add API contract verfication

Author: PSeitz

src/docset.rs

@@ -40,6 +40,8 @@ pub trait DocSet: Send {
     /// of `DocSet` should support it.
     ///
     /// Calling `seek(TERMINATED)` is also legal and is the normal way to consume a `DocSet`.
+    ///
+    /// `target` has to be larger or equal to `.doc()` when calling `seek`.
     fn seek(&mut self, target: DocId) -> DocId {
         let mut doc = self.doc();
         debug_assert!(doc <= target);
@@ -58,11 +60,19 @@ pub trait DocSet: Send {
     ///
     /// ## API Behaviour
     /// If `seek_exact` is returning true, a call to `doc()` has to return target.
-    /// If `seek_exact` is returning false, a call to `doc()` may return the previous doc,
-    /// which may be lower than target.
+    /// If `seek_exact` is returning false, a call to `doc()` may return any doc and should not be
+    /// used until `seek_exact` returns true again.
+    ///
+    /// Consecutive calls are not allowed to have decreasing `target` values.
+    ///
+    /// # Warning
+    /// This is an advanced API used by intersection. The API contract is tricky, avoid using it.
     fn seek_exact(&mut self, target: DocId) -> bool {
-        let doc = self.seek(target);
-        doc == target
+        let current_doc = self.doc();
+        if current_doc < target {
+            self.seek(target);
+        }
+        self.doc() == target
     }
 
     /// Fills a given mutable buffer with the next doc ids from the
@@ -103,8 +113,11 @@ pub trait DocSet: Send {
     /// length of the docset.
     fn size_hint(&self) -> u32;
 
-    /// Returns a best-effort hint of the
-    /// cost to drive the docset.
+    /// Returns a best-effort hint of the cost to consume the entire docset.
+    ///
+    /// Consuming means calling advance until [`TERMINATED`] is returned.
+    /// The cost should be relative to the cost of driving a Term query,
+    /// which would be the number of documents in the DocSet.
     ///
     /// By default this returns `size_hint()`.
     ///

src/query/intersection.rs

@@ -12,7 +12,13 @@ use crate::{DocId, Score};
 /// For better performance, the function uses a
 /// specialized implementation if the two
 /// shortest scorers are `TermScorer`s.
-pub fn intersect_scorers(mut scorers: Vec<Box<dyn Scorer>>, num_docs: u32) -> Box<dyn Scorer> {
+///
+/// num_docs_segment is the number of documents in the segment. It is used for estimating the
+/// `size_hint` of the intersection.
+pub fn intersect_scorers(
+    mut scorers: Vec<Box<dyn Scorer>>,
+    num_docs_segment: u32,
+) -> Box<dyn Scorer> {
     if scorers.is_empty() {
         return Box::new(EmptyScorer);
     }
@@ -35,14 +41,14 @@ pub fn intersect_scorers(mut scorers: Vec<Box<dyn Scorer>>, num_docs: u32) -> Bo
             left: *(left.downcast::<TermScorer>().map_err(|_| ()).unwrap()),
             right: *(right.downcast::<TermScorer>().map_err(|_| ()).unwrap()),
             others: scorers,
-            num_docs,
+            num_docs: num_docs_segment,
         });
     }
     Box::new(Intersection {
         left,
         right,
         others: scorers,
-        num_docs,
+        num_docs: num_docs_segment,
     })
 }
 

src/query/range_query/fast_field_range_doc_set.rs

@@ -183,7 +183,14 @@ impl<T: Send + Sync + PartialOrd + Copy + Debug + 'static> DocSet for RangeDocSe
     fn cost(&self) -> u64 {
         // Advancing the docset is pretty expensive since it scans the whole column, there is no
         // index currently (will change with an kd-tree)
-        // Since we use SIMD to scan the fast field range query we lower the cost a little bit.
+        // Since we use SIMD to scan the fast field range query we lower the cost a little bit,
+        // assuming that we hit 10% of the docs like in size_hint.
+        //
+        // If we would return a cost higher than num_docs, we would never choose ff range query as
+        // the driver in a DocSet, when intersecting a term query with a fast field. But
+        // it's the faster choice when the term query has a lot of docids and the range
+        // query has not.
+        //
         // Ideally this would take the fast field codec into account
         (self.column.num_docs() as f64 * 0.8) as u64
     }

src/query/size_hint