?;
         Ok(index_writer)
@@ -366,23 +371,11 @@ impl IndexWriter {
                 })?;
         }
 
-        let result = self
+        self
             .segment_updater
-            .wait_merging_thread()
-            .map_err(|_| TantivyError::ErrorInThread("Failed to join merging thread.".into()));
+            .wait_merging_thread();
 
-        if let Err(ref e) = result {
-            error!("Some merging thread failed {:?}", e);
-        }
-
-        result
-    }
-
-    #[doc(hidden)]
-    pub fn add_segment(&self, segment_meta: SegmentMeta) -> crate::Result<()> {
-        let delete_cursor = self.delete_queue.cursor();
-        let segment_entry = SegmentEntry::new(segment_meta, delete_cursor, None);
-        block_on(self.segment_updater.schedule_add_segment(segment_entry))
+        Ok(())
     }
 
     /// Creates a new segment.
@@ -405,8 +398,9 @@ impl IndexWriter {
 
         let mut delete_cursor = self.delete_queue.cursor();
 
-        let mem_budget = self.heap_size_in_bytes_per_thread;
         let index = self.index.clone();
+        let config = self.config.clone();
+        let memory_manager = self.memory_manager.clone();
         let join_handle: JoinHandle<crate::Result<()>> = thread::Builder::new()
             .name(format!("thrd-tantivy-index{}", self.worker_id))
             .spawn(move || {
@@ -435,11 +429,13 @@ impl IndexWriter {
                     }
                     let segment = index.new_segment();
                     index_documents(
-                        mem_budget,
+                        config.clone(),
                         segment,
                         &mut document_iterator,
                         &mut segment_updater,
+                        index.tokenizers(),
                         delete_cursor.clone(),
+                        memory_manager.clone()
                     )?;
                 }
             })?;
@@ -459,7 +455,7 @@ impl IndexWriter {
     }
 
     fn start_workers(&mut self) -> crate::Result<()> {
-        for _ in 0..self.num_threads {
+        for _ in 0..self.config.max_indexing_threads {
             self.add_indexing_worker()?;
         }
         Ok(())
@@ -564,12 +560,8 @@ impl IndexWriter {
             .take()
             .expect("The IndexWriter does not have any lock. This is a bug, please report.");
 
-        let new_index_writer: IndexWriter = IndexWriter::new(
-            &self.index,
-            self.num_threads,
-            self.heap_size_in_bytes_per_thread,
-            directory_lock,
-        )?;
+        let new_index_writer: IndexWriter =
+            IndexWriter::new(&self.index, self.config.clone(), directory_lock)?;
 
         // the current `self` is dropped right away because of this call.
         //
@@ -608,7 +600,7 @@ impl IndexWriter {
     /// It is also possible to add a payload to the `commit`
     /// using this API.
     /// See [`PreparedCommit::set_payload()`](PreparedCommit.html)
-    pub fn prepare_commit(&mut self) -> crate::Result<PreparedCommit> {
+    pub fn prepare_commit(&mut self, soft_commit: bool) -> crate::Result<PreparedCommit> {
         // Here, because we join all of the worker threads,
         // all of the segment update for this commit have been
         // sent.
@@ -636,7 +628,7 @@ impl IndexWriter {
         }
 
         let commit_opstamp = self.stamper.stamp();
-        let prepared_commit = PreparedCommit::new(self, commit_opstamp);
+        let prepared_commit = PreparedCommit::new(self, commit_opstamp, soft_commit);
         info!("Prepared commit {}", commit_opstamp);
         Ok(prepared_commit)
     }
@@ -656,7 +648,25 @@ impl IndexWriter {
     /// that made it in the commit.
     ///
     pub fn commit(&mut self) -> crate::Result<Opstamp> {
-        self.prepare_commit()?.commit()
+        self.prepare_commit(false)?.commit()
+    }
+
+    pub fn soft_commit(&mut self) -> crate::Result<Opstamp> {
+        self.prepare_commit(true)?.commit()
+    }
+
+    pub(crate) fn trigger_commit(&self) -> impl Future<Output = ()> {
+        self.on_commit.broadcast()
+    }
+
+    pub fn reader(&self, num_searchers: usize) -> crate::Result<IndexReader> {
+        let nrt_reader = NRTReader::create(
+            num_searchers,
+            self.index.clone(),
+            self.segment_registers.clone(),
+            &self.on_commit,
+        )?;
+        Ok(IndexReader::NRT(nrt_reader))
     }
 
     pub(crate) fn segment_updater(&self) -> &SegmentUpdater {
@@ -1054,7 +1064,8 @@ mod tests {
                 index_writer.add_document(doc!(text_field => "a"));
             }
             {
-                let mut prepared_commit = index_writer.prepare_commit().expect("commit failed");
+                let mut prepared_commit =
+                    index_writer.prepare_commit(false).expect("commit failed");
                 prepared_commit.set_payload("first commit");
                 prepared_commit.commit().expect("commit failed");
             }
@@ -1087,7 +1098,8 @@ mod tests {
                 index_writer.add_document(doc!(text_field => "a"));
             }
             {
-                let mut prepared_commit = index_writer.prepare_commit().expect("commit failed");
+                let mut prepared_commit =
+                    index_writer.prepare_commit(false).expect("commit failed");
                 prepared_commit.set_payload("first commit");
                 prepared_commit.abort().expect("commit failed");
             }
@@ -1265,4 +1277,41 @@ mod tests {
         let commit = index_writer.commit();
         assert!(commit.is_ok());
     }
+
+    #[test]
+    fn test_index_writer_reader() {
+        let mut schema_builder = schema::Schema::builder();
+        let idfield = schema_builder.add_text_field("id", STRING);
+        schema_builder.add_text_field("optfield", STRING);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        index_writer.add_document(doc!(idfield=>"myid"));
+        assert!(index_writer.commit().is_ok());
+        let reader = index_writer.reader(2).unwrap();
+        let searcher = reader.searcher();
+        assert_eq!(searcher.num_docs(), 1u64);
+        index_writer.add_document(doc!(idfield=>"myid"));
+        assert!(index_writer.commit().is_ok());
+        assert_eq!(reader.searcher().num_docs(), 2u64);
+        assert_eq!(searcher.num_docs(), 1u64);
+    }
+
+    #[test]
+    fn test_index_writer_reader_soft_commit() {
+        let mut schema_builder = schema::Schema::builder();
+        let idfield = schema_builder.add_text_field("id", STRING);
+        schema_builder.add_text_field("optfield", STRING);
+        let index = Index::create_in_ram(schema_builder.build());
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        index_writer.add_document(doc!(idfield=>"myid"));
+        assert!(index_writer.soft_commit().is_ok());
+        let nrt_reader = index_writer.reader(2).unwrap();
+        let normal_reader = index.reader_builder().try_into().unwrap();
+        assert_eq!(nrt_reader.searcher().num_docs(), 1u64);
+        assert_eq!(normal_reader.searcher().num_docs(), 0u64);
+        assert!(index_writer.commit().is_ok());
+        assert!(normal_reader.reload().is_ok());
+        assert_eq!(nrt_reader.searcher().num_docs(), 1u64);
+        assert_eq!(normal_reader.searcher().num_docs(), 1u64);
+    }
 }

src/indexer/index_writer_config.rs

@@ -0,0 +1,133 @@
+use serde::{Deserialize, Serialize};
+
+// Size of the margin for the heap. A segment is closed when the remaining memory
+// in the heap goes below MARGIN_IN_BYTES.
+const MARGIN_IN_BYTES: u64 = 1_000_000;
+
+// We impose the memory per thread to be at least 3 MB.
+const HEAP_SIZE_MIN: u64 = MARGIN_IN_BYTES * 3u64;
+const HEAP_SIZE_MAX: u64 = u32::max_value() as u64 - MARGIN_IN_BYTES;
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct IndexWriterConfig {
+    pub max_indexing_threads: usize,
+    pub max_merging_threads: usize,
+    pub memory_budget: u64,
+    pub store_flush_num_bytes: u64,
+    pub persist_low: u64,
+    pub persist_high: u64,
+}
+
+impl Default for IndexWriterConfig {
+    fn default() -> Self {
+        IndexWriterConfig {
+            max_indexing_threads: 1,
+            max_merging_threads: 3,
+            memory_budget: 50_000_000u64,
+            store_flush_num_bytes: 10_000_000u64,
+            persist_low: 10_000_000u64,
+            persist_high: 50_000_000u64
+        }
+    }
+}
+
+impl IndexWriterConfig {
+    #[cfg(test)]
+    pub fn for_test() -> IndexWriterConfig {
+        IndexWriterConfig {
+            max_indexing_threads: 1,
+            max_merging_threads: 5,
+            memory_budget: 4_000_000u64,
+            store_flush_num_bytes: 500_000u64,
+            persist_low: 2_000_000u64,
+            persist_high: 3_000_000u64,
+        }
+    }
+
+    // Ensures the `IndexWriterConfig` is correct.
+    //
+    // This method checks that the values in the `IndexWriterConfig`
+    // are valid. If it is not, it may mutate some of the values (like `max_num_threads`) to
+    // fit the contracts or return an error with an explicit error message.
+    //
+    // If called twice, the config is guaranteed to not be updated the second time.
+    pub fn validate(&mut self) -> crate::Result<()> {
+        if self.memory_budget < HEAP_SIZE_MIN {
+            let err_msg = format!(
+                "The heap size per thread needs to be at least {}.",
+                HEAP_SIZE_MIN
+            );
+            return Err(crate::TantivyError::InvalidArgument(err_msg));
+        }
+        let heap_size_in_bytes_per_thread = self.heap_size_in_byte_per_thread();
+        if heap_size_in_bytes_per_thread >= HEAP_SIZE_MAX {
+            let err_msg = format!("The heap size per thread cannot exceed {}", HEAP_SIZE_MAX);
+            return Err(crate::TantivyError::InvalidArgument(err_msg));
+        }
+        if heap_size_in_bytes_per_thread < HEAP_SIZE_MIN {
+            self.max_indexing_threads = (self.memory_budget / HEAP_SIZE_MIN) as usize;
+        }
+        Ok(())
+    }
+
+    pub fn heap_size_in_byte_per_thread(&self) -> u64 {
+        self.memory_budget / self.max_indexing_threads as u64
+    }
+
+    pub fn heap_size_before_flushing(&self) -> u64 {
+        self.heap_size_in_byte_per_thread() - MARGIN_IN_BYTES
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use crate::IndexWriterConfig;
+
+    #[test]
+    fn test_index_writer_config_simple() {
+        let mut index = IndexWriterConfig {
+            max_indexing_threads: 3,
+            memory_budget: super::HEAP_SIZE_MIN * 3,
+            ..Default::default()
+        };
+        assert!(index.validate().is_ok());
+        assert_eq!(index.max_indexing_threads, 3);
+        assert_eq!(index.heap_size_in_byte_per_thread(), super::HEAP_SIZE_MIN);
+    }
+
+    #[test]
+    fn test_index_writer_config_reduce_num_threads() {
+        let mut index = IndexWriterConfig {
+            max_indexing_threads: 3,
+            memory_budget: super::HEAP_SIZE_MIN,
+            ..Default::default()
+        };
+        assert!(index.validate().is_ok());
+        assert_eq!(index.max_indexing_threads, 1);
+        assert_eq!(index.heap_size_in_byte_per_thread(), super::HEAP_SIZE_MIN);
+    }
+
+    #[test]
+    fn test_index_writer_config_not_enough_memory() {
+        let mut index = IndexWriterConfig {
+            max_indexing_threads: 1,
+            memory_budget: super::HEAP_SIZE_MIN - 1,
+            ..Default::default()
+        };
+        assert!(
+            matches!(index.validate(), Err(crate::TantivyError::InvalidArgument(msg) ) if msg.contains("The heap size per thread needs to be at least"))
+        );
+    }
+
+    #[test]
+    fn test_index_writer_config_too_much_memory() {
+        let mut index = IndexWriterConfig {
+            max_indexing_threads: 1,
+            memory_budget: (u32::max_value() as u64) + 1,
+            ..Default::default()
+        };
+        assert!(
+            matches!(index.validate(), Err(crate::TantivyError::InvalidArgument(msg) ) if msg.contains("The heap size per thread cannot exceed"))
+        );
+    }
+}

src/indexer/merge_operation.rs

@@ -1,17 +1,22 @@
+use crate::indexer::resource_manager::{Allocation, ResourceManager};
 use crate::Opstamp;
 use crate::SegmentId;
 use census::{Inventory, TrackedObject};
 use std::collections::HashSet;
+use std::fmt;
 use std::ops::Deref;
 
-#[derive(Default)]
-pub(crate) struct MergeOperationInventory(Inventory<InnerMergeOperation>);
+#[derive(Default, Clone)]
+pub(crate) struct MergeOperationInventory {
+    inventory: Inventory<InnerMergeOperation>,
+    num_merge_watcher: ResourceManager,
+}
 
 impl Deref for MergeOperationInventory {
     type Target = Inventory<InnerMergeOperation>;
 
     fn deref(&self) -> &Self::Target {
-        &self.0
+        &self.inventory
     }
 }
 
@@ -25,6 +30,10 @@ impl MergeOperationInventory {
         }
         segment_in_merge
     }
+
+    pub fn wait_until_empty(&self) {
+        let _ = self.num_merge_watcher.wait_until_in_range(0..1);
+    }
 }
 
 /// A `MergeOperation` has two roles.
@@ -47,6 +56,17 @@ pub struct MergeOperation {
 pub(crate) struct InnerMergeOperation {
     target_opstamp: Opstamp,
     segment_ids: Vec<SegmentId>,
+    _allocation: Allocation,
+}
+
+impl fmt::Debug for InnerMergeOperation {
+    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
+        write!(
+            f,
+            "MergeOp(target_opstamp={:?}, segment_ids={:?})",
+            self.target_opstamp, self.segment_ids
+        )
+    }
 }
 
 impl MergeOperation {
@@ -55,9 +75,11 @@ impl MergeOperation {
         target_opstamp: Opstamp,
         segment_ids: Vec<SegmentId>,
     ) -> MergeOperation {
+        let allocation = inventory.num_merge_watcher.allocate(1);
         let inner_merge_operation = InnerMergeOperation {
             target_opstamp,
             segment_ids,
+            _allocation: allocation,
         };
         MergeOperation {
             inner: inventory.track(inner_merge_operation),

src/indexer/merger.rs

@@ -2,6 +2,8 @@ use crate::common::MAX_DOC_LIMIT;
 use crate::core::Segment;
 use crate::core::SegmentReader;
 use crate::core::SerializableSegment;
+use crate::directory::TerminatingWrite;
+use crate::directory::WritePtr;
 use crate::docset::DocSet;
 use crate::fastfield::BytesFastFieldReader;
 use crate::fastfield::DeleteBitSet;
@@ -660,7 +662,8 @@ impl IndexMerger {
         Ok(term_ordinal_mappings)
     }
 
-    fn write_storable_fields(&self, store_writer: &mut StoreWriter) -> crate::Result<()> {
+    pub fn write_storable_fields(&self, store_wrt: WritePtr) -> crate::Result<()> {
+        let mut store_writer = StoreWriter::new(store_wrt);
         for reader in &self.readers {
             let store_reader = reader.get_store_reader();
             if reader.num_deleted_docs() > 0 {
@@ -672,6 +675,8 @@ impl IndexMerger {
                 store_writer.stack(&store_reader)?;
             }
         }
+        let store_wrt = store_writer.close()?;
+        store_wrt.terminate()?;
         Ok(())
     }
 }
@@ -681,7 +686,6 @@ impl SerializableSegment for IndexMerger {
         let term_ord_mappings = self.write_postings(serializer.get_postings_serializer())?;
         self.write_fieldnorms(serializer.get_fieldnorms_serializer())?;
         self.write_fast_fields(serializer.get_fast_field_serializer(), term_ord_mappings)?;
-        self.write_storable_fields(serializer.get_store_writer())?;
         serializer.close()?;
         Ok(self.max_doc)
     }

src/indexer/mod.rs

@@ -2,12 +2,14 @@ pub mod delete_queue;
 
 mod doc_opstamp_mapping;
 pub mod index_writer;
+mod index_writer_config;
 mod log_merge_policy;
 mod merge_operation;
 pub mod merge_policy;
 pub mod merger;
 pub mod operation;
 mod prepared_commit;
+mod resource_manager;
 mod segment_entry;
 mod segment_manager;
 mod segment_register;
@@ -16,13 +18,17 @@ pub mod segment_updater;
 mod segment_writer;
 mod stamper;
 
+pub(crate) use self::resource_manager::{Allocation, ResourceManager};
+pub(crate) use self::merge_operation::MergeOperationInventory;
 pub use self::index_writer::IndexWriter;
+pub use self::index_writer_config::IndexWriterConfig;
 pub use self::log_merge_policy::LogMergePolicy;
 pub use self::merge_operation::MergeOperation;
 pub use self::merge_policy::{MergeCandidate, MergePolicy, NoMergePolicy};
 pub use self::prepared_commit::PreparedCommit;
 pub use self::segment_entry::SegmentEntry;
 pub use self::segment_manager::SegmentManager;
+pub(crate) use self::segment_manager::SegmentRegisters;
 pub use self::segment_serializer::SegmentSerializer;
 pub use self::segment_writer::SegmentWriter;
 

src/indexer/prepared_commit.rs

@@ -7,14 +7,20 @@ pub struct PreparedCommit<'a> {
     index_writer: &'a mut IndexWriter,
     payload: Option<String>,
     opstamp: Opstamp,
+    soft_commit: bool,
 }
 
 impl<'a> PreparedCommit<'a> {
-    pub(crate) fn new(index_writer: &'a mut IndexWriter, opstamp: Opstamp) -> PreparedCommit<'_> {
+    pub(crate) fn new(
+        index_writer: &'a mut IndexWriter,
+        opstamp: Opstamp,
+        soft_commit: bool,
+    ) -> PreparedCommit<'_> {
         PreparedCommit {
             index_writer,
             payload: None,
             opstamp,
+            soft_commit,
         }
     }
 
@@ -32,11 +38,12 @@ impl<'a> PreparedCommit<'a> {
 
     pub fn commit(self) -> crate::Result<Opstamp> {
         info!("committing {}", self.opstamp);
-        let _ = block_on(
-            self.index_writer
-                .segment_updater()
-                .schedule_commit(self.opstamp, self.payload),
-        );
+        block_on(self.index_writer.segment_updater().schedule_commit(
+            self.opstamp,
+            self.payload,
+            self.soft_commit,
+        ))?;
+        block_on(self.index_writer.trigger_commit());
         Ok(self.opstamp)
     }
 }

src/indexer/resource_manager.rs

@@ -0,0 +1,213 @@
+use std::ops::RangeBounds;
+use std::sync::{Arc, Condvar, Mutex, MutexGuard, RwLock};
+
+struct LockedData {
+    count: u64,
+    enabled: bool
+}
+
+impl Default for LockedData {
+    fn default() -> Self {
+        LockedData {
+            count: 0u64,
+            enabled: true
+        }
+    }
+}
+
+#[derive(Default)]
+struct Inner {
+    resource_level: Mutex<LockedData>,
+    convdvar: Condvar,
+}
+
+
+/// The resource manager makes it possible to track the amount of level of a given resource.
+/// There is no magic here : it is to the description of the user to declare how much
+/// of the resource is being held.
+///
+/// Allocation of a resource is bound to the lifetime of a `Allocation` instance.
+///
+/// ```rust
+/// let resource_manager = ResourceManager::default();
+///
+/// ```
+///
+/// In tantivy, this is used to check the number of merging thread and the number of memory
+/// used by the volatile segments.
+///
+#[derive(Clone, Default)]
+pub struct ResourceManager {
+    inner: Arc<Inner>,
+}
+
+impl ResourceManager {
+    /// Return the total amount of reousrce allocated
+    pub fn total_amount(&self) -> u64 {
+        self.lock().count
+    }
+
+    fn lock(&self) -> MutexGuard<LockedData> {
+        self.inner
+            .resource_level
+            .lock()
+            .expect("Failed to obtain lock for ReservedMemory. This should never happen.")
+    }
+
+    fn record_delta(&self, delta: i64) {
+        if delta == 0i64 {
+            return;
+        }
+        let mut lock = self.lock();
+        let new_val = lock.count as i64 + delta;
+        lock.count = new_val as u64;
+        self.inner.convdvar.notify_all();
+    }
+
+    /// Records a new allocation.
+    ///
+    /// The returned allocate object is used to automatically release the allocated resource
+    /// on drop.
+    pub fn allocate(&self, amount: u64) -> Allocation {
+        self.record_delta(amount as i64);
+        Allocation {
+            resource_manager: self.clone(),
+            amount: RwLock::new(amount),
+        }
+    }
+
+    /// Stops the resource manager.
+    ///
+    /// If any thread is waiting via `.wait_until_in_range(...)`, the method will stop
+    /// being blocking and will return an error.
+    pub fn terminate(&self) {
+        self.lock().enabled = false;
+        self.inner.convdvar.notify_all();
+    }
+
+    /// Blocks the current thread until the resource level reaches the given range,
+    /// in a cpu-efficient way.
+    ///
+    /// This method does not necessarily wakes up the current thread at every transition
+    /// into the targetted range, but any durable entry in the range will be detected.
+    pub fn wait_until_in_range<R: RangeBounds<u64>>(&self, range: R) -> Result<u64, u64> {
+        let mut levels = self.lock();
+        if !levels.enabled {
+            return Err(levels.count)
+        }
+        while !range.contains(&levels.count) {
+           levels = self.inner.convdvar.wait(levels).unwrap();
+            if !levels.enabled {
+                return Err(levels.count)
+            }
+        }
+        Ok(levels.count)
+    }
+}
+
+pub struct Allocation {
+    resource_manager: ResourceManager,
+    amount: RwLock<u64>,
+}
+
+impl Allocation {
+    pub fn amount(&self) -> u64 {
+        *self.amount.read().unwrap()
+    }
+
+    pub fn modify(&self, new_amount: u64) {
+        let mut wlock = self.amount.write().unwrap();
+        let delta = new_amount as i64 - *wlock as i64;
+        *wlock = new_amount;
+        self.resource_manager.record_delta(delta);
+    }
+}
+
+impl Drop for Allocation {
+    fn drop(&mut self) {
+        let amount = self.amount();
+        self.resource_manager.record_delta(-(amount as i64))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::ResourceManager;
+    use futures::channel::oneshot;
+    use futures::executor::block_on;
+    use std::{mem, thread};
+
+    #[test]
+    fn test_simple_allocation() {
+        let memory = ResourceManager::default();
+        assert_eq!(memory.total_amount(), 0u64);
+        let _allocation = memory.allocate(10u64);
+        assert_eq!(memory.total_amount(), 10u64);
+    }
+
+    #[test]
+    fn test_multiple_allocation() {
+        let memory = ResourceManager::default();
+        assert_eq!(memory.total_amount(), 0u64);
+        let _allocation = memory.allocate(10u64);
+        let _allocation_2 = memory.allocate(11u64);
+        assert_eq!(memory.total_amount(), 21u64);
+    }
+
+    #[test]
+    fn test_release_on_drop() {
+        let memory = ResourceManager::default();
+        assert_eq!(memory.total_amount(), 0u64);
+        let allocation = memory.allocate(10u64);
+        let allocation_2 = memory.allocate(11u64);
+        assert_eq!(memory.total_amount(), 21u64);
+        mem::drop(allocation);
+        assert_eq!(memory.total_amount(), 11u64);
+        mem::drop(allocation_2);
+        assert_eq!(memory.total_amount(), 0u64);
+    }
+
+    #[test]
+    fn test_wait_until() {
+        let memory = ResourceManager::default();
+        let (send, recv) = oneshot::channel::<()>();
+        let memory_clone = memory.clone();
+        thread::spawn(move || {
+            let _allocation1 = memory_clone.allocate(2u64);
+            let _allocation2 = memory_clone.allocate(3u64);
+            let _allocation3 = memory_clone.allocate(4u64);
+            std::mem::drop(_allocation3);
+            assert!(block_on(recv).is_ok());
+        });
+        assert_eq!(memory.wait_until_in_range(5u64..8u64), Ok(5u64));
+        assert!(send.send(()).is_ok());
+    }
+
+    #[test]
+    fn test_modify_amount() {
+        let memory = ResourceManager::default();
+        let alloc = memory.allocate(2u64);
+        assert_eq!(memory.total_amount(), 2u64);
+        assert_eq!(alloc.amount(), 2u64);
+        let alloc2 = memory.allocate(3u64);
+        assert_eq!(memory.total_amount(), 2u64 + 3u64);
+        assert_eq!(alloc2.amount(), 3u64);
+        alloc.modify(14u64);
+        assert_eq!(alloc.amount(), 14u64);
+        assert_eq!(memory.total_amount(), 14u64 + 3u64)
+    }
+    
+    #[test]
+    fn test_stop_resource_manager() {
+        let resource_manager = ResourceManager::default();
+        let resource_manager_clone = resource_manager.clone();
+        let (sender, recv) = oneshot::channel();
+        let join_handle = thread::spawn(move || {
+            assert!(sender.send(()).is_ok());
+            resource_manager_clone.wait_until_in_range(10..20)
+        });
+        let _ = block_on(recv);
+        resource_manager.terminate();
+        assert_eq!(join_handle.join().unwrap(), Err(0u64));
+    }
+}

src/indexer/segment_entry.rs

@@ -1,7 +1,9 @@
 use crate::common::BitSet;
 use crate::core::SegmentId;
 use crate::core::SegmentMeta;
+use crate::directory::ManagedDirectory;
 use crate::indexer::delete_queue::DeleteCursor;
+use crate::{Opstamp, Segment};
 use std::fmt;
 
 /// A segment entry describes the state of
@@ -19,7 +21,7 @@ use std::fmt;
 /// in the .del file or in the `delete_bitset`.
 #[derive(Clone)]
 pub struct SegmentEntry {
-    meta: SegmentMeta,
+    segment: Segment,
     delete_bitset: Option<BitSet>,
     delete_cursor: DeleteCursor,
 }
@@ -27,47 +29,67 @@ pub struct SegmentEntry {
 impl SegmentEntry {
     /// Create a new `SegmentEntry`
     pub fn new(
-        segment_meta: SegmentMeta,
+        segment: Segment,
         delete_cursor: DeleteCursor,
         delete_bitset: Option<BitSet>,
     ) -> SegmentEntry {
         SegmentEntry {
-            meta: segment_meta,
+            segment,
             delete_bitset,
             delete_cursor,
         }
     }
 
-    /// Return a reference to the segment entry deleted bitset.
-    ///
-    /// `DocId` in this bitset are flagged as deleted.
-    pub fn delete_bitset(&self) -> Option<&BitSet> {
-        self.delete_bitset.as_ref()
+    pub fn persist(&mut self, dest_directory: ManagedDirectory) -> crate::Result<()> {
+        // TODO take in account delete bitset?
+        self.segment.persist(dest_directory)?;
+        Ok(())
     }
 
-    /// Set the `SegmentMeta` for this segment.
-    pub fn set_meta(&mut self, segment_meta: SegmentMeta) {
-        self.meta = segment_meta;
+    pub fn set_delete_cursor(&mut self, delete_cursor: DeleteCursor) {
+        self.delete_cursor = delete_cursor;
+    }
+
+    /// `Takes` (as in Option::take) the delete bitset of a segment entry.
+    /// `DocId` in this bitset are flagged as deleted.
+    pub fn take_delete_bitset(&mut self) -> Option<BitSet> {
+        self.delete_bitset.take()
+    }
+
+    /// Reset the delete information in this segment.
+    ///
+    /// The `SegmentEntry` segment's `SegmentMeta` gets updated, and
+    /// any delete bitset is drop and set to None.
+    pub fn reset_delete_meta(&mut self, num_deleted_docs: u32, target_opstamp: Opstamp) {
+        self.segment = self
+            .segment
+            .clone()
+            .with_delete_meta(num_deleted_docs, target_opstamp);
+        self.delete_bitset = None;
     }
 
     /// Return a reference to the segment_entry's delete cursor
-    pub fn delete_cursor(&mut self) -> &mut DeleteCursor {
-        &mut self.delete_cursor
+    pub fn delete_cursor(&mut self) -> DeleteCursor {
+        self.delete_cursor.clone()
     }
 
     /// Returns the segment id.
     pub fn segment_id(&self) -> SegmentId {
-        self.meta.id()
+        self.meta().id()
     }
 
+    /// Returns the `segment` associated to the `SegmentEntry`.
+    pub fn segment(&self) -> &Segment {
+        &self.segment
+    }
     /// Accessor to the `SegmentMeta`
     pub fn meta(&self) -> &SegmentMeta {
-        &self.meta
+        self.segment.meta()
     }
 }
 
 impl fmt::Debug for SegmentEntry {
     fn fmt(&self, formatter: &mut fmt::Formatter<'_>) -> fmt::Result {
-        write!(formatter, "SegmentEntry({:?})", self.meta)
+        write!(formatter, "SegmentEntry({:?})", self.meta())
     }
 }

src/indexer/segment_manager.rs

@@ -2,15 +2,15 @@ use super::segment_register::SegmentRegister;
 use crate::core::SegmentId;
 use crate::core::SegmentMeta;
 use crate::error::TantivyError;
-use crate::indexer::delete_queue::DeleteCursor;
-use crate::indexer::SegmentEntry;
+use crate::indexer::{SegmentEntry, MergeOperationInventory, MergeCandidate, MergeOperation};
+use crate::{Segment, Opstamp};
 use std::collections::hash_set::HashSet;
 use std::fmt::{self, Debug, Formatter};
-use std::sync::RwLock;
+use std::sync::{Arc, RwLock};
 use std::sync::{RwLockReadGuard, RwLockWriteGuard};
 
 #[derive(Default)]
-struct SegmentRegisters {
+pub(crate) struct SegmentRegisters {
     uncommitted: SegmentRegister,
     committed: SegmentRegister,
 }
@@ -22,6 +22,17 @@ pub(crate) enum SegmentsStatus {
 }
 
 impl SegmentRegisters {
+    pub fn new(committed: SegmentRegister) -> SegmentRegisters {
+        SegmentRegisters {
+            uncommitted: Default::default(),
+            committed,
+        }
+    }
+
+    pub fn committed_segment(&self) -> Vec<Segment> {
+        self.committed.segments()
+    }
+
     /// Check if all the segments are committed or uncommited.
     ///
     /// If some segment is missing or segments are in a different state (this should not happen
@@ -44,7 +55,8 @@ impl SegmentRegisters {
 /// changes (merges especially)
 #[derive(Default)]
 pub struct SegmentManager {
-    registers: RwLock<SegmentRegisters>,
+    registers: Arc<RwLock<SegmentRegisters>>,
+    merge_operations: MergeOperationInventory,
 }
 
 impl Debug for SegmentManager {
@@ -58,34 +70,28 @@ impl Debug for SegmentManager {
     }
 }
 
-pub fn get_mergeable_segments(
-    in_merge_segment_ids: &HashSet<SegmentId>,
-    segment_manager: &SegmentManager,
-) -> (Vec<SegmentMeta>, Vec<SegmentMeta>) {
-    let registers_lock = segment_manager.read();
-    (
-        registers_lock
-            .committed
-            .get_mergeable_segments(in_merge_segment_ids),
-        registers_lock
-            .uncommitted
-            .get_mergeable_segments(in_merge_segment_ids),
-    )
-}
 
 impl SegmentManager {
-    pub fn from_segments(
-        segment_metas: Vec<SegmentMeta>,
-        delete_cursor: &DeleteCursor,
-    ) -> SegmentManager {
+    pub(crate) fn new(registers: Arc<RwLock<SegmentRegisters>>) -> SegmentManager {
         SegmentManager {
-            registers: RwLock::new(SegmentRegisters {
-                uncommitted: SegmentRegister::default(),
-                committed: SegmentRegister::new(segment_metas, delete_cursor),
-            }),
+            registers,
+            merge_operations: Default::default()
         }
     }
 
+    pub fn new_merge_operation(&self, opstamp: Opstamp, merge_candidate: MergeCandidate) -> MergeOperation {
+        MergeOperation::new(
+            &self.merge_operations,
+            opstamp,
+            merge_candidate.0
+
+        )
+    }
+
+    pub fn wait_merging_thread(&self) {
+        self.merge_operations.wait_until_empty()
+    }
+
     /// Returns all of the segment entries (committed or uncommitted)
     pub fn segment_entries(&self) -> Vec<SegmentEntry> {
         let registers_lock = self.read();
@@ -94,6 +100,34 @@ impl SegmentManager {
         segment_entries
     }
 
+    /// Returns the segments that are currently not in merge.
+    ///
+    /// They are split over two `Vec`. The committed segments on one hand
+    /// and the uncommitted ones on the other hand.
+    ///
+    /// This method is useful when searching for merge candidate or segments
+    /// to persists.
+    pub fn segments_not_in_merge(&self) -> (Vec<SegmentMeta>, Vec<SegmentMeta>) {
+        let in_merge_segment_ids: HashSet<SegmentId> = self.merge_operations.segment_in_merge();
+        let registers_lock = self.read();
+        (
+            registers_lock
+                .committed
+                .get_mergeable_segments(&in_merge_segment_ids),
+            registers_lock
+                .uncommitted
+                .get_mergeable_segments(&in_merge_segment_ids),
+        )
+    }
+
+    pub fn largest_segment_not_in_merge(&self) -> Option<SegmentMeta> {
+        let (committed, uncommitted) = self.segments_not_in_merge();
+        let mut segments = vec![];
+        segments.extend(committed);
+        segments.extend(uncommitted);
+        None
+    }
+
     // Lock poisoning should never happen :
     // The lock is acquired and released within this class,
     // and the operations cannot panic.

src/indexer/segment_register.rs

@@ -1,7 +1,10 @@
 use crate::core::SegmentId;
 use crate::core::SegmentMeta;
+use crate::directory::ManagedDirectory;
 use crate::indexer::delete_queue::DeleteCursor;
 use crate::indexer::segment_entry::SegmentEntry;
+use crate::schema::Schema;
+use crate::Segment;
 use std::collections::HashMap;
 use std::collections::HashSet;
 use std::fmt::{self, Debug, Formatter};
@@ -46,6 +49,13 @@ impl SegmentRegister {
             .collect()
     }
 
+    pub fn segments(&self) -> Vec<Segment> {
+        self.segment_states
+            .values()
+            .map(|segment_entry| segment_entry.segment().clone())
+            .collect()
+    }
+
     pub fn segment_entries(&self) -> Vec<SegmentEntry> {
         self.segment_states.values().cloned().collect()
     }
@@ -79,11 +89,17 @@ impl SegmentRegister {
         self.segment_states.get(segment_id).cloned()
     }
 
-    pub fn new(segment_metas: Vec<SegmentMeta>, delete_cursor: &DeleteCursor) -> SegmentRegister {
+    pub fn new(
+        directory: &ManagedDirectory,
+        schema: &Schema,
+        segment_metas: Vec<SegmentMeta>,
+        delete_cursor: &DeleteCursor,
+    ) -> SegmentRegister {
         let mut segment_states = HashMap::new();
         for segment_meta in segment_metas {
             let segment_id = segment_meta.id();
-            let segment_entry = SegmentEntry::new(segment_meta, delete_cursor.clone(), None);
+            let segment = Segment::new_persisted(segment_meta, directory.clone(), schema.clone());
+            let segment_entry = SegmentEntry::new(segment, delete_cursor.clone(), None);
             segment_states.insert(segment_id, segment_entry);
         }
         SegmentRegister { segment_states }
@@ -95,6 +111,7 @@ mod tests {
     use super::*;
     use crate::core::{SegmentId, SegmentMetaInventory};
     use crate::indexer::delete_queue::*;
+    use crate::indexer::ResourceManager;
 
     fn segment_ids(segment_register: &SegmentRegister) -> Vec<SegmentId> {
         segment_register
@@ -108,28 +125,34 @@ mod tests {
     fn test_segment_register() {
         let inventory = SegmentMetaInventory::default();
         let delete_queue = DeleteQueue::new();
+        let schema = Schema::builder().build();
 
         let mut segment_register = SegmentRegister::default();
         let segment_id_a = SegmentId::generate_random();
         let segment_id_b = SegmentId::generate_random();
         let segment_id_merged = SegmentId::generate_random();
 
+        let memory_manager = ResourceManager::default();
+
         {
-            let segment_meta = inventory.new_segment_meta(segment_id_a, 0u32);
-            let segment_entry = SegmentEntry::new(segment_meta, delete_queue.cursor(), None);
+            let meta = inventory.new_segment_meta(segment_id_a, 0u32);
+            let segment = Segment::new_volatile(meta, schema.clone(), memory_manager.clone());
+            let segment_entry = SegmentEntry::new(segment, delete_queue.cursor(), None);
             segment_register.add_segment_entry(segment_entry);
         }
         assert_eq!(segment_ids(&segment_register), vec![segment_id_a]);
         {
-            let segment_meta = inventory.new_segment_meta(segment_id_b, 0u32);
-            let segment_entry = SegmentEntry::new(segment_meta, delete_queue.cursor(), None);
+            let meta = inventory.new_segment_meta(segment_id_b, 0u32);
+            let segment = Segment::new_volatile(meta, schema.clone(), memory_manager.clone());
+            let segment_entry = SegmentEntry::new(segment, delete_queue.cursor(), None);
             segment_register.add_segment_entry(segment_entry);
         }
         segment_register.remove_segment(&segment_id_a);
         segment_register.remove_segment(&segment_id_b);
         {
             let segment_meta_merged = inventory.new_segment_meta(segment_id_merged, 0u32);
-            let segment_entry = SegmentEntry::new(segment_meta_merged, delete_queue.cursor(), None);
+            let segment_merged = Segment::new_volatile(segment_meta_merged, schema.clone(), memory_manager.clone());
+            let segment_entry = SegmentEntry::new(segment_merged, delete_queue.cursor(), None);
             segment_register.add_segment_entry(segment_entry);
         }
         assert_eq!(segment_ids(&segment_register), vec![segment_id_merged]);

src/indexer/segment_serializer.rs

@@ -3,12 +3,10 @@ use crate::core::SegmentComponent;
 use crate::fastfield::FastFieldSerializer;
 use crate::fieldnorm::FieldNormsSerializer;
 use crate::postings::InvertedIndexSerializer;
-use crate::store::StoreWriter;
 
 /// Segment serializer is in charge of laying out on disk
 /// the data accumulated and sorted by the `SegmentWriter`.
 pub struct SegmentSerializer {
-    store_writer: StoreWriter,
     fast_field_serializer: FastFieldSerializer,
     fieldnorms_serializer: FieldNormsSerializer,
     postings_serializer: InvertedIndexSerializer,
@@ -17,8 +15,6 @@ pub struct SegmentSerializer {
 impl SegmentSerializer {
     /// Creates a new `SegmentSerializer`.
     pub fn for_segment(segment: &mut Segment) -> crate::Result<SegmentSerializer> {
-        let store_write = segment.open_write(SegmentComponent::STORE)?;
-
         let fast_field_write = segment.open_write(SegmentComponent::FASTFIELDS)?;
         let fast_field_serializer = FastFieldSerializer::from_write(fast_field_write)?;
 
@@ -27,7 +23,6 @@ impl SegmentSerializer {
 
         let postings_serializer = InvertedIndexSerializer::open(segment)?;
         Ok(SegmentSerializer {
-            store_writer: StoreWriter::new(store_write),
             fast_field_serializer,
             fieldnorms_serializer,
             postings_serializer,
@@ -49,16 +44,10 @@ impl SegmentSerializer {
         &mut self.fieldnorms_serializer
     }
 
-    /// Accessor to the `StoreWriter`.
-    pub fn get_store_writer(&mut self) -> &mut StoreWriter {
-        &mut self.store_writer
-    }
-
     /// Finalize the segment serialization.
     pub fn close(self) -> crate::Result<()> {
         self.fast_field_serializer.close()?;
         self.postings_serializer.close()?;
-        self.store_writer.close()?;
         self.fieldnorms_serializer.close()?;
         Ok(())
     }

src/indexer/segment_updater.rs

@@ -1,4 +1,4 @@
-use super::segment_manager::{get_mergeable_segments, SegmentManager};
+use super::segment_manager::SegmentManager;
 use crate::core::Index;
 use crate::core::IndexMeta;
 use crate::core::Segment;
@@ -7,22 +7,21 @@ use crate::core::SegmentMeta;
 use crate::core::SerializableSegment;
 use crate::core::META_FILEPATH;
 use crate::directory::{Directory, DirectoryClone, GarbageCollectionResult};
-use crate::indexer::delete_queue::DeleteCursor;
 use crate::indexer::index_writer::advance_deletes;
-use crate::indexer::merge_operation::MergeOperationInventory;
 use crate::indexer::merger::IndexMerger;
-use crate::indexer::segment_manager::SegmentsStatus;
+use crate::indexer::segment_manager::{SegmentRegisters, SegmentsStatus};
 use crate::indexer::stamper::Stamper;
 use crate::indexer::SegmentEntry;
 use crate::indexer::SegmentSerializer;
 use crate::indexer::{DefaultMergePolicy, MergePolicy};
 use crate::indexer::{MergeCandidate, MergeOperation};
 use crate::schema::Schema;
-use crate::Opstamp;
+use crate::{Opstamp, SegmentComponent};
 use futures::channel::oneshot;
 use futures::executor::{ThreadPool, ThreadPoolBuilder};
 use futures::future::Future;
 use futures::future::TryFutureExt;
+use serde_json;
 use std::borrow::BorrowMut;
 use std::collections::HashSet;
 use std::io::Write;
@@ -116,15 +115,14 @@ fn merge(
 
     // First we apply all of the delet to the merged segment, up to the target opstamp.
     for segment_entry in &mut segment_entries {
-        let segment = index.segment(segment_entry.meta().clone());
-        advance_deletes(segment, segment_entry, target_opstamp)?;
+        advance_deletes(segment_entry, target_opstamp)?;
     }
 
-    let delete_cursor = segment_entries[0].delete_cursor().clone();
+    let delete_cursor = segment_entries[0].delete_cursor();
 
     let segments: Vec<Segment> = segment_entries
         .iter()
-        .map(|segment_entry| index.segment(segment_entry.meta().clone()))
+        .map(|segment_entry| segment_entry.segment().clone())
         .collect();
 
     // An IndexMerger is like a "view" of our merged segments.
@@ -133,11 +131,16 @@ fn merge(
     // ... we just serialize this index merger in our new segment to merge the two segments.
     let segment_serializer = SegmentSerializer::for_segment(&mut merged_segment)?;
 
-    let num_docs = merger.write(segment_serializer)?;
+    let store_wrt = merged_segment.open_write(SegmentComponent::STORE)?;
+    merger.write_storable_fields(store_wrt)?;
 
-    let segment_meta = index.new_segment_meta(merged_segment.id(), num_docs);
+    let max_doc = merger.write(segment_serializer)?;
 
-    Ok(SegmentEntry::new(segment_meta, delete_cursor, None))
+    Ok(SegmentEntry::new(
+        merged_segment.with_max_doc(max_doc),
+        delete_cursor,
+        None,
+    ))
 }
 
 pub(crate) struct InnerSegmentUpdater {
@@ -156,17 +159,15 @@ pub(crate) struct InnerSegmentUpdater {
     merge_policy: RwLock<Arc<Box<dyn MergePolicy>>>,
     killed: AtomicBool,
     stamper: Stamper,
-    merge_operations: MergeOperationInventory,
 }
 
 impl SegmentUpdater {
     pub fn create(
+        segment_registers: Arc<RwLock<SegmentRegisters>>,
         index: Index,
         stamper: Stamper,
-        delete_cursor: &DeleteCursor,
     ) -> crate::Result<SegmentUpdater> {
-        let segments = index.searchable_segment_metas()?;
-        let segment_manager = SegmentManager::from_segments(segments, delete_cursor);
+        let segment_manager = SegmentManager::new(segment_registers);
         let pool = ThreadPoolBuilder::new()
             .name_prefix("segment_updater")
             .pool_size(1)
@@ -195,7 +196,6 @@ impl SegmentUpdater {
             merge_policy: RwLock::new(Arc::new(Box::new(DefaultMergePolicy::default()))),
             killed: AtomicBool::new(false),
             stamper,
-            merge_operations: Default::default(),
         })))
     }
 
@@ -261,8 +261,7 @@ impl SegmentUpdater {
     fn purge_deletes(&self, target_opstamp: Opstamp) -> crate::Result<Vec<SegmentEntry>> {
         let mut segment_entries = self.segment_manager.segment_entries();
         for segment_entry in &mut segment_entries {
-            let segment = self.index.segment(segment_entry.meta().clone());
-            advance_deletes(segment, segment_entry, target_opstamp)?;
+            advance_deletes(segment_entry, target_opstamp)?;
         }
         Ok(segment_entries)
     }
@@ -330,12 +329,21 @@ impl SegmentUpdater {
         &self,
         opstamp: Opstamp,
         payload: Option<String>,
+        soft_commit: bool,
     ) -> impl Future<Output = crate::Result<()>> {
         let segment_updater: SegmentUpdater = self.clone();
+        let directory = self.index.directory().clone();
         self.schedule_future(async move {
-            let segment_entries = segment_updater.purge_deletes(opstamp)?;
+            let mut segment_entries = segment_updater.purge_deletes(opstamp)?;
+            if !soft_commit {
+                for segment_entry in &mut segment_entries {
+                    segment_entry.persist(directory.clone())?;
+                }
+            }
             segment_updater.segment_manager.commit(segment_entries);
-            segment_updater.save_metas(opstamp, payload)?;
+            if !soft_commit {
+                segment_updater.save_metas(opstamp, payload)?;
+            }
             let _ = garbage_collect_files(segment_updater.clone()).await;
             segment_updater.consider_merge_options().await;
             Ok(())
@@ -352,7 +360,7 @@ impl SegmentUpdater {
 
     pub(crate) fn make_merge_operation(&self, segment_ids: &[SegmentId]) -> MergeOperation {
         let commit_opstamp = self.load_metas().opstamp;
-        MergeOperation::new(&self.merge_operations, commit_opstamp, segment_ids.to_vec())
+        self.segment_manager.new_merge_operation(commit_opstamp, MergeCandidate(segment_ids.to_vec()))
     }
 
     // Starts a merge operation. This function will block until the merge operation is effectively
@@ -426,9 +434,8 @@ impl SegmentUpdater {
     }
 
     async fn consider_merge_options(&self) {
-        let merge_segment_ids: HashSet<SegmentId> = self.merge_operations.segment_in_merge();
         let (committed_segments, uncommitted_segments) =
-            get_mergeable_segments(&merge_segment_ids, &self.segment_manager);
+            self.segment_manager.segments_not_in_merge();
 
         // Committed segments cannot be merged with uncommitted_segments.
         // We therefore consider merges using these two sets of segments independently.
@@ -439,7 +446,7 @@ impl SegmentUpdater {
             .compute_merge_candidates(&uncommitted_segments)
             .into_iter()
             .map(|merge_candidate| {
-                MergeOperation::new(&self.merge_operations, current_opstamp, merge_candidate.0)
+                self.segment_manager.new_merge_operation(current_opstamp, merge_candidate)
             })
             .collect();
 
@@ -448,7 +455,7 @@ impl SegmentUpdater {
             .compute_merge_candidates(&committed_segments)
             .into_iter()
             .map(|merge_candidate: MergeCandidate| {
-                MergeOperation::new(&self.merge_operations, commit_opstamp, merge_candidate.0)
+                self.segment_manager.new_merge_operation(commit_opstamp, merge_candidate)
             })
             .collect::<Vec<_>>();
         merge_candidates.extend(committed_merge_candidates.into_iter());
@@ -473,17 +480,13 @@ impl SegmentUpdater {
         let end_merge_future = self.schedule_future(async move {
             info!("End merge {:?}", after_merge_segment_entry.meta());
             {
-                let mut delete_cursor = after_merge_segment_entry.delete_cursor().clone();
+                let mut delete_cursor = after_merge_segment_entry.delete_cursor();
                 if let Some(delete_operation) = delete_cursor.get() {
                     let committed_opstamp = segment_updater.load_metas().opstamp;
                     if delete_operation.opstamp < committed_opstamp {
-                        let index = &segment_updater.index;
-                        let segment = index.segment(after_merge_segment_entry.meta().clone());
-                        if let Err(e) = advance_deletes(
-                            segment,
-                            &mut after_merge_segment_entry,
-                            committed_opstamp,
-                        ) {
+                        if let Err(e) =
+                            advance_deletes(&mut after_merge_segment_entry, committed_opstamp)
+                        {
                             error!(
                                 "Merge of {:?} was cancelled (advancing deletes failed): {:?}",
                                 merge_operation.segment_ids(),
@@ -532,9 +535,8 @@ impl SegmentUpdater {
     ///
     /// Obsolete files will eventually be cleaned up
     /// by the directory garbage collector.
-    pub fn wait_merging_thread(&self) -> crate::Result<()> {
-        self.merge_operations.wait_until_empty();
-        Ok(())
+    pub fn wait_merging_thread(&self) {
+        self.segment_manager.wait_merging_thread()
     }
 }
 
@@ -543,7 +545,8 @@ mod tests {
 
     use crate::indexer::merge_policy::tests::MergeWheneverPossible;
     use crate::schema::*;
-    use crate::Index;
+    use crate::{Index, SegmentId};
+    use futures::executor::block_on;
 
     #[test]
     fn test_delete_during_merge() {
@@ -694,4 +697,27 @@ mod tests {
             .segment_entries();
         assert!(seg_vec.is_empty());
     }
+
+    #[test]
+    fn test_merge_over_soft_commit() {
+        let mut schema_builder = Schema::builder();
+        let text_field = schema_builder.add_text_field("text", TEXT);
+        let schema = schema_builder.build();
+        let index = Index::create_in_ram(schema);
+        // writing the segment
+        let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
+        index_writer.add_document(doc!(text_field=>"a"));
+        assert!(index_writer.soft_commit().is_ok());
+        index_writer.add_document(doc!(text_field=>"a"));
+        assert!(index_writer.soft_commit().is_ok());
+
+        let reader = index_writer.reader(1).unwrap();
+        let segment_ids: Vec<SegmentId> = reader
+            .searcher()
+            .segment_readers()
+            .iter()
+            .map(|reader| reader.segment_id())
+            .collect();
+        assert!(block_on(index_writer.merge(&segment_ids)).is_ok());
+    }
 }

src/indexer/segment_writer.rs

@@ -1,9 +1,11 @@
 use super::operation::AddOperation;
 use crate::core::Segment;
 use crate::core::SerializableSegment;
+use crate::directory::{SpillingResult, SpillingWriter, TerminatingWrite};
 use crate::fastfield::FastFieldsWriter;
 use crate::fieldnorm::FieldNormsWriter;
 use crate::indexer::segment_serializer::SegmentSerializer;
+use crate::indexer::IndexWriterConfig;
 use crate::postings::compute_table_size;
 use crate::postings::MultiFieldPostingsWriter;
 use crate::schema::FieldType;
@@ -11,21 +13,24 @@ use crate::schema::Schema;
 use crate::schema::Term;
 use crate::schema::Value;
 use crate::schema::{Field, FieldEntry};
+use crate::store::StoreWriter;
 use crate::tokenizer::{BoxTokenStream, PreTokenizedStream};
-use crate::tokenizer::{FacetTokenizer, TextAnalyzer};
+use crate::tokenizer::{FacetTokenizer, TextAnalyzer, TokenizerManager};
 use crate::tokenizer::{TokenStreamChain, Tokenizer};
-use crate::DocId;
 use crate::Opstamp;
+use crate::{DocId, SegmentComponent};
 use std::io;
+use std::io::Write;
 use std::str;
+use crate::indexer::resource_manager::ResourceManager;
 
 /// Computes the initial size of the hash table.
 ///
 /// Returns a number of bit `b`, such that the recommended initial table size is 2^b.
-fn initial_table_size(per_thread_memory_budget: usize) -> crate::Result<usize> {
-    let table_memory_upper_bound = per_thread_memory_budget / 3;
+fn initial_table_size(per_thread_memory_budget: u64) -> crate::Result<usize> {
+    let table_memory_upper_bound = per_thread_memory_budget / 3u64;
     if let Some(limit) = (10..)
-        .take_while(|num_bits: &usize| compute_table_size(*num_bits) < table_memory_upper_bound)
+        .take_while(|num_bits| compute_table_size(*num_bits) < table_memory_upper_bound)
         .last()
     {
         Ok(limit.min(19)) // we cap it at 2^19 = 512K.
@@ -43,11 +48,13 @@ fn initial_table_size(per_thread_memory_budget: usize) -> crate::Result<usize> {
 pub struct SegmentWriter {
     max_doc: DocId,
     multifield_postings: MultiFieldPostingsWriter,
-    segment_serializer: SegmentSerializer,
+    segment: Segment,
     fast_field_writers: FastFieldsWriter,
     fieldnorms_writer: FieldNormsWriter,
     doc_opstamps: Vec<Opstamp>,
     tokenizers: Vec<Option<TextAnalyzer>>,
+    store_writer: StoreWriter<SpillingWriter>,
+    memory_manager: ResourceManager,
 }
 
 impl SegmentWriter {
@@ -61,12 +68,13 @@ impl SegmentWriter {
     /// - segment: The segment being written
     /// - schema
     pub fn for_segment(
-        memory_budget: usize,
-        mut segment: Segment,
+        config: &IndexWriterConfig,
+        segment: Segment,
         schema: &Schema,
+        tokenizer_manager: &TokenizerManager,
+        memory_manager: ResourceManager
     ) -> crate::Result<SegmentWriter> {
-        let table_num_bits = initial_table_size(memory_budget)?;
-        let segment_serializer = SegmentSerializer::for_segment(&mut segment)?;
+        let table_num_bits = initial_table_size(config.heap_size_in_byte_per_thread())?;
         let multifield_postings = MultiFieldPostingsWriter::new(schema, table_num_bits);
         let tokenizers = schema
             .fields()
@@ -76,20 +84,32 @@ impl SegmentWriter {
                         .get_indexing_options()
                         .and_then(|text_index_option| {
                             let tokenizer_name = &text_index_option.tokenizer();
-                            segment.index().tokenizers().get(tokenizer_name)
+                            tokenizer_manager.get(tokenizer_name)
                         }),
                     _ => None,
                 },
             )
             .collect();
+        let mut segment_clone = segment.clone();
+        let spilling_wrt = SpillingWriter::new(
+            50_000_000,
+            Box::new(move || {
+                segment_clone
+                    .open_write(SegmentComponent::STORE)
+                    .map_err(|err| io::Error::new(io::ErrorKind::Other, err))
+            }),
+        );
+        let store_writer = StoreWriter::new(spilling_wrt);
         Ok(SegmentWriter {
             max_doc: 0,
             multifield_postings,
             fieldnorms_writer: FieldNormsWriter::for_schema(schema),
-            segment_serializer,
+            segment,
             fast_field_writers: FastFieldsWriter::from_schema(schema),
             doc_opstamps: Vec::with_capacity(1_000),
             tokenizers,
+            store_writer,
+            memory_manager
         })
     }
 
@@ -97,18 +117,33 @@ impl SegmentWriter {
     ///
     /// Finalize consumes the `SegmentWriter`, so that it cannot
     /// be used afterwards.
-    pub fn finalize(mut self) -> crate::Result<Vec<u64>> {
+    pub fn finalize(mut self) -> crate::Result<(Segment, Vec<u64>)> {
         self.fieldnorms_writer.fill_up_to_max_doc(self.max_doc);
+        let spilling_wrt = self.store_writer.close()?;
+        let mut segment: Segment;
+        match spilling_wrt.finalize()? {
+            SpillingResult::Spilled => {
+                segment = self.segment.clone();
+            }
+            SpillingResult::Buffer(buf) => {
+                segment = self.segment.into_volatile(self.memory_manager.clone());
+                let mut store_wrt = segment.open_write(SegmentComponent::STORE)?;
+                store_wrt.write_all(&buf[..])?;
+                store_wrt.terminate()?;
+            }
+        }
+        let segment_serializer = SegmentSerializer::for_segment(&mut segment)?;
+        segment = segment.with_max_doc(self.max_doc);
         write(
             &self.multifield_postings,
             &self.fast_field_writers,
             &self.fieldnorms_writer,
-            self.segment_serializer,
+            segment_serializer,
         )?;
-        Ok(self.doc_opstamps)
+        Ok((segment, self.doc_opstamps))
     }
 
-    pub fn mem_usage(&self) -> usize {
+    pub fn mem_usage(&self) -> u64 {
         self.multifield_postings.mem_usage()
     }
 
@@ -246,8 +281,7 @@ impl SegmentWriter {
         }
         doc.filter_fields(|field| schema.get_field_entry(field).is_stored());
         doc.prepare_for_store();
-        let doc_writer = self.segment_serializer.get_store_writer();
-        doc_writer.store(&doc)?;
+        self.store_writer.store(&doc)?;
         self.max_doc += 1;
         Ok(())
     }

src/lib.rs

@@ -163,7 +163,7 @@ pub use crate::core::{Index, IndexMeta, Searcher, Segment, SegmentId, SegmentMet
 pub use crate::core::{InvertedIndexReader, SegmentReader};
 pub use crate::directory::Directory;
 pub use crate::indexer::operation::UserOperation;
-pub use crate::indexer::IndexWriter;
+pub use crate::indexer::{IndexWriter, IndexWriterConfig};
 pub use crate::postings::Postings;
 pub use crate::reader::LeasedItem;
 pub use crate::schema::{Document, Term};

src/postings/block_max_postings.rs

@@ -1,16 +0,0 @@
-use crate::postings::Postings;
-use crate::DocId;
-
-/// Inverted list with additional information about the maximum term frequency
-/// within a block, as well as globally within the list.
-pub trait BlockMaxPostings: Postings {
-    /// Returns the maximum frequency in the entire list.
-    fn max_term_freq(&self) -> u32;
-    /// Returns the maximum frequency in the current block.
-    fn block_max_term_freq(&mut self) -> u32;
-    /// Returns the document with the largest frequency.
-    fn max_doc(&self) -> DocId;
-    /// Returns the document with the largest frequency within the current
-    /// block.
-    fn block_max_doc(&self) -> DocId;
-}

src/postings/block_max_segment_postings.rs

@@ -1,76 +0,0 @@
-use crate::postings::{BlockMaxPostings, Postings, SegmentPostings};
-use crate::{DocId, DocSet, SkipResult};
-
-/// A wrapper over [`SegmentPostings`](./struct.SegmentPostings.html)
-/// with max block frequencies.
-pub struct BlockMaxSegmentPostings {
-    postings: SegmentPostings,
-    max_blocks: SegmentPostings,
-    doc_with_max_term_freq: DocId,
-    max_term_freq: u32,
-}
-
-impl BlockMaxSegmentPostings {
-    /// Constructs a new segment postings with block-max information.
-    pub fn new(
-        postings: SegmentPostings,
-        max_blocks: SegmentPostings,
-        doc_with_max_term_freq: DocId,
-        max_term_freq: u32,
-    ) -> Self {
-        Self {
-            postings,
-            max_blocks,
-            doc_with_max_term_freq,
-            max_term_freq,
-        }
-    }
-}
-
-impl DocSet for BlockMaxSegmentPostings {
-    fn advance(&mut self) -> bool {
-        self.postings.advance()
-    }
-
-    fn doc(&self) -> DocId {
-        self.postings.doc()
-    }
-
-    fn size_hint(&self) -> u32 {
-        self.postings.size_hint()
-    }
-
-    fn skip_next(&mut self, target: DocId) -> SkipResult {
-        self.postings.skip_next(target)
-    }
-}
-
-impl Postings for BlockMaxSegmentPostings {
-    fn term_freq(&self) -> u32 {
-        self.postings.term_freq()
-    }
-    fn positions_with_offset(&mut self, offset: u32, output: &mut Vec<u32>) {
-        self.postings.positions_with_offset(offset, output);
-    }
-    fn positions(&mut self, output: &mut Vec<u32>) {
-        self.postings.positions(output);
-    }
-}
-
-impl BlockMaxPostings for BlockMaxSegmentPostings {
-    fn max_term_freq(&self) -> u32 {
-        self.max_term_freq
-    }
-    fn block_max_term_freq(&mut self) -> u32 {
-        if let SkipResult::End = self.max_blocks.skip_next(self.doc()) {
-            panic!("Max blocks corrupted: reached end of max block");
-        }
-        self.max_blocks.term_freq()
-    }
-    fn max_doc(&self) -> DocId {
-        self.doc_with_max_term_freq
-    }
-    fn block_max_doc(&self) -> DocId {
-        self.max_blocks.doc()
-    }
-}

src/postings/block_segment_postings.rs

@@ -1,316 +0,0 @@
-use crate::DocId;
-use tantivy_fst::Streamer;
-use crate::postings::{SkipReader, FreqReadingOption, USE_SKIP_INFO_LIMIT};
-use owned_read::OwnedRead;
-use crate::postings::compression::{BlockDecoder, COMPRESSION_BLOCK_SIZE, VIntDecoder, compressed_block_size, AlignedBuffer};
-use crate::schema::IndexRecordOption;
-use crate::common::{VInt, BinarySerializable};
-
-
-fn split_into_skips_and_postings(
-    doc_freq: u32,
-    mut data: OwnedRead,
-) -> (Option<OwnedRead>, OwnedRead) {
-    if doc_freq >= USE_SKIP_INFO_LIMIT {
-        let skip_len = VInt::deserialize(&mut data).expect("Data corrupted").0 as usize;
-        let mut postings_data = data.clone();
-        postings_data.advance(skip_len);
-        data.clip(skip_len);
-        (Some(data), postings_data)
-    } else {
-        (None, data)
-    }
-}
-
-/// `BlockSegmentPostings` is a cursor iterating over blocks
-/// of documents.
-///
-/// # Warning
-///
-/// While it is useful for some very specific high-performance
-/// use cases, you should prefer using `SegmentPostings` for most usage.
-pub struct BlockSegmentPostings {
-    doc_decoder: BlockDecoder,
-    freq_decoder: BlockDecoder,
-    freq_reading_option: FreqReadingOption,
-
-    doc_freq: usize,
-    doc_offset: DocId,
-
-    num_vint_docs: usize,
-
-    remaining_data: OwnedRead,
-    skip_reader: SkipReader,
-}
-
-
-#[derive(Debug, Eq, PartialEq)]
-pub enum BlockSegmentPostingsSkipResult {
-    Terminated,
-    Success(u32), //< number of term freqs to skip
-}
-
-impl BlockSegmentPostings {
-    pub(crate) fn from_data(
-        doc_freq: u32,
-        data: OwnedRead,
-        record_option: IndexRecordOption,
-        requested_option: IndexRecordOption,
-    ) -> BlockSegmentPostings {
-        let freq_reading_option = match (record_option, requested_option) {
-            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
-            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
-            (_, _) => FreqReadingOption::ReadFreq,
-        };
-
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
-        let skip_reader = match skip_data_opt {
-            Some(skip_data) => SkipReader::new(skip_data, record_option),
-            None => SkipReader::new(OwnedRead::new(&[][..]), record_option),
-        };
-        let doc_freq = doc_freq as usize;
-        let num_vint_docs = doc_freq % COMPRESSION_BLOCK_SIZE;
-        BlockSegmentPostings {
-            num_vint_docs,
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option,
-            doc_offset: 0,
-            doc_freq,
-            remaining_data: postings_data,
-            skip_reader,
-        }
-    }
-
-    // Resets the block segment postings on another position
-    // in the postings file.
-    //
-    // This is useful for enumerating through a list of terms,
-    // and consuming the associated posting lists while avoiding
-    // reallocating a `BlockSegmentPostings`.
-    //
-    // # Warning
-    //
-    // This does not reset the positions list.
-    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: OwnedRead) {
-        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
-        let num_vint_docs = (doc_freq as usize) & (COMPRESSION_BLOCK_SIZE - 1);
-        self.num_vint_docs = num_vint_docs;
-        self.remaining_data = postings_data;
-        if let Some(skip_data) = skip_data_opt {
-            self.skip_reader.reset(skip_data);
-        } else {
-            self.skip_reader.reset(OwnedRead::new(&[][..]))
-        }
-        self.doc_offset = 0;
-        self.doc_freq = doc_freq as usize;
-    }
-
-    /// Returns the document frequency associated to this block postings.
-    ///
-    /// This `doc_freq` is simply the sum of the length of all of the blocks
-    /// length, and it does not take in account deleted documents.
-    pub fn doc_freq(&self) -> usize {
-        self.doc_freq
-    }
-
-    /// Returns the array of docs in the current block.
-    ///
-    /// Before the first call to `.advance()`, the block
-    /// returned by `.docs()` is empty.
-    #[inline]
-    pub fn docs(&self) -> &[DocId] {
-        self.doc_decoder.output_array()
-    }
-
-    pub(crate) fn docs_aligned(&self) -> (&AlignedBuffer, usize) {
-        self.doc_decoder.output_aligned()
-    }
-
-    /// Return the document at index `idx` of the block.
-    #[inline]
-    pub fn doc(&self, idx: usize) -> u32 {
-        self.doc_decoder.output(idx)
-    }
-
-    /// Return the array of `term freq` in the block.
-    #[inline]
-    pub fn freqs(&self) -> &[u32] {
-        self.freq_decoder.output_array()
-    }
-
-    /// Return the frequency at index `idx` of the block.
-    #[inline]
-    pub fn freq(&self, idx: usize) -> u32 {
-        self.freq_decoder.output(idx)
-    }
-
-    /// Returns the length of the current block.
-    ///
-    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
-    /// except the last block that may have a length
-    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
-    #[inline]
-    pub(crate) fn block_len(&self) -> usize {
-        self.doc_decoder.output_len
-    }
-
-    /// position on a block that may contains `doc_id`.
-    /// Always advance the current block.
-    ///
-    /// Returns true if a block that has an element greater or equal to the target is found.
-    /// Returning true does not guarantee that the smallest element of the block is smaller
-    /// than the target. It only guarantees that the last element is greater or equal.
-    ///
-    /// Returns false iff all of the document remaining are smaller than
-    /// `doc_id`. In that case, all of these document are consumed.
-    ///
-    pub fn skip_to(&mut self, target_doc: DocId) -> BlockSegmentPostingsSkipResult {
-        let mut skip_freqs = 0u32;
-        while self.skip_reader.advance() {
-            if self.skip_reader.doc() >= target_doc {
-                // the last document of the current block is larger
-                // than the target.
-                //
-                // We found our block!
-                let num_bits = self.skip_reader.doc_num_bits();
-                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                    self.remaining_data.as_ref(),
-                    self.doc_offset,
-                    num_bits,
-                );
-                self.remaining_data.advance(num_consumed_bytes);
-                let tf_num_bits = self.skip_reader.tf_num_bits();
-                match self.freq_reading_option {
-                    FreqReadingOption::NoFreq => {}
-                    FreqReadingOption::SkipFreq => {
-                        let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                        self.remaining_data.advance(num_bytes_to_skip);
-                    }
-                    FreqReadingOption::ReadFreq => {
-                        let num_consumed_bytes = self
-                            .freq_decoder
-                            .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                        self.remaining_data.advance(num_consumed_bytes);
-                    }
-                }
-                self.doc_offset = self.skip_reader.doc();
-                return BlockSegmentPostingsSkipResult::Success(skip_freqs);
-            } else {
-                skip_freqs += self.skip_reader.tf_sum();
-                let advance_len = self.skip_reader.total_block_len();
-                self.doc_offset = self.skip_reader.doc();
-                self.remaining_data.advance(advance_len);
-            }
-        }
-
-        // we are now on the last, incomplete, variable encoded block.
-        if self.num_vint_docs > 0 {
-            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                self.num_vint_docs,
-            );
-            self.remaining_data.advance(num_compressed_bytes);
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-                FreqReadingOption::ReadFreq => {
-                    self.freq_decoder
-                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-                }
-            }
-            self.num_vint_docs = 0;
-            return self
-                .docs()
-                .last()
-                .map(|last_doc| {
-                    if *last_doc >= target_doc {
-                        BlockSegmentPostingsSkipResult::Success(skip_freqs)
-                    } else {
-                        BlockSegmentPostingsSkipResult::Terminated
-                    }
-                })
-                .unwrap_or(BlockSegmentPostingsSkipResult::Terminated);
-        }
-        BlockSegmentPostingsSkipResult::Terminated
-    }
-
-    /// Advance to the next block.
-    ///
-    /// Returns false iff there was no remaining blocks.
-    pub fn advance(&mut self) -> bool {
-        if self.skip_reader.advance() {
-            let num_bits = self.skip_reader.doc_num_bits();
-            let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                num_bits,
-            );
-            self.remaining_data.advance(num_consumed_bytes);
-            let tf_num_bits = self.skip_reader.tf_num_bits();
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq => {}
-                FreqReadingOption::SkipFreq => {
-                    let num_bytes_to_skip = compressed_block_size(tf_num_bits);
-                    self.remaining_data.advance(num_bytes_to_skip);
-                }
-                FreqReadingOption::ReadFreq => {
-                    let num_consumed_bytes = self
-                        .freq_decoder
-                        .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
-                    self.remaining_data.advance(num_consumed_bytes);
-                }
-            }
-            // it will be used as the next offset.
-            self.doc_offset = self.doc_decoder.output(COMPRESSION_BLOCK_SIZE - 1);
-            true
-        } else if self.num_vint_docs > 0 {
-            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
-                self.remaining_data.as_ref(),
-                self.doc_offset,
-                self.num_vint_docs,
-            );
-            self.remaining_data.advance(num_compressed_bytes);
-            match self.freq_reading_option {
-                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
-                FreqReadingOption::ReadFreq => {
-                    self.freq_decoder
-                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
-                }
-            }
-            self.num_vint_docs = 0;
-            true
-        } else {
-            false
-        }
-    }
-
-    /// Returns an empty segment postings object
-    pub fn empty() -> BlockSegmentPostings {
-        BlockSegmentPostings {
-            num_vint_docs: 0,
-
-            doc_decoder: BlockDecoder::new(),
-            freq_decoder: BlockDecoder::with_val(1),
-            freq_reading_option: FreqReadingOption::NoFreq,
-
-            doc_offset: 0,
-            doc_freq: 0,
-
-            remaining_data: OwnedRead::new(vec![]),
-            skip_reader: SkipReader::new(OwnedRead::new(vec![]), IndexRecordOption::Basic),
-        }
-    }
-}
-
-impl<'a> Streamer<'a> for BlockSegmentPostings {
-    type Item = &'a [DocId];
-
-    fn next(&'a mut self) -> Option<&'a [DocId]> {
-        if self.advance() {
-            Some(self.docs())
-        } else {
-            None
-        }
-    }
-}

src/postings/mod.rs

@@ -2,8 +2,6 @@
 Postings module (also called inverted index)
 */
 
-mod block_max_postings;
-mod block_max_segment_postings;
 mod block_search;
 pub(crate) mod compression;
 /// Postings module
@@ -14,7 +12,6 @@ mod postings;
 mod postings_writer;
 mod recorder;
 mod segment_postings;
-mod block_segment_postings;
 mod serializer;
 mod skip;
 mod stacker;
@@ -30,11 +27,7 @@ pub use self::postings::Postings;
 pub(crate) use self::skip::SkipReader;
 pub use self::term_info::TermInfo;
 
-pub use self::segment_postings::SegmentPostings;
-
-pub use self::block_segment_postings::BlockSegmentPostings;
-pub use self::block_max_postings::BlockMaxPostings;
-pub use self::block_max_segment_postings::BlockMaxSegmentPostings;
+pub use self::segment_postings::{BlockSegmentPostings, SegmentPostings};
 
 pub(crate) use self::stacker::compute_table_size;
 
@@ -61,7 +54,7 @@ pub mod tests {
     use crate::docset::{DocSet, SkipResult};
     use crate::fieldnorm::FieldNormReader;
     use crate::indexer::operation::AddOperation;
-    use crate::indexer::SegmentWriter;
+    use crate::indexer::{IndexWriterConfig, SegmentWriter, ResourceManager};
     use crate::merge_policy::NoMergePolicy;
     use crate::query::Scorer;
     use crate::schema::{Document, Schema, Term, INDEXED, STRING, TEXT};
@@ -223,11 +216,17 @@ pub mod tests {
         let text_field = schema_builder.add_text_field("text", TEXT);
         let schema = schema_builder.build();
         let index = Index::create_in_ram(schema.clone());
-        let segment = index.new_segment();
 
-        {
-            let mut segment_writer =
-                SegmentWriter::for_segment(3_000_000, segment.clone(), &schema).unwrap();
+        let resource_manager = ResourceManager::default();
+        let segment = {
+            let mut segment_writer = SegmentWriter::for_segment(
+                &IndexWriterConfig::for_test(),
+                index.new_segment(),
+                &schema,
+                index.tokenizers(),
+                resource_manager
+            )
+            .unwrap();
             {
                 let mut doc = Document::default();
                 // checking that position works if the field has two values
@@ -259,8 +258,9 @@ pub mod tests {
                 };
                 segment_writer.add_document(op, &schema).unwrap();
             }
-            segment_writer.finalize().unwrap();
-        }
+            let (segment, _) = segment_writer.finalize().unwrap();
+            segment
+        };
         {
             let segment_reader = SegmentReader::open(&segment).unwrap();
             {

src/postings/postings_writer.rs

@@ -95,7 +95,7 @@ impl MultiFieldPostingsWriter {
         }
     }
 
-    pub fn mem_usage(&self) -> usize {
+    pub fn mem_usage(&self) -> u64 {
         self.term_index.mem_usage() + self.heap.mem_usage()
     }
 

src/postings/segment_postings.rs

@@ -1,16 +1,21 @@
 use crate::common::BitSet;
 use crate::common::HasLen;
+use crate::common::{BinarySerializable, VInt};
 use crate::docset::{DocSet, SkipResult};
 use crate::positions::PositionReader;
-use crate::postings::compression::COMPRESSION_BLOCK_SIZE;
+use crate::postings::compression::{compressed_block_size, AlignedBuffer};
+use crate::postings::compression::{BlockDecoder, VIntDecoder, COMPRESSION_BLOCK_SIZE};
 use crate::postings::serializer::PostingsSerializer;
-use crate::postings::{BlockSearcher, BlockSegmentPostings};
+use crate::postings::BlockSearcher;
+use crate::postings::FreqReadingOption;
 use crate::postings::Postings;
+use crate::postings::SkipReader;
+use crate::postings::USE_SKIP_INFO_LIMIT;
 use crate::schema::IndexRecordOption;
 use crate::DocId;
 use owned_read::OwnedRead;
 use std::cmp::Ordering;
-use crate::postings::block_segment_postings::BlockSegmentPostingsSkipResult;
+use tantivy_fst::Streamer;
 
 struct PositionComputer {
     // store the amount of position int
@@ -294,6 +299,313 @@ impl Postings for SegmentPostings {
     }
 }
 
+/// `BlockSegmentPostings` is a cursor iterating over blocks
+/// of documents.
+///
+/// # Warning
+///
+/// While it is useful for some very specific high-performance
+/// use cases, you should prefer using `SegmentPostings` for most usage.
+pub struct BlockSegmentPostings {
+    doc_decoder: BlockDecoder,
+    freq_decoder: BlockDecoder,
+    freq_reading_option: FreqReadingOption,
+
+    doc_freq: usize,
+    doc_offset: DocId,
+
+    num_vint_docs: usize,
+
+    remaining_data: OwnedRead,
+    skip_reader: SkipReader,
+}
+
+fn split_into_skips_and_postings(
+    doc_freq: u32,
+    mut data: OwnedRead,
+) -> (Option<OwnedRead>, OwnedRead) {
+    if doc_freq >= USE_SKIP_INFO_LIMIT {
+        let skip_len = VInt::deserialize(&mut data).expect("Data corrupted").0 as usize;
+        let mut postings_data = data.clone();
+        postings_data.advance(skip_len);
+        data.clip(skip_len);
+        (Some(data), postings_data)
+    } else {
+        (None, data)
+    }
+}
+
+#[derive(Debug, Eq, PartialEq)]
+pub enum BlockSegmentPostingsSkipResult {
+    Terminated,
+    Success(u32), //< number of term freqs to skip
+}
+
+impl BlockSegmentPostings {
+    pub(crate) fn from_data(
+        doc_freq: u32,
+        data: OwnedRead,
+        record_option: IndexRecordOption,
+        requested_option: IndexRecordOption,
+    ) -> BlockSegmentPostings {
+        let freq_reading_option = match (record_option, requested_option) {
+            (IndexRecordOption::Basic, _) => FreqReadingOption::NoFreq,
+            (_, IndexRecordOption::Basic) => FreqReadingOption::SkipFreq,
+            (_, _) => FreqReadingOption::ReadFreq,
+        };
+
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, data);
+        let skip_reader = match skip_data_opt {
+            Some(skip_data) => SkipReader::new(skip_data, record_option),
+            None => SkipReader::new(OwnedRead::new(&[][..]), record_option),
+        };
+        let doc_freq = doc_freq as usize;
+        let num_vint_docs = doc_freq % COMPRESSION_BLOCK_SIZE;
+        BlockSegmentPostings {
+            num_vint_docs,
+            doc_decoder: BlockDecoder::new(),
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option,
+            doc_offset: 0,
+            doc_freq,
+            remaining_data: postings_data,
+            skip_reader,
+        }
+    }
+
+    // Resets the block segment postings on another position
+    // in the postings file.
+    //
+    // This is useful for enumerating through a list of terms,
+    // and consuming the associated posting lists while avoiding
+    // reallocating a `BlockSegmentPostings`.
+    //
+    // # Warning
+    //
+    // This does not reset the positions list.
+    pub(crate) fn reset(&mut self, doc_freq: u32, postings_data: OwnedRead) {
+        let (skip_data_opt, postings_data) = split_into_skips_and_postings(doc_freq, postings_data);
+        let num_vint_docs = (doc_freq as usize) & (COMPRESSION_BLOCK_SIZE - 1);
+        self.num_vint_docs = num_vint_docs;
+        self.remaining_data = postings_data;
+        if let Some(skip_data) = skip_data_opt {
+            self.skip_reader.reset(skip_data);
+        } else {
+            self.skip_reader.reset(OwnedRead::new(&[][..]))
+        }
+        self.doc_offset = 0;
+        self.doc_freq = doc_freq as usize;
+    }
+
+    /// Returns the document frequency associated to this block postings.
+    ///
+    /// This `doc_freq` is simply the sum of the length of all of the blocks
+    /// length, and it does not take in account deleted documents.
+    pub fn doc_freq(&self) -> usize {
+        self.doc_freq
+    }
+
+    /// Returns the array of docs in the current block.
+    ///
+    /// Before the first call to `.advance()`, the block
+    /// returned by `.docs()` is empty.
+    #[inline]
+    pub fn docs(&self) -> &[DocId] {
+        self.doc_decoder.output_array()
+    }
+
+    pub(crate) fn docs_aligned(&self) -> (&AlignedBuffer, usize) {
+        self.doc_decoder.output_aligned()
+    }
+
+    /// Return the document at index `idx` of the block.
+    #[inline]
+    pub fn doc(&self, idx: usize) -> u32 {
+        self.doc_decoder.output(idx)
+    }
+
+    /// Return the array of `term freq` in the block.
+    #[inline]
+    pub fn freqs(&self) -> &[u32] {
+        self.freq_decoder.output_array()
+    }
+
+    /// Return the frequency at index `idx` of the block.
+    #[inline]
+    pub fn freq(&self, idx: usize) -> u32 {
+        self.freq_decoder.output(idx)
+    }
+
+    /// Returns the length of the current block.
+    ///
+    /// All blocks have a length of `NUM_DOCS_PER_BLOCK`,
+    /// except the last block that may have a length
+    /// of any number between 1 and `NUM_DOCS_PER_BLOCK - 1`
+    #[inline]
+    fn block_len(&self) -> usize {
+        self.doc_decoder.output_len
+    }
+
+    /// position on a block that may contains `doc_id`.
+    /// Always advance the current block.
+    ///
+    /// Returns true if a block that has an element greater or equal to the target is found.
+    /// Returning true does not guarantee that the smallest element of the block is smaller
+    /// than the target. It only guarantees that the last element is greater or equal.
+    ///
+    /// Returns false iff all of the document remaining are smaller than
+    /// `doc_id`. In that case, all of these document are consumed.
+    ///
+    pub fn skip_to(&mut self, target_doc: DocId) -> BlockSegmentPostingsSkipResult {
+        let mut skip_freqs = 0u32;
+        while self.skip_reader.advance() {
+            if self.skip_reader.doc() >= target_doc {
+                // the last document of the current block is larger
+                // than the target.
+                //
+                // We found our block!
+                let num_bits = self.skip_reader.doc_num_bits();
+                let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
+                    self.remaining_data.as_ref(),
+                    self.doc_offset,
+                    num_bits,
+                );
+                self.remaining_data.advance(num_consumed_bytes);
+                let tf_num_bits = self.skip_reader.tf_num_bits();
+                match self.freq_reading_option {
+                    FreqReadingOption::NoFreq => {}
+                    FreqReadingOption::SkipFreq => {
+                        let num_bytes_to_skip = compressed_block_size(tf_num_bits);
+                        self.remaining_data.advance(num_bytes_to_skip);
+                    }
+                    FreqReadingOption::ReadFreq => {
+                        let num_consumed_bytes = self
+                            .freq_decoder
+                            .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
+                        self.remaining_data.advance(num_consumed_bytes);
+                    }
+                }
+                self.doc_offset = self.skip_reader.doc();
+                return BlockSegmentPostingsSkipResult::Success(skip_freqs);
+            } else {
+                skip_freqs += self.skip_reader.tf_sum();
+                let advance_len = self.skip_reader.total_block_len();
+                self.doc_offset = self.skip_reader.doc();
+                self.remaining_data.advance(advance_len);
+            }
+        }
+
+        // we are now on the last, incomplete, variable encoded block.
+        if self.num_vint_docs > 0 {
+            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                self.num_vint_docs,
+            );
+            self.remaining_data.advance(num_compressed_bytes);
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
+                FreqReadingOption::ReadFreq => {
+                    self.freq_decoder
+                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
+                }
+            }
+            self.num_vint_docs = 0;
+            return self
+                .docs()
+                .last()
+                .map(|last_doc| {
+                    if *last_doc >= target_doc {
+                        BlockSegmentPostingsSkipResult::Success(skip_freqs)
+                    } else {
+                        BlockSegmentPostingsSkipResult::Terminated
+                    }
+                })
+                .unwrap_or(BlockSegmentPostingsSkipResult::Terminated);
+        }
+        BlockSegmentPostingsSkipResult::Terminated
+    }
+
+    /// Advance to the next block.
+    ///
+    /// Returns false iff there was no remaining blocks.
+    pub fn advance(&mut self) -> bool {
+        if self.skip_reader.advance() {
+            let num_bits = self.skip_reader.doc_num_bits();
+            let num_consumed_bytes = self.doc_decoder.uncompress_block_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                num_bits,
+            );
+            self.remaining_data.advance(num_consumed_bytes);
+            let tf_num_bits = self.skip_reader.tf_num_bits();
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq => {}
+                FreqReadingOption::SkipFreq => {
+                    let num_bytes_to_skip = compressed_block_size(tf_num_bits);
+                    self.remaining_data.advance(num_bytes_to_skip);
+                }
+                FreqReadingOption::ReadFreq => {
+                    let num_consumed_bytes = self
+                        .freq_decoder
+                        .uncompress_block_unsorted(self.remaining_data.as_ref(), tf_num_bits);
+                    self.remaining_data.advance(num_consumed_bytes);
+                }
+            }
+            // it will be used as the next offset.
+            self.doc_offset = self.doc_decoder.output(COMPRESSION_BLOCK_SIZE - 1);
+            true
+        } else if self.num_vint_docs > 0 {
+            let num_compressed_bytes = self.doc_decoder.uncompress_vint_sorted(
+                self.remaining_data.as_ref(),
+                self.doc_offset,
+                self.num_vint_docs,
+            );
+            self.remaining_data.advance(num_compressed_bytes);
+            match self.freq_reading_option {
+                FreqReadingOption::NoFreq | FreqReadingOption::SkipFreq => {}
+                FreqReadingOption::ReadFreq => {
+                    self.freq_decoder
+                        .uncompress_vint_unsorted(self.remaining_data.as_ref(), self.num_vint_docs);
+                }
+            }
+            self.num_vint_docs = 0;
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Returns an empty segment postings object
+    pub fn empty() -> BlockSegmentPostings {
+        BlockSegmentPostings {
+            num_vint_docs: 0,
+
+            doc_decoder: BlockDecoder::new(),
+            freq_decoder: BlockDecoder::with_val(1),
+            freq_reading_option: FreqReadingOption::NoFreq,
+
+            doc_offset: 0,
+            doc_freq: 0,
+
+            remaining_data: OwnedRead::new(vec![]),
+            skip_reader: SkipReader::new(OwnedRead::new(vec![]), IndexRecordOption::Basic),
+        }
+    }
+}
+
+impl<'b> Streamer<'b> for BlockSegmentPostings {
+    type Item = &'b [DocId];
+
+    fn next(&'b mut self) -> Option<&'b [DocId]> {
+        if self.advance() {
+            Some(self.docs())
+        } else {
+            None
+        }
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::BlockSegmentPostings;

src/postings/stacker/memory_arena.rs

@@ -106,8 +106,8 @@ impl MemoryArena {
     ///
     /// Internally, it counts a number of `1MB` pages
     /// and therefore delivers an upperbound.
-    pub fn mem_usage(&self) -> usize {
-        self.pages.len() * PAGE_SIZE
+    pub fn mem_usage(&self) -> u64 {
+        (self.pages.len() as u64) * (PAGE_SIZE as u64)
     }
 
     pub fn write_at<Item: Copy + 'static>(&mut self, addr: Addr, val: Item) {

src/postings/stacker/term_hashmap.rs

@@ -1,4 +1,6 @@
-use murmurhash32::murmurhash2;
+use murmurhash32;
+
+use self::murmurhash32::murmurhash2;
 
 use super::{Addr, MemoryArena};
 use crate::postings::stacker::memory_arena::store;
@@ -10,8 +12,8 @@ use std::slice;
 
 /// Returns the actual memory size in bytes
 /// required to create a table of size $2^num_bits$.
-pub fn compute_table_size(num_bits: usize) -> usize {
-    (1 << num_bits) * mem::size_of::<KeyValue>()
+pub fn compute_table_size(num_bits: usize) -> u64 {
+    (1u64 << num_bits as u64) * mem::size_of::<KeyValue>() as u64
 }
 
 /// `KeyValue` is the item stored in the hash table.
@@ -114,8 +116,8 @@ impl TermHashMap {
         QuadraticProbing::compute(hash as usize, self.mask)
     }
 
-    pub fn mem_usage(&self) -> usize {
-        self.table.len() * mem::size_of::<KeyValue>()
+    pub fn mem_usage(&self) -> u64 {
+        self.table.len() as u64 * mem::size_of::<KeyValue>() as u64
     }
 
     fn is_saturated(&self) -> bool {

src/query/block_max_scorer.rs

@@ -1,31 +0,0 @@
-use crate::docset::DocSet;
-use crate::DocId;
-use crate::Score;
-use downcast_rs::impl_downcast;
-use crate::query::Scorer;
-
-/// A set of documents matching a query within a specific segment
-/// and having a maximum score within certain blocks.
-///
-/// See [`Query`](./trait.Query.html) and [`Scorer`](./trait.Scorer.html).
-pub trait BlockMaxScorer: downcast_rs::Downcast + DocSet + Scorer + 'static {
-    /// Returns the maximum score within the current block.
-    ///
-    /// The blocks are defined when indexing. For example, blocks can be
-    /// have a specific number postings each, or can be optimized for
-    /// retrieval speed. Read more in
-    /// [Faster BlockMax WAND with Variable-sized Blocks][vbmw]
-    ///
-    /// This method will perform a bit of computation and is not cached.
-    ///
-    /// [vbmw]: https://dl.acm.org/doi/abs/10.1145/3077136.3080780
-    fn block_max_score(&mut self) -> Score;
-
-    /// Returns the last document in the current block.
-    fn block_max_doc(&mut self) -> DocId;
-
-    /// Returns the maximum possible score within the entire document set.
-    fn max_score(&self) -> Score;
-}
-
-impl_downcast!(BlockMaxScorer);

src/query/block_max_wand.rs

@@ -1,613 +0,0 @@
-use crate::docset::{DocSet, SkipResult};
-use crate::query::score_combiner::ScoreCombiner;
-use crate::query::{BlockMaxScorer, Scorer};
-use crate::DocId;
-use crate::Score;
-use crate::query::weight::PruningScorer;
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-struct Pivot {
-    position: usize,
-    first_occurrence: usize,
-    doc: DocId,
-}
-
-
-/// Find the position in the sorted list of posting lists of the **pivot**.
-///
-/// docsets need to be advanced, and are required to be sorted by the doc they point to.
-///
-/// The pivot is then defined as the lowest DocId that has a chance of matching our condition.
-fn find_pivot_position<'a, TScorer>(
-    mut docsets: impl Iterator<Item = &'a TScorer>,
-    lower_bound_score: Score,
-) -> Option<Pivot>
-    where TScorer: BlockMaxScorer
-{
-    let mut position = 0;
-    let mut upper_bound = Score::default();
-    while let Some(docset) = docsets.next() {
-        upper_bound += docset.max_score();
-        if lower_bound_score < upper_bound {
-            let pivot_doc = docset.doc();
-            let first_occurrence = position;
-            while let Some(docset) = docsets.next() {
-                if docset.doc() != pivot_doc {
-                    break;
-                } else {
-                    position += 1;
-                }
-            }
-            return Some(Pivot {
-                position,
-                doc: pivot_doc,
-                first_occurrence,
-            });
-        }
-        position += 1;
-    }
-    None
-}
-
-/// Given an iterator over all ordered lists up to the pivot (inclusive) and the following list (if
-/// exists), it returns the next document ID that can be possibly relevant, based on the block max
-/// scores.
-fn find_next_relevant_doc<T, TScorer>(
-    docsets_up_to_pivot: &mut [T],
-    pivot_docset: &mut T,
-    docset_after_pivot: Option<&mut T>,
-) -> DocId
-    where
-        T: AsMut<TScorer>,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    let pivot_docset = pivot_docset.as_mut();
-    let mut next_doc = 1 + docsets_up_to_pivot
-        .iter_mut()
-        .map(|docset| docset.as_mut().block_max_doc())
-        .chain(std::iter::once(pivot_docset.block_max_doc()))
-        .min()
-        .unwrap();
-    if let Some(docset) = docset_after_pivot {
-        let doc = docset.as_mut().doc();
-        if doc < next_doc {
-            next_doc = doc;
-        }
-    }
-    if next_doc <= pivot_docset.doc() {
-        pivot_docset.doc() + 1
-    } else {
-        next_doc
-    }
-}
-
-/// Sifts down the first element of the slice.
-///
-/// `docsets[1..]` are assumed sorted.
-/// This function swaps `docsets[0]` with its right
-/// neighbor successively -bubble sort style- until it reaches the first
-/// position such that `docsets` is sorted.
-fn sift_down<T, TScorer>(docsets: &mut [T])
-    where
-        T: AsRef<TScorer>,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    for idx in 1..docsets.len() {
-        if docsets[idx].as_ref().doc() >= docsets[idx - 1].as_ref().doc() {
-            return;
-        }
-        docsets.swap(idx, idx - 1);
-    }
-}
-
-/// Creates a `DocSet` that iterates through the union of two or more `DocSet`s,
-/// applying [BlockMaxWand] dynamic pruning.
-///
-/// [BlockMaxWand]: https://dl.acm.org/doi/10.1145/2009916.2010048
-pub struct BlockMaxWand<TScorer, TScoreCombiner> {
-    docsets: Vec<Box<TScorer>>,
-    doc: DocId,
-    score: Score,
-    combiner: TScoreCombiner,
-}
-
-impl<TScorer, TScoreCombiner> BlockMaxWand<TScorer, TScoreCombiner>
-    where
-        TScoreCombiner: ScoreCombiner,
-        TScorer: BlockMaxScorer + Scorer,
-{
-    fn new(
-        docsets: Vec<TScorer>,
-        combiner: TScoreCombiner,
-    ) -> BlockMaxWand<TScorer, TScoreCombiner> {
-        let mut non_empty_docsets: Vec<_> = docsets
-            .into_iter()
-            .flat_map(|mut docset| {
-                if docset.advance() {
-                    Some(Box::new(docset))
-                } else {
-                    None
-                }
-            })
-            .collect();
-        non_empty_docsets.sort_by_key(Box::<TScorer>::doc);
-        BlockMaxWand {
-            docsets: non_empty_docsets,
-            combiner,
-            doc: 0u32,
-            score: 0f32
-        }
-    }
-
-    /// Find the position in the sorted list of posting lists of the **pivot**.
-    fn find_pivot_position(&self, lower_bound_score: Score) -> Option<Pivot> {
-        find_pivot_position(
-            self.docsets.iter().map(|docset| docset.as_ref()),
-            lower_bound_score)
-    }
-
-    fn advance_with_pivot(&mut self, pivot: Pivot, lower_bound_score: Score) -> SkipResult {
-        let block_upper_bound: Score = self.docsets[..=pivot.position]
-            .iter_mut()
-            .map(|docset| docset.block_max_score())
-            .sum();
-        if block_upper_bound > lower_bound_score {
-            if pivot.doc == self.docsets[0].doc() {
-                // Since self.docsets is sorted by their current doc, in this branch, all
-                // docsets in [0..=pivot] are positioned on pivot.doc.
-                //
-                // Lets compute the actual score for this doc.
-                //
-                // NOTE(elshize): One additional check needs to be done to improve performance:
-                // update block-wise bound while accumulating score with the actual score,
-                // and check each time if still above threshold.
-                self.combiner.clear();
-                for idx in (0..=pivot.position).rev() {
-                    self.combiner.update(self.docsets[idx].as_mut());
-                    if !self.docsets[idx].advance() {
-                        self.docsets.swap_remove(idx);
-                    }
-                }
-                self.score = self.combiner.score();
-                self.doc = pivot.doc;
-                self.docsets.sort_by_key(Box::<TScorer>::doc);
-                SkipResult::Reached
-            } else {
-                // The substraction does not underflow because otherwise we would go to the other
-                // branch.
-                //
-                // `advanced_idx` is the last idx that is not positionned on the pivot yet.
-                let advanced_idx = pivot.first_occurrence - 1;
-                if !self.docsets[advanced_idx].advance() {
-                    self.docsets.swap_remove(advanced_idx);
-                }
-                if self.docsets.is_empty() {
-                    return SkipResult::End;
-                }
-                sift_down(&mut self.docsets[advanced_idx..]);
-                SkipResult::OverStep
-            }
-        } else {
-            let (up_to_pivot, pivot_and_rest) = self.docsets.split_at_mut(pivot.position as usize);
-            let (pivot, after_pivot) = pivot_and_rest.split_first_mut().unwrap();
-            let next_doc = find_next_relevant_doc(up_to_pivot, pivot, after_pivot.first_mut());
-            // NOTE(elshize): It might be more efficient to advance the list with the higher
-            // max score, but let's advance the first one for now for simplicity.
-            if self.docsets[0].skip_next(next_doc) == SkipResult::End {
-                self.docsets.swap_remove(0);
-            }
-            if self.docsets.is_empty() {
-                return SkipResult::End;
-            }
-            sift_down(&mut self.docsets[..]);
-            SkipResult::OverStep
-        }
-    }
-}
-
-impl<TScorer, TScoreCombiner> PruningScorer
-for BlockMaxWand<TScorer, TScoreCombiner>
-    where
-        TScoreCombiner: ScoreCombiner,
-        TScorer: Scorer + BlockMaxScorer,
-{
-    fn doc(&self) -> DocId {
-        self.doc
-    }
-
-    fn score(&self) -> Score {
-        self.score
-    }
-
-    fn advance_with_pruning(&mut self, lower_bound_score: f32) -> bool {
-        while let Some(pivot) = self.find_pivot_position(lower_bound_score) {
-            match self.advance_with_pivot(pivot, lower_bound_score) {
-                SkipResult::End => { return false },
-                SkipResult::Reached=> { return true; }
-                SkipResult::OverStep => {}
-            }
-        }
-        false
-    }
-}
-
-
-#[cfg(test)]
-mod tests {
-
-    use super::*;
-    use crate::common::HasLen;
-    use crate::docset::DocSet;
-    use crate::query::score_combiner::SumCombiner;
-    use crate::query::Union;
-    use crate::query::{BlockMaxScorer, Scorer};
-    use crate::{DocId, Score};
-    use float_cmp::approx_eq;
-    use proptest::strategy::Strategy;
-    use std::cmp::Ordering;
-    use std::num::Wrapping;
-    use crate::collector::{SegmentCollector, TopScoreSegmentCollector};
-
-    /*
-    #[derive(Debug, Clone)]
-    pub struct VecDocSet {
-        postings: Vec<(DocId, Score)>,
-        cursor: Wrapping<usize>,
-        block_max_scores: Vec<(DocId, Score)>,
-        max_score: Score,
-        block_size: usize,
-    }
-
-    impl VecDocSet {
-        fn new(postings: Vec<(DocId, Score)>, block_size: usize) -> VecDocSet {
-            let block_max_scores: Vec<(DocId, f32)> = postings
-                .chunks(block_size)
-                .into_iter()
-                .map(|block| {
-                    (
-                        block.iter().last().unwrap().0,
-                        block
-                            .iter()
-                            .map(|(_, s)| *s)
-                            .fold(-f32::INFINITY, |left, right| left.max(right))
-                    )
-                })
-                .collect();
-            let max_score = block_max_scores
-                .iter()
-                .copied()
-                .map(|(_, s)| s)
-                .fold(-f32::INFINITY, |left, right| left.max(right));
-            VecDocSet {
-                postings,
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores,
-                max_score,
-                block_size,
-            }
-        }
-        /// Constructs a new set and advances it.
-        fn started(postings: Vec<(DocId, Score)>, block_size: usize) -> VecDocSet {
-            let mut docset = VecDocSet::new(postings, block_size);
-            docset.advance();
-            docset
-        }
-    }
-
-    impl DocSet for VecDocSet {
-        fn advance(&mut self) -> bool {
-            self.cursor += Wrapping(1);
-            self.postings.len() > self.cursor.0
-        }
-
-        fn doc(&self) -> DocId {
-            self.postings[self.cursor.0].0
-        }
-
-        fn size_hint(&self) -> u32 {
-            self.len() as u32
-        }
-    }
-
-    impl HasLen for VecDocSet {
-        fn len(&self) -> usize {
-            self.postings.len()
-        }
-    }
-
-    impl BlockMaxScorer for VecDocSet {
-        fn max_score(&self) -> Score {
-            self.max_score
-        }
-        fn block_max_score(&mut self) -> Score {
-            self.block_max_scores[self.cursor.0 / self.block_size].1
-        }
-        fn block_max_doc(&mut self) -> DocId {
-            self.block_max_scores[self.cursor.0 / self.block_size].0
-        }
-    }
-
-    impl Scorer for VecDocSet {
-        fn score(&mut self) -> Score {
-            self.postings[self.cursor.0].1
-        }
-    }
-
-    #[derive(Debug)]
-    struct ComparableDoc<T, D> {
-        feature: T,
-        doc: D,
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> PartialOrd for ComparableDoc<T, D> {
-        fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
-            Some(self.cmp(other))
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> Ord for ComparableDoc<T, D> {
-        #[inline]
-        fn cmp(&self, other: &Self) -> Ordering {
-            // Reversed to make BinaryHeap work as a min-heap
-            let by_feature = other
-                .feature
-                .partial_cmp(&self.feature)
-                .unwrap_or(Ordering::Equal);
-
-            let lazy_by_doc_address =
-                || self.doc.partial_cmp(&other.doc).unwrap_or(Ordering::Equal);
-
-            // In case of a tie on the feature, we sort by ascending
-            // `DocAddress` in order to ensure a stable sorting of the
-            // documents.
-            by_feature.then_with(lazy_by_doc_address)
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> PartialEq for ComparableDoc<T, D> {
-        fn eq(&self, other: &Self) -> bool {
-            self.cmp(other) == Ordering::Equal
-        }
-    }
-
-    impl<T: PartialOrd, D: PartialOrd> Eq for ComparableDoc<T, D> {}
-
-    fn union_vs_bmw(posting_lists: Vec<VecDocSet>) {
-        let mut union = Union::<VecDocSet, SumCombiner>::from(posting_lists.clone());
-        let mut top_union = TopScoreSegmentCollector::new(0, 10);
-        while union.advance() {
-            top_union.collect(union.doc(), union.score());
-        }
-        let top_bmw = TopScoreSegmentCollector::new(0, 10 );
-        let mut bmw = BlockMaxWand::new(posting_lists, SumCombiner::default());
-        let top_docs_bnw = top_bmw.collect_scorer(&mut bmw, None);
-        for ((expected_score, expected_doc), (actual_score, actual_doc)) in
-        top_union.harvest().into_iter().zip( top_docs_bnw )
-        {
-            assert!(approx_eq!(
-                f32,
-                expected_score,
-                actual_score,
-                epsilon = 0.0001
-            ));
-            assert_eq!(expected_doc, actual_doc);
-        }
-    }
-
-    #[test]
-    fn test_bmw_0() {
-        union_vs_bmw(vec![
-            VecDocSet {
-                postings: vec![
-                    (0, 1.0),
-                    (23, 1.0),
-                    (28, 1.0),
-                    (56, 1.0),
-                    (59, 1.0),
-                    (66, 1.0),
-                    (93, 1.0),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(93, 1.0)],
-                max_score: 1.0,
-                block_size: 16,
-            },
-            VecDocSet {
-                postings: vec![
-                    (2, 1.6549665),
-                    (43, 2.6958032),
-                    (53, 3.5309567),
-                    (71, 2.7688136),
-                    (87, 3.4279852),
-                    (96, 3.9028034),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(96, 3.9028034)],
-                max_score: 3.9028034,
-                block_size: 16,
-            },
-        ])
-    }
-
-    #[test]
-    fn test_bmw_1() {
-        union_vs_bmw(vec![
-            VecDocSet {
-                postings: vec![(73, 1.0), (82, 1.0)],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(82, 1.0)],
-                max_score: 1.0,
-                block_size: 16,
-            },
-            VecDocSet {
-                postings: vec![
-                    (21, 3.582513),
-                    (23, 1.6928024),
-                    (27, 3.887647),
-                    (42, 1.5469292),
-                    (61, 1.7317574),
-                    (62, 1.2968783),
-                    (82, 2.4040694),
-                    (85, 3.1487892),
-                ],
-                cursor: Wrapping(0_usize) - Wrapping(1_usize),
-                block_max_scores: vec![(85, 3.887647)],
-                max_score: 3.887647,
-                block_size: 16,
-            },
-        ])
-    }
-
-    proptest::proptest! {
-        #[test]
-        fn test_union_vs_bmw(postings in proptest::collection::vec(
-            proptest::collection::vec(0_u32..100, 1..10)
-                .prop_flat_map(|v| {
-                    let scores = proptest::collection::vec(1_f32..4_f32, v.len()..=v.len());
-                    scores.prop_map(move |s| {
-                        let mut postings: Vec<_> = v.iter().copied().zip(s.iter().copied()).collect();
-                        postings.sort_by_key(|p| p.0);
-                        postings.dedup_by_key(|p| p.0);
-                        VecDocSet::new(postings, 16)
-                    })
-                }),
-            2..5)
-        ) {
-            union_vs_bmw(postings);
-        }
-    }
-
-    #[test]
-    fn test_find_pivot_position() {
-        let postings = vec![
-            VecDocSet::started(vec![(0, 2.0)], 1),
-            VecDocSet::started(vec![(1, 3.0)], 1),
-            VecDocSet::started(vec![(2, 4.0)], 1),
-            VecDocSet::started(vec![(3, 5.0)], 1),
-            VecDocSet::started(vec![(3, 6.0)], 1),
-        ];
-        assert_eq!(
-            find_pivot_position(postings.iter(), 2.0f32),
-            Some(Pivot {
-                position: 1,
-                doc: 1,
-                first_occurrence: 1,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 5.0f32),
-            Some(Pivot {
-                position: 2,
-                doc: 2,
-                first_occurrence: 2,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 9.0f32),
-            Some(Pivot {
-                position: 4,
-                doc: 3,
-                first_occurrence: 3,
-            })
-        );
-        assert_eq!(
-            find_pivot_position(postings.iter(), 20.0f32),
-            None
-        );
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_before_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (3, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (4, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (6, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 4);
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_prefix_smaller_than_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (3, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (4, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(5, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 6);
-    }
-
-    #[test]
-    fn test_find_next_relevant_doc_after_pivot() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        let (up_to_pivot, rest) = postings.split_at_mut(2);
-        let (pivot, after_pivot) = rest.split_first_mut().unwrap();
-        let next_doc = find_next_relevant_doc(up_to_pivot, pivot, Some(&mut after_pivot[0]));
-        assert_eq!(next_doc, 5);
-    }
-
-    #[test]
-    fn test_sift_down_already_sifted() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(2, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 2, 5, 6]
-        );
-    }
-
-    #[test]
-    fn test_sift_down_sift_one_down() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(7, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 5, 6, 7]
-        );
-    }
-
-    #[test]
-    fn test_sift_down_to_bottom() {
-        let mut postings = vec![
-            Box::new(VecDocSet::started(vec![(0, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(1, 0.0), (8, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(7, 0.0), (8, 0.0)], 2)), // pivot
-            Box::new(VecDocSet::started(vec![(5, 0.0), (7, 0.0)], 2)),
-            Box::new(VecDocSet::started(vec![(6, 0.0), (7, 0.0)], 2)),
-        ];
-        sift_down(&mut postings[2..]);
-        assert_eq!(
-            postings.into_iter().map(|p| p.doc()).collect::<Vec<_>>(),
-            vec![0, 1, 5, 6, 7]
-        );
-    }
-    
-     */
-}

src/query/boolean_query/mod.rs

@@ -31,11 +31,24 @@ mod tests {
             // writing the segment
             let mut index_writer = index.writer_with_num_threads(1, 3_000_000).unwrap();
             {
-                index_writer.add_document(doc!(text_field => "a b c"));
-                index_writer.add_document(doc!(text_field => "a c"));
-                index_writer.add_document(doc!(text_field => "b c"));
-                index_writer.add_document(doc!(text_field => "a b c d"));
-                index_writer.add_document(doc!(text_field => "d"));
+                let doc = doc!(text_field => "a b c");
+                index_writer.add_document(doc);
+            }
+            {
+                let doc = doc!(text_field => "a c");
+                index_writer.add_document(doc);
+            }
+            {
+                let doc = doc!(text_field => "b c");
+                index_writer.add_document(doc);
+            }
+            {
+                let doc = doc!(text_field => "a b c d");
+                index_writer.add_document(doc);
+            }
+            {
+                let doc = doc!(text_field => "d");
+                index_writer.add_document(doc);
             }
             assert!(index_writer.commit().is_ok());
         }

src/query/fuzzy_query.rs

@@ -2,36 +2,10 @@ use crate::query::{AutomatonWeight, Query, Weight};
 use crate::schema::Term;
 use crate::Searcher;
 use crate::TantivyError::InvalidArgument;
-use levenshtein_automata::{Distance, LevenshteinAutomatonBuilder, DFA};
+use levenshtein_automata::{LevenshteinAutomatonBuilder, DFA};
 use once_cell::sync::Lazy;
 use std::collections::HashMap;
 use std::ops::Range;
-use tantivy_fst::Automaton;
-
-pub(crate) struct DFAWrapper(pub DFA);
-
-impl Automaton for DFAWrapper {
-    type State = u32;
-
-    fn start(&self) -> Self::State {
-        self.0.initial_state()
-    }
-
-    fn is_match(&self, state: &Self::State) -> bool {
-        match self.0.distance(*state) {
-            Distance::Exact(_) => true,
-            Distance::AtLeast(_) => false,
-        }
-    }
-
-    fn can_match(&self, state: &u32) -> bool {
-        *state != levenshtein_automata::SINK_STATE
-    }
-
-    fn accept(&self, state: &Self::State, byte: u8) -> Self::State {
-        self.0.transition(*state, byte)
-    }
-}
 
 /// A range of Levenshtein distances that we will build DFAs for our terms
 /// The computation is exponential, so best keep it to low single digits
@@ -127,20 +101,13 @@ impl FuzzyTermQuery {
         }
     }
 
-    fn specialized_weight(&self) -> crate::Result<AutomatonWeight<DFAWrapper>> {
+    fn specialized_weight(&self) -> crate::Result<AutomatonWeight<DFA>> {
         // LEV_BUILDER is a HashMap, whose `get` method returns an Option
         match LEV_BUILDER.get(&(self.distance, false)) {
             // Unwrap the option and build the Ok(AutomatonWeight)
             Some(automaton_builder) => {
-                let automaton = if self.prefix {
-                    automaton_builder.build_prefix_dfa(self.term.text())
-                } else {
-                    automaton_builder.build_dfa(self.term.text())
-                };
-                Ok(AutomatonWeight::new(
-                    self.term.field(),
-                    DFAWrapper(automaton),
-                ))
+                let automaton = automaton_builder.build_dfa(self.term.text());
+                Ok(AutomatonWeight::new(self.term.field(), automaton))
             }
             None => Err(InvalidArgument(format!(
                 "Levenshtein distance of {} is not allowed. Choose a value in the {:?} range",
@@ -199,17 +166,5 @@ mod test {
             let (score, _) = top_docs[0];
             assert_nearly_equals(1f32, score);
         }
-
-        {
-            let term = Term::from_field_text(country_field, "jap");
-
-            let fuzzy_query = FuzzyTermQuery::new_prefix(term, 1, true);
-            let top_docs = searcher
-                .search(&fuzzy_query, &TopDocs::with_limit(2))
-                .unwrap();
-            assert_eq!(top_docs.len(), 1, "Expected only 1 document");
-            let (score, _) = top_docs[0];
-            assert_nearly_equals(1f32, score);
-        }
     }
 }

src/query/mod.rs

@@ -1,10 +1,10 @@
-/*! Query Module */
+/*!
+Query
+*/
 
 mod all_query;
 mod automaton_weight;
 mod bitset;
-mod block_max_scorer;
-mod block_max_wand;
 mod bm25;
 mod boolean_query;
 mod boost_query;
@@ -37,14 +37,11 @@ pub use self::vec_docset::VecDocSet;
 pub use self::all_query::{AllQuery, AllScorer, AllWeight};
 pub use self::automaton_weight::AutomatonWeight;
 pub use self::bitset::BitSetDocSet;
-pub use self::block_max_scorer::BlockMaxScorer;
 pub use self::boolean_query::BooleanQuery;
 pub use self::boost_query::BoostQuery;
 pub use self::empty_query::{EmptyQuery, EmptyScorer, EmptyWeight};
 pub use self::exclude::Exclude;
 pub use self::explanation::Explanation;
-#[cfg(test)]
-pub(crate) use self::fuzzy_query::DFAWrapper;
 pub use self::fuzzy_query::FuzzyTermQuery;
 pub use self::intersection::intersect_scorers;
 pub use self::phrase_query::PhraseQuery;
@@ -59,8 +56,6 @@ pub use self::scorer::Scorer;
 pub use self::term_query::TermQuery;
 pub use self::weight::Weight;
 pub use tantivy_query_grammar::Occur;
-pub use self::weight::PruningScorerIfPossible;
-
 
 #[cfg(test)]
 mod tests {

src/query/phrase_query/mod.rs

@@ -12,6 +12,7 @@ pub mod tests {
     use super::*;
     use crate::collector::tests::{TEST_COLLECTOR_WITHOUT_SCORE, TEST_COLLECTOR_WITH_SCORE};
     use crate::core::Index;
+    use crate::error::TantivyError;
     use crate::schema::{Schema, Term, TEXT};
     use crate::tests::assert_nearly_equals;
     use crate::DocAddress;
@@ -126,16 +127,21 @@ pub mod tests {
             Term::from_field_text(text_field, "a"),
             Term::from_field_text(text_field, "b"),
         ]);
-
-        let search_result = searcher
+        match searcher
             .search(&phrase_query, &TEST_COLLECTOR_WITH_SCORE)
-            .map(|_| ());
-        assert!(matches!(
-            search_result,
-            Err(crate::TantivyError::SchemaError(msg))
-            if msg == "Applied phrase query on field \"text\", which does not have positions \
-            indexed"
-        ));
+            .map(|_| ())
+            .unwrap_err()
+        {
+            TantivyError::SchemaError(ref msg) => {
+                assert_eq!(
+                    "Applied phrase query on field \"text\", which does not have positions indexed",
+                    msg.as_str()
+                );
+            }
+            _ => {
+                panic!("Should have returned an error");
+            }
+        }
     }
 
     #[test]

src/query/query_parser/query_parser.rs

@@ -174,16 +174,6 @@ pub struct QueryParser {
     boost: HashMap<Field, f32>,
 }
 
-fn all_negative(ast: &LogicalAST) -> bool {
-    match ast {
-        LogicalAST::Leaf(_) => false,
-        LogicalAST::Boost(ref child_ast, _) => all_negative(&*child_ast),
-        LogicalAST::Clause(children) => children
-            .iter()
-            .all(|(ref occur, child)| (*occur == Occur::MustNot) || all_negative(child)),
-    }
-}
-
 impl QueryParser {
     /// Creates a `QueryParser`, given
     /// * schema - index Schema
@@ -263,13 +253,8 @@ impl QueryParser {
         &self,
         user_input_ast: UserInputAST,
     ) -> Result<LogicalAST, QueryParserError> {
-        let ast = self.compute_logical_ast_with_occur(user_input_ast)?;
-        if let LogicalAST::Clause(children) = &ast {
-            if children.is_empty() {
-                return Ok(ast);
-            }
-        }
-        if all_negative(&ast) {
+        let (occur, ast) = self.compute_logical_ast_with_occur(user_input_ast)?;
+        if occur == Occur::MustNot {
             return Err(QueryParserError::AllButQueryForbidden);
         }
         Ok(ast)
@@ -425,23 +410,31 @@ impl QueryParser {
     fn compute_logical_ast_with_occur(
         &self,
         user_input_ast: UserInputAST,
-    ) -> Result<LogicalAST, QueryParserError> {
+    ) -> Result<(Occur, LogicalAST), QueryParserError> {
         match user_input_ast {
             UserInputAST::Clause(sub_queries) => {
                 let default_occur = self.default_occur();
                 let mut logical_sub_queries: Vec<(Occur, LogicalAST)> = Vec::new();
-                for (occur_opt, sub_ast) in sub_queries {
-                    let sub_ast = self.compute_logical_ast_with_occur(sub_ast)?;
-                    let occur = occur_opt.unwrap_or(default_occur);
-                    logical_sub_queries.push((occur, sub_ast));
+                for sub_query in sub_queries {
+                    let (occur, sub_ast) = self.compute_logical_ast_with_occur(sub_query)?;
+                    let new_occur = Occur::compose(default_occur, occur);
+                    logical_sub_queries.push((new_occur, sub_ast));
                 }
-                Ok(LogicalAST::Clause(logical_sub_queries))
+                Ok((Occur::Should, LogicalAST::Clause(logical_sub_queries)))
+            }
+            UserInputAST::Unary(left_occur, subquery) => {
+                let (right_occur, logical_sub_queries) =
+                    self.compute_logical_ast_with_occur(*subquery)?;
+                Ok((Occur::compose(left_occur, right_occur), logical_sub_queries))
             }
             UserInputAST::Boost(ast, boost) => {
-                let ast = self.compute_logical_ast_with_occur(*ast)?;
-                Ok(ast.boost(boost))
+                let (occur, ast_without_occur) = self.compute_logical_ast_with_occur(*ast)?;
+                Ok((occur, ast_without_occur.boost(boost)))
+            }
+            UserInputAST::Leaf(leaf) => {
+                let result_ast = self.compute_logical_ast_from_leaf(*leaf)?;
+                Ok((Occur::Should, result_ast))
             }
-            UserInputAST::Leaf(leaf) => self.compute_logical_ast_from_leaf(*leaf),
         }
     }
 
@@ -789,20 +782,6 @@ mod test {
         );
     }
 
-    #[test]
-    fn test_parse_query_to_ast_ab_c() {
-        test_parse_query_to_logical_ast_helper(
-            "(+title:a +title:b) title:c",
-            "((+Term(field=0,bytes=[97]) +Term(field=0,bytes=[98])) Term(field=0,bytes=[99]))",
-            false,
-        );
-        test_parse_query_to_logical_ast_helper(
-            "(+title:a +title:b) title:c",
-            "(+(+Term(field=0,bytes=[97]) +Term(field=0,bytes=[98])) +Term(field=0,bytes=[99]))",
-            true,
-        );
-    }
-
     #[test]
     pub fn test_parse_query_to_ast_single_term() {
         test_parse_query_to_logical_ast_helper(
@@ -822,13 +801,11 @@ mod test {
              Term(field=1,bytes=[116, 105, 116, 105])))",
             false,
         );
-    }
-
-    #[test]
-    fn test_single_negative_term() {
-        assert_matches!(
-            parse_query_to_logical_ast("-title:toto", false),
-            Err(QueryParserError::AllButQueryForbidden)
+        assert_eq!(
+            parse_query_to_logical_ast("-title:toto", false)
+                .err()
+                .unwrap(),
+            QueryParserError::AllButQueryForbidden
         );
     }
 
@@ -988,18 +965,6 @@ mod test {
         assert!(query_parser.parse_query("with_stop_words:the").is_ok());
     }
 
-    #[test]
-    pub fn test_parse_query_single_negative_term_through_error() {
-        assert_matches!(
-            parse_query_to_logical_ast("-title:toto", true),
-            Err(QueryParserError::AllButQueryForbidden)
-        );
-        assert_matches!(
-            parse_query_to_logical_ast("-title:toto", false),
-            Err(QueryParserError::AllButQueryForbidden)
-        );
-    }
-
     #[test]
     pub fn test_parse_query_to_ast_conjunction() {
         test_parse_query_to_logical_ast_helper(
@@ -1019,6 +984,12 @@ mod test {
              Term(field=1,bytes=[116, 105, 116, 105])))",
             true,
         );
+        assert_eq!(
+            parse_query_to_logical_ast("-title:toto", true)
+                .err()
+                .unwrap(),
+            QueryParserError::AllButQueryForbidden
+        );
         test_parse_query_to_logical_ast_helper(
             "title:a b",
             "(+Term(field=0,bytes=[97]) \
@@ -1042,26 +1013,4 @@ mod test {
             false
         );
     }
-
-    #[test]
-    fn test_and_default_regardless_of_default_conjunctive() {
-        for &default_conjunction in &[false, true] {
-            test_parse_query_to_logical_ast_helper(
-                "title:a AND title:b",
-                "(+Term(field=0,bytes=[97]) +Term(field=0,bytes=[98]))",
-                default_conjunction,
-            );
-        }
-    }
-
-    #[test]
-    fn test_or_default_conjunctive() {
-        for &default_conjunction in &[false, true] {
-            test_parse_query_to_logical_ast_helper(
-                "title:a OR title:b",
-                "(Term(field=0,bytes=[97]) Term(field=0,bytes=[98]))",
-                default_conjunction,
-            );
-        }
-    }
 }

src/query/term_query/block_max_term_scorer.rs

@@ -1,98 +0,0 @@
-use crate::docset::{DocSet, SkipResult};
-use crate::query::{Explanation, Scorer};
-use crate::DocId;
-use crate::Score;
-
-use crate::fieldnorm::FieldNormReader;
-use crate::postings::Postings;
-use crate::postings::{BlockMaxPostings, BlockMaxSegmentPostings};
-use crate::query::bm25::BM25Weight;
-use crate::query::BlockMaxScorer;
-
-pub struct BlockMaxTermScorer {
-    postings: BlockMaxSegmentPostings,
-    fieldnorm_reader: FieldNormReader,
-    similarity_weight: BM25Weight,
-}
-
-impl BlockMaxTermScorer {
-    pub fn new(
-        postings: BlockMaxSegmentPostings,
-        fieldnorm_reader: FieldNormReader,
-        similarity_weight: BM25Weight,
-    ) -> Self {
-        Self {
-            postings,
-            fieldnorm_reader,
-            similarity_weight,
-        }
-    }
-}
-
-impl BlockMaxTermScorer {
-    fn _score(&self, fieldnorm_id: u8, term_freq: u32) -> Score {
-        self.similarity_weight.score(fieldnorm_id, term_freq)
-    }
-
-    pub fn term_freq(&self) -> u32 {
-        self.postings.term_freq()
-    }
-
-    pub fn fieldnorm_id(&self) -> u8 {
-        self.fieldnorm_reader.fieldnorm_id(self.doc())
-    }
-
-    pub fn explain(&self) -> Explanation {
-        let fieldnorm_id = self.fieldnorm_id();
-        let term_freq = self.term_freq();
-        self.similarity_weight.explain(fieldnorm_id, term_freq)
-    }
-}
-
-impl DocSet for BlockMaxTermScorer {
-    fn advance(&mut self) -> bool {
-        self.postings.advance()
-    }
-
-    fn skip_next(&mut self, target: DocId) -> SkipResult {
-        self.postings.skip_next(target)
-    }
-
-    fn doc(&self) -> DocId {
-        self.postings.doc()
-    }
-
-    fn size_hint(&self) -> u32 {
-        self.postings.size_hint()
-    }
-}
-
-impl Scorer for BlockMaxTermScorer {
-    fn score(&mut self) -> Score {
-        self._score(
-            self.fieldnorm_reader.fieldnorm_id(self.doc()),
-            self.postings.term_freq(),
-        )
-    }
-}
-
-impl BlockMaxScorer for BlockMaxTermScorer {
-    fn block_max_score(&mut self) -> Score {
-        self._score(
-            self.fieldnorm_reader
-                .fieldnorm_id(self.postings.block_max_doc()),
-            self.postings.term_freq(),
-        )
-    }
-
-    fn block_max_doc(&mut self) -> DocId {
-        self.postings.block_max_doc()
-    }
-
-    fn max_score(&self) -> Score {
-        self._score(
-            self.fieldnorm_reader.fieldnorm_id(self.postings.max_doc()),
-            self.postings.max_term_freq(),
-        )
-    }
-}

src/query/term_query/mod.rs

@@ -1,9 +1,7 @@
-mod block_max_term_scorer;
 mod term_query;
 mod term_scorer;
 mod term_weight;
 
-pub use self::block_max_term_scorer::BlockMaxTermScorer;
 pub use self::term_query::TermQuery;
 pub use self::term_scorer::TermScorer;
 pub use self::term_weight::TermWeight;

src/query/union.rs

@@ -1,7 +1,7 @@
 use crate::common::TinySet;
 use crate::docset::{DocSet, SkipResult};
 use crate::query::score_combiner::{DoNothingCombiner, ScoreCombiner};
-use crate::query::{Scorer, BlockMaxScorer};
+use crate::query::Scorer;
 use crate::DocId;
 use crate::Score;
 use std::cmp::Ordering;
@@ -9,99 +9,6 @@ use std::cmp::Ordering;
 const HORIZON_NUM_TINYBITSETS: usize = 64;
 const HORIZON: u32 = 64u32 * HORIZON_NUM_TINYBITSETS as u32;
 
-
-#[derive(Debug, Copy, Clone, PartialEq, Eq)]
-struct Pivot {
-    position: usize,
-    first_occurrence: usize,
-    doc: DocId,
-}
-
-/// Find the position in the sorted list of posting lists of the **pivot**.
-///
-/// docsets need to be advanced, and are required to be sorted by the doc they point to.
-///
-/// The pivot is then defined as the lowest DocId that has a chance of matching our condition.
-fn find_pivot_position<'a, TScorer>(
-    mut docsets: impl Iterator<Item = &'a TScorer>,
-    lower_bound_score: Score,
-) -> Option<Pivot>
-    where TScorer: BlockMaxScorer
-{
-    let mut position = 0;
-    let mut upper_bound = Score::default();
-    while let Some(docset) = docsets.next() {
-        upper_bound += docset.max_score();
-        if lower_bound_score < upper_bound {
-            let pivot_doc = docset.doc();
-            let first_occurrence = position;
-            while let Some(docset) = docsets.next() {
-                if docset.doc() != pivot_doc {
-                    break;
-                } else {
-                    position += 1;
-                }
-            }
-            return Some(Pivot {
-                position,
-                doc: pivot_doc,
-                first_occurrence,
-            });
-        }
-        position += 1;
-    }
-    None
-}
-
-/// Sifts down the first element of the slice.
-///
-/// `docsets[1..]` are assumed sorted.
-/// This function swaps `docsets[0]` with its right
-/// neighbor successively -bubble sort style- until it reaches the first
-/// position such that `docsets` is sorted.
-fn sift_down<TScorer>(docsets: &mut [TScorer])
-    where
-        TScorer: BlockMaxScorer + Scorer,
-{
-    for idx in 1..docsets.len() {
-        if docsets[idx].doc() >= docsets[idx - 1].doc() {
-            return;
-        }
-        docsets.swap(idx, idx - 1);
-    }
-}
-
-
-/// Given an iterator over all ordered lists up to the pivot (inclusive) and the following list (if
-/// exists), it returns the next document ID that can be possibly relevant, based on the block max
-/// scores.
-fn find_next_relevant_doc<TScorer>(
-    docsets_up_to_pivot: &mut [TScorer],
-    pivot_docset: &mut TScorer,
-    docset_after_pivot: Option<&mut TScorer>,
-) -> DocId
-    where
-        TScorer: BlockMaxScorer + Scorer,
-{
-    let mut next_doc = 1 + docsets_up_to_pivot
-        .iter_mut()
-        .map(|docset| docset.block_max_doc())
-        .chain(std::iter::once(pivot_docset.block_max_doc()))
-        .min()
-        .unwrap();
-    if let Some(docset) = docset_after_pivot {
-        let doc = docset.doc();
-        if doc < next_doc {
-            next_doc = doc;
-        }
-    }
-    if next_doc <= pivot_docset.doc() {
-        pivot_docset.doc() + 1
-    } else {
-        next_doc
-    }
-}
-
 // `drain_filter` is not stable yet.
 // This function is similar except that it does is not unstable, and
 // it does not keep the original vector ordering.
@@ -132,7 +39,6 @@ pub struct Union<TScorer, TScoreCombiner = DoNothingCombiner> {
     score: Score,
 }
 
-
 impl<TScorer, TScoreCombiner> From<Vec<TScorer>> for Union<TScorer, TScoreCombiner>
 where
     TScoreCombiner: ScoreCombiner,
@@ -220,76 +126,6 @@ impl<TScorer: Scorer, TScoreCombiner: ScoreCombiner> Union<TScorer, TScoreCombin
         }
         false
     }
-
-
-}
-
-
-impl<TScorer: BlockMaxScorer, TScoreCombiner: ScoreCombiner> Union<TScorer, TScoreCombiner> {
-    fn advance_with_pivot(&mut self, pivot: Pivot, lower_bound_score: Score) -> SkipResult {
-        let block_upper_bound: Score = self.docsets[..=pivot.position]
-            .iter_mut()
-            .map(|docset| docset.block_max_score())
-            .sum();
-        if block_upper_bound > lower_bound_score {
-            if pivot.doc == self.docsets[0].doc() {
-                // Since self.docsets is sorted by their current doc, in this branch, all
-                // docsets in [0..=pivot] are positioned on pivot.doc.
-                //
-                // Lets compute the actual score for this doc.
-                //
-                // NOTE(elshize): One additional check needs to be done to improve performance:
-                // update block-wise bound while accumulating score with the actual score,
-                // and check each time if still above threshold.
-                let mut combiner = TScoreCombiner::default();
-                for idx in (0..=pivot.position).rev() {
-                    combiner.update(&mut self.docsets[idx]);
-                    if !self.docsets[idx].advance() {
-                        self.docsets.swap_remove(idx);
-                    }
-                }
-                self.score = combiner.score();
-                self.doc = pivot.doc;
-                self.docsets.sort_by_key(TScorer::doc);
-                SkipResult::Reached
-            } else {
-                // The substraction does not underflow because otherwise we would go to the other
-                // branch.
-                //
-                // `advanced_idx` is the last idx that is not positionned on the pivot yet.
-                let advanced_idx = pivot.first_occurrence - 1;
-                if !self.docsets[advanced_idx].advance() {
-                    self.docsets.swap_remove(advanced_idx);
-                }
-                if self.docsets.is_empty() {
-                    return SkipResult::End;
-                }
-                sift_down(&mut self.docsets[advanced_idx..]);
-                SkipResult::OverStep
-            }
-        } else {
-            let (up_to_pivot, pivot_and_rest) = self.docsets.split_at_mut(pivot.position as usize);
-            let (pivot, after_pivot) = pivot_and_rest.split_first_mut().unwrap();
-            let next_doc = find_next_relevant_doc(up_to_pivot, pivot, after_pivot.first_mut());
-            // NOTE(elshize): It might be more efficient to advance the list with the higher
-            // max score, but let's advance the first one for now for simplicity.
-            if self.docsets[0].skip_next(next_doc) == SkipResult::End {
-                self.docsets.swap_remove(0);
-            }
-            if self.docsets.is_empty() {
-                return SkipResult::End;
-            }
-            sift_down(&mut self.docsets[..]);
-            SkipResult::OverStep
-        }
-    }
-
-    /// Find the position in the sorted list of posting lists of the **pivot**.
-    fn find_pivot_position(&self, lower_bound_score: Score) -> Option<Pivot> {
-        find_pivot_position(
-            self.docsets.iter().map(|docset| docset),
-            lower_bound_score)
-    }
 }
 
 impl<TScorer, TScoreCombiner> DocSet for Union<TScorer, TScoreCombiner>

src/query/weight.rs

@@ -1,26 +1,7 @@
 use super::Scorer;
 use crate::core::SegmentReader;
 use crate::query::Explanation;
-use crate::{DocId, Score};
-
-pub trait PruningScorer {
-
-    fn doc(&self) -> DocId;
-
-    fn score(&self) -> Score;
-
-    /// Advance to the next document that has a score strictly greater than
-    /// `lower_bound_score`.
-    fn advance_with_pruning(&mut self, score_lower_bound: f32) -> bool;
-    fn advance(&mut self) -> bool {
-        self.advance_with_pruning(std::f32::NEG_INFINITY)
-    }
-}
-
-pub enum PruningScorerIfPossible {
-    Pruning(Box<dyn PruningScorer>),
-    NonPruning(Box<dyn Scorer>)
-}
+use crate::DocId;
 
 /// A Weight is the specialization of a Query
 /// for a given set of segments.
@@ -34,11 +15,6 @@ pub trait Weight: Send + Sync + 'static {
     /// See [`Query`](./trait.Query.html).
     fn scorer(&self, reader: &SegmentReader, boost: f32) -> crate::Result<Box<dyn Scorer>>;
 
-    fn pruning_scorer(&self, reader: &SegmentReader, boost: f32) -> crate::Result<PruningScorerIfPossible> {
-        let scorer = self.scorer(reader, boost)?;
-        Ok(PruningScorerIfPossible::NonPruning(Box::new(scorer)))
-    }
-
     /// Returns an `Explanation` for the given document.
     fn explain(&self, reader: &SegmentReader, doc: DocId) -> crate::Result<Explanation>;
 

src/reader/index_writer_reader.rs

@@ -0,0 +1,84 @@
+use crate::directory::{WatchCallbackList, WatchHandle};
+use crate::indexer::SegmentRegisters;
+use crate::reader::pool::Pool;
+use crate::{Index, LeasedItem, Searcher, Segment, SegmentReader};
+use std::iter::repeat_with;
+use std::sync::{Arc, RwLock, Weak};
+
+struct InnerNRTReader {
+    num_searchers: usize,
+    index: Index,
+    searcher_pool: Pool<Searcher>,
+    segment_registers: Arc<RwLock<SegmentRegisters>>,
+}
+
+impl InnerNRTReader {
+    fn load_segment_readers(&self) -> crate::Result<Vec<SegmentReader>> {
+        let segments: Vec<Segment> = self
+            .segment_registers
+            .read()
+            .expect("lock should never be polluted. Please report.")
+            .committed_segment();
+        segments
+            .iter()
+            .map(SegmentReader::open)
+            .collect::<crate::Result<Vec<SegmentReader>>>()
+    }
+
+    pub fn reload(&self) -> crate::Result<()> {
+        let segment_readers: Vec<SegmentReader> = self.load_segment_readers()?;
+        let schema = self.index.schema();
+        let searchers = repeat_with(|| {
+            Searcher::new(schema.clone(), self.index.clone(), segment_readers.clone())
+        })
+        .take(self.num_searchers)
+        .collect();
+        self.searcher_pool.publish_new_generation(searchers);
+        Ok(())
+    }
+
+    pub fn searcher(&self) -> LeasedItem<Searcher> {
+        self.searcher_pool.acquire()
+    }
+}
+
+#[derive(Clone)]
+pub struct NRTReader {
+    inner: Arc<InnerNRTReader>,
+    watch_handle: WatchHandle,
+}
+
+impl NRTReader {
+    pub fn reload(&self) -> crate::Result<()> {
+        self.inner.reload()
+    }
+
+    pub fn searcher(&self) -> LeasedItem<Searcher> {
+        self.inner.searcher()
+    }
+
+    pub(crate) fn create(
+        num_searchers: usize,
+        index: Index,
+        segment_registers: Arc<RwLock<SegmentRegisters>>,
+        watch_callback_list: &WatchCallbackList,
+    ) -> crate::Result<Self> {
+        let inner_reader: Arc<InnerNRTReader> = Arc::new(InnerNRTReader {
+            num_searchers,
+            index,
+            searcher_pool: Pool::new(),
+            segment_registers,
+        });
+        let inner_reader_weak: Weak<InnerNRTReader> = Arc::downgrade(&inner_reader);
+        let watch_handle = watch_callback_list.subscribe(Box::new(move || {
+            if let Some(nrt_reader_arc) = inner_reader_weak.upgrade() {
+                let _ = nrt_reader_arc.reload();
+            }
+        }));
+        inner_reader.reload()?;
+        Ok(NRTReader {
+            inner: inner_reader,
+            watch_handle,
+        })
+    }
+}

src/reader/meta_file_reader.rs

@@ -0,0 +1,177 @@
+use super::pool::Pool;
+use crate::core::Segment;
+use crate::directory::Directory;
+use crate::directory::WatchHandle;
+use crate::directory::META_LOCK;
+use crate::Searcher;
+use crate::SegmentReader;
+use crate::{Index, LeasedItem};
+use crate::{IndexReader, Result};
+use std::iter::repeat_with;
+use std::sync::Arc;
+
+/// Defines when a new version of the index should be reloaded.
+///
+/// Regardless of whether you search and index in the same process, tantivy does not necessarily
+/// reflects the change that are commited to your index. `ReloadPolicy` precisely helps you define
+/// when you want your index to be reloaded.
+#[derive(Clone, Copy)]
+pub enum ReloadPolicy {
+    /// The index is entirely reloaded manually.
+    /// All updates of the index should be manual.
+    ///
+    /// No change is reflected automatically. You are required to call `.load_seacher()` manually.
+    Manual,
+    /// The index is reloaded within milliseconds after a new commit is available.
+    /// This is made possible by watching changes in the `meta.json` file.
+    OnCommit, // TODO add NEAR_REAL_TIME(target_ms)
+}
+
+/// `IndexReader` builder
+///
+/// It makes it possible to set the following values.
+///
+/// - `num_searchers` (by default, the number of detected CPU threads):
+///
+///   When `num_searchers` queries are requested at the same time, the `num_searchers` will block
+///   until the one of the searcher in-use gets released.
+/// - `reload_policy` (by default `ReloadPolicy::OnCommit`):
+///
+///   See [`ReloadPolicy`](./enum.ReloadPolicy.html) for more details.
+#[derive(Clone)]
+pub struct IndexReaderBuilder {
+    num_searchers: usize,
+    reload_policy: ReloadPolicy,
+    index: Index,
+}
+
+impl IndexReaderBuilder {
+    pub(crate) fn new(index: Index) -> IndexReaderBuilder {
+        IndexReaderBuilder {
+            num_searchers: num_cpus::get(),
+            reload_policy: ReloadPolicy::Manual,
+            index,
+        }
+    }
+
+    /// Sets the reload_policy.
+    ///
+    /// See [`ReloadPolicy`](./enum.ReloadPolicy.html) for more details.
+    pub fn reload_policy(mut self, reload_policy: ReloadPolicy) -> IndexReaderBuilder {
+        self.reload_policy = reload_policy;
+        self
+    }
+
+    /// Sets the number of `Searcher` in the searcher pool.
+    pub fn num_searchers(mut self, num_searchers: usize) -> IndexReaderBuilder {
+        self.num_searchers = num_searchers;
+        self
+    }
+
+    /// Building the reader is a non-trivial operation that requires
+    /// to open different segment readers. It may take hundreds of milliseconds
+    /// of time and it may return an error.
+    pub fn try_into(self) -> crate::Result<IndexReader> {
+        let inner_reader = MetaFileIndexReaderInner {
+            index: self.index,
+            num_searchers: self.num_searchers,
+            searcher_pool: Pool::new(),
+        };
+        inner_reader.reload()?;
+        let inner_reader_arc = Arc::new(inner_reader);
+        let watch_handle_opt: Option<WatchHandle>;
+        match self.reload_policy {
+            ReloadPolicy::Manual => {
+                // No need to set anything...
+                watch_handle_opt = None;
+            }
+            ReloadPolicy::OnCommit => {
+                let inner_reader_arc_clone = inner_reader_arc.clone();
+                let callback = move || {
+                    if let Err(err) = inner_reader_arc_clone.reload() {
+                        error!(
+                            "Error while loading searcher after commit was detected. {:?}",
+                            err
+                        );
+                    }
+                };
+                let watch_handle = inner_reader_arc
+                    .index
+                    .directory()
+                    .watch(Box::new(callback))?;
+                watch_handle_opt = Some(watch_handle);
+            }
+        }
+        Ok(IndexReader::from(MetaFileIndexReader {
+            inner: inner_reader_arc,
+            watch_handle_opt,
+        }))
+    }
+}
+
+struct MetaFileIndexReaderInner {
+    num_searchers: usize,
+    searcher_pool: Pool<Searcher>,
+    index: Index,
+}
+
+impl MetaFileIndexReaderInner {
+    fn load_segment_readers(&self) -> crate::Result<Vec<SegmentReader>> {
+        // We keep the lock until we have effectively finished opening the
+        // the `SegmentReader` because it prevents a diffferent process
+        // to garbage collect these file while we open them.
+        //
+        // Once opened, on linux & mac, the mmap will remain valid after
+        // the file has been deleted
+        // On windows, the file deletion will fail.
+        let _meta_lock = self.index.directory().acquire_lock(&META_LOCK)?;
+        let searchable_segments = self.searchable_segments()?;
+        searchable_segments
+            .iter()
+            .map(SegmentReader::open)
+            .collect::<Result<_>>()
+    }
+
+    fn reload(&self) -> crate::Result<()> {
+        let segment_readers: Vec<SegmentReader> = self.load_segment_readers()?;
+        let schema = self.index.schema();
+        let searchers = repeat_with(|| {
+            Searcher::new(schema.clone(), self.index.clone(), segment_readers.clone())
+        })
+        .take(self.num_searchers)
+        .collect();
+        self.searcher_pool.publish_new_generation(searchers);
+        Ok(())
+    }
+
+    /// Returns the list of segments that are searchable
+    fn searchable_segments(&self) -> crate::Result<Vec<Segment>> {
+        self.index.searchable_segments()
+    }
+
+    fn searcher(&self) -> LeasedItem<Searcher> {
+        self.searcher_pool.acquire()
+    }
+}
+
+/// `IndexReader` is your entry point to read and search the index.
+///
+/// It controls when a new version of the index should be loaded and lends
+/// you instances of `Searcher` for the last loaded version.
+///
+/// `Clone` does not clone the different pool of searcher. `IndexReader`
+/// just wraps and `Arc`.
+#[derive(Clone)]
+pub struct MetaFileIndexReader {
+    inner: Arc<MetaFileIndexReaderInner>,
+    watch_handle_opt: Option<WatchHandle>,
+}
+
+impl MetaFileIndexReader {
+    pub fn reload(&self) -> crate::Result<()> {
+        self.inner.reload()
+    }
+    pub fn searcher(&self) -> LeasedItem<Searcher> {
+        self.inner.searcher()
+    }
+}

src/reader/mod.rs

@@ -1,159 +1,14 @@
+mod index_writer_reader;
+mod meta_file_reader;
 mod pool;
 
+use self::meta_file_reader::MetaFileIndexReader;
+pub use self::meta_file_reader::{IndexReaderBuilder, ReloadPolicy};
 pub use self::pool::LeasedItem;
-use self::pool::Pool;
-use crate::core::Segment;
-use crate::directory::Directory;
-use crate::directory::WatchHandle;
-use crate::directory::META_LOCK;
-use crate::Index;
+
+pub(crate) use crate::reader::index_writer_reader::NRTReader;
+
 use crate::Searcher;
-use crate::SegmentReader;
-use std::convert::TryInto;
-use std::sync::Arc;
-
-/// Defines when a new version of the index should be reloaded.
-///
-/// Regardless of whether you search and index in the same process, tantivy does not necessarily
-/// reflects the change that are commited to your index. `ReloadPolicy` precisely helps you define
-/// when you want your index to be reloaded.
-#[derive(Clone, Copy)]
-pub enum ReloadPolicy {
-    /// The index is entirely reloaded manually.
-    /// All updates of the index should be manual.
-    ///
-    /// No change is reflected automatically. You are required to call `.load_seacher()` manually.
-    Manual,
-    /// The index is reloaded within milliseconds after a new commit is available.
-    /// This is made possible by watching changes in the `meta.json` file.
-    OnCommit, // TODO add NEAR_REAL_TIME(target_ms)
-}
-
-/// `IndexReader` builder
-///
-/// It makes it possible to set the following values.
-///
-/// - `num_searchers` (by default, the number of detected CPU threads):
-///
-///   When `num_searchers` queries are requested at the same time, the `num_searchers` will block
-///   until the one of the searcher in-use gets released.
-/// - `reload_policy` (by default `ReloadPolicy::OnCommit`):
-///
-///   See [`ReloadPolicy`](./enum.ReloadPolicy.html) for more details.
-#[derive(Clone)]
-pub struct IndexReaderBuilder {
-    num_searchers: usize,
-    reload_policy: ReloadPolicy,
-    index: Index,
-}
-
-impl IndexReaderBuilder {
-    pub(crate) fn new(index: Index) -> IndexReaderBuilder {
-        IndexReaderBuilder {
-            num_searchers: num_cpus::get(),
-            reload_policy: ReloadPolicy::OnCommit,
-            index,
-        }
-    }
-
-    /// Builds the reader.
-    ///
-    /// Building the reader is a non-trivial operation that requires
-    /// to open different segment readers. It may take hundreds of milliseconds
-    /// of time and it may return an error.
-    pub fn try_into(self) -> crate::Result<IndexReader> {
-        let inner_reader = InnerIndexReader {
-            index: self.index,
-            num_searchers: self.num_searchers,
-            searcher_pool: Pool::new(),
-        };
-        inner_reader.reload()?;
-        let inner_reader_arc = Arc::new(inner_reader);
-        let watch_handle_opt: Option<WatchHandle>;
-        match self.reload_policy {
-            ReloadPolicy::Manual => {
-                // No need to set anything...
-                watch_handle_opt = None;
-            }
-            ReloadPolicy::OnCommit => {
-                let inner_reader_arc_clone = inner_reader_arc.clone();
-                let callback = move || {
-                    if let Err(err) = inner_reader_arc_clone.reload() {
-                        error!(
-                            "Error while loading searcher after commit was detected. {:?}",
-                            err
-                        );
-                    }
-                };
-                let watch_handle = inner_reader_arc
-                    .index
-                    .directory()
-                    .watch(Box::new(callback))?;
-                watch_handle_opt = Some(watch_handle);
-            }
-        }
-        Ok(IndexReader {
-            inner: inner_reader_arc,
-            watch_handle_opt,
-        })
-    }
-
-    /// Sets the reload_policy.
-    ///
-    /// See [`ReloadPolicy`](./enum.ReloadPolicy.html) for more details.
-    pub fn reload_policy(mut self, reload_policy: ReloadPolicy) -> IndexReaderBuilder {
-        self.reload_policy = reload_policy;
-        self
-    }
-
-    /// Sets the number of `Searcher` in the searcher pool.
-    pub fn num_searchers(mut self, num_searchers: usize) -> IndexReaderBuilder {
-        self.num_searchers = num_searchers;
-        self
-    }
-}
-
-impl TryInto<IndexReader> for IndexReaderBuilder {
-    type Error = crate::TantivyError;
-
-    fn try_into(self) -> crate::Result<IndexReader> {
-        IndexReaderBuilder::try_into(self)
-    }
-}
-
-struct InnerIndexReader {
-    num_searchers: usize,
-    searcher_pool: Pool<Searcher>,
-    index: Index,
-}
-
-impl InnerIndexReader {
-    fn reload(&self) -> crate::Result<()> {
-        let segment_readers: Vec<SegmentReader> = {
-            let _meta_lock = self.index.directory().acquire_lock(&META_LOCK)?;
-            let searchable_segments = self.searchable_segments()?;
-            searchable_segments
-                .iter()
-                .map(SegmentReader::open)
-                .collect::<crate::Result<_>>()?
-        };
-        let schema = self.index.schema();
-        let searchers = (0..self.num_searchers)
-            .map(|_| Searcher::new(schema.clone(), self.index.clone(), segment_readers.clone()))
-            .collect();
-        self.searcher_pool.publish_new_generation(searchers);
-        Ok(())
-    }
-
-    /// Returns the list of segments that are searchable
-    fn searchable_segments(&self) -> crate::Result<Vec<Segment>> {
-        self.index.searchable_segments()
-    }
-
-    fn searcher(&self) -> LeasedItem<Searcher> {
-        self.searcher_pool.acquire()
-    }
-}
 
 /// `IndexReader` is your entry point to read and search the index.
 ///
@@ -163,17 +18,12 @@ impl InnerIndexReader {
 /// `Clone` does not clone the different pool of searcher. `IndexReader`
 /// just wraps and `Arc`.
 #[derive(Clone)]
-pub struct IndexReader {
-    inner: Arc<InnerIndexReader>,
-    watch_handle_opt: Option<WatchHandle>,
+pub enum IndexReader {
+    FromMetaFile(MetaFileIndexReader),
+    NRT(NRTReader),
 }
 
 impl IndexReader {
-    #[cfg(test)]
-    pub(crate) fn index(&self) -> Index {
-        self.inner.index.clone()
-    }
-
     /// Update searchers so that they reflect the state of the last
     /// `.commit()`.
     ///
@@ -184,7 +34,10 @@ impl IndexReader {
     /// This automatic reload can take 10s of milliseconds to kick in however, and in unit tests
     /// it can be nice to deterministically force the reload of searchers.
     pub fn reload(&self) -> crate::Result<()> {
-        self.inner.reload()
+        match self {
+            IndexReader::FromMetaFile(meta_file_reader) => meta_file_reader.reload(),
+            IndexReader::NRT(nrt_reader) => nrt_reader.reload(),
+        }
     }
 
     /// Returns a searcher
@@ -198,6 +51,21 @@ impl IndexReader {
     /// The same searcher must be used for a given query, as it ensures
     /// the use of a consistent segment set.
     pub fn searcher(&self) -> LeasedItem<Searcher> {
-        self.inner.searcher()
+        match self {
+            IndexReader::FromMetaFile(meta_file_reader) => meta_file_reader.searcher(),
+            IndexReader::NRT(nrt_reader) => nrt_reader.searcher(),
+        }
+    }
+}
+
+impl From<MetaFileIndexReader> for IndexReader {
+    fn from(meta_file_reader: MetaFileIndexReader) -> Self {
+        IndexReader::FromMetaFile(meta_file_reader)
+    }
+}
+
+impl From<NRTReader> for IndexReader {
+    fn from(nrt_reader: NRTReader) -> Self {
+        IndexReader::NRT(nrt_reader)
     }
 }

src/schema/document.rs

@@ -3,6 +3,7 @@ use crate::common::BinarySerializable;
 use crate::common::VInt;
 use crate::tokenizer::PreTokenizedString;
 use crate::DateTime;
+use serde;
 use std::io::{self, Read, Write};
 use std::mem;
 

src/schema/field.rs

@@ -1,4 +1,5 @@
 use crate::common::BinarySerializable;
+use serde;
 use std::io;
 use std::io::Read;
 use std::io::Write;

src/schema/field_value.rs

@@ -1,7 +1,10 @@
 use crate::common::BinarySerializable;
 use crate::schema::Field;
 use crate::schema::Value;
-use std::io::{self, Read, Write};
+use serde;
+use std::io;
+use std::io::Read;
+use std::io::Write;
 
 /// `FieldValue` holds together a `Field` and its `Value`.
 #[derive(Debug, Clone, Ord, PartialEq, Eq, PartialOrd, serde::Serialize, serde::Deserialize)]

src/schema/text_options.rs

@@ -156,17 +156,30 @@ mod tests {
 
     #[test]
     fn test_field_options() {
-        let field_options = STORED | TEXT;
-        assert!(field_options.is_stored());
-        assert!(field_options.get_indexing_options().is_some());
-        let mut schema_builder = Schema::builder();
-        schema_builder.add_text_field("body", TEXT);
-        let schema = schema_builder.build();
-        let field = schema.get_field("body").unwrap();
-        let field_entry = schema.get_field_entry(field);
-        assert!(matches!(field_entry.field_type(),
-                &FieldType::Str(ref text_options)
-                if text_options.get_indexing_options().unwrap().tokenizer() == "default"));
+        {
+            let field_options = STORED | TEXT;
+            assert!(field_options.is_stored());
+            assert!(field_options.get_indexing_options().is_some());
+        }
+        {
+            let mut schema_builder = Schema::builder();
+            schema_builder.add_text_field("body", TEXT);
+            let schema = schema_builder.build();
+            let field = schema.get_field("body").unwrap();
+            let field_entry = schema.get_field_entry(field);
+            match field_entry.field_type() {
+                &FieldType::Str(ref text_options) => {
+                    assert!(text_options.get_indexing_options().is_some());
+                    assert_eq!(
+                        text_options.get_indexing_options().unwrap().tokenizer(),
+                        "default"
+                    );
+                }
+                _ => {
+                    panic!("");
+                }
+            }
+        }
     }
 
     #[test]

src/store/compression_snap.rs

@@ -1,3 +1,5 @@
+use snap;
+
 use std::io::{self, Read, Write};
 
 /// Name of the compression scheme used in the doc store.

src/store/writer.rs

@@ -3,8 +3,6 @@ use super::skiplist::SkipListBuilder;
 use super::StoreReader;
 use crate::common::CountingWriter;
 use crate::common::{BinarySerializable, VInt};
-use crate::directory::TerminatingWrite;
-use crate::directory::WritePtr;
 use crate::schema::Document;
 use crate::DocId;
 use std::io::{self, Write};
@@ -19,20 +17,20 @@ const BLOCK_SIZE: usize = 16_384;
 ///
 /// The skip list index on the other hand, is built in memory.
 ///
-pub struct StoreWriter {
+pub struct StoreWriter<W: io::Write> {
     doc: DocId,
     offset_index_writer: SkipListBuilder<u64>,
-    writer: CountingWriter<WritePtr>,
+    writer: CountingWriter<W>,
     intermediary_buffer: Vec<u8>,
     current_block: Vec<u8>,
 }
 
-impl StoreWriter {
+impl<W: io::Write> StoreWriter<W> {
     /// Create a store writer.
     ///
     /// The store writer will writes blocks on disc as
     /// document are added.
-    pub fn new(writer: WritePtr) -> StoreWriter {
+    pub fn new(writer: W) -> StoreWriter<W> {
         StoreWriter {
             doc: 0,
             offset_index_writer: SkipListBuilder::new(4),
@@ -102,7 +100,9 @@ impl StoreWriter {
     ///
     /// Compress the last unfinished block if any,
     /// and serializes the skip list index on disc.
-    pub fn close(mut self) -> io::Result<()> {
+    ///
+    /// The returned writer is not flushed.
+    pub fn close(mut self) -> io::Result<W> {
         if !self.current_block.is_empty() {
             self.write_and_compress_block()?;
         }
@@ -110,6 +110,7 @@ impl StoreWriter {
         self.offset_index_writer.write(&mut self.writer)?;
         header_offset.serialize(&mut self.writer)?;
         self.doc.serialize(&mut self.writer)?;
-        self.writer.terminate()
+        let (wrt, _) = self.writer.finish()?;
+        Ok(wrt)
     }
 }

src/termdict/mod.rs

@@ -434,7 +434,6 @@ mod tests {
 
     #[test]
     fn test_automaton_search() {
-        use crate::query::DFAWrapper;
         use levenshtein_automata::LevenshteinAutomatonBuilder;
 
         const COUNTRIES: [&'static str; 7] = [
@@ -464,7 +463,7 @@ mod tests {
 
         // We can now build an entire dfa.
         let lev_automaton_builder = LevenshteinAutomatonBuilder::new(2, true);
-        let automaton = DFAWrapper(lev_automaton_builder.build_dfa("Spaen"));
+        let automaton = lev_automaton_builder.build_dfa("Spaen");
 
         let mut range = term_dict.search(automaton).into_stream();
 

src/termdict/termdict.rs

@@ -7,6 +7,7 @@ use crate::postings::TermInfo;
 use crate::termdict::TermOrdinal;
 use once_cell::sync::Lazy;
 use std::io::{self, Write};
+use tantivy_fst;
 use tantivy_fst::raw::Fst;
 use tantivy_fst::Automaton;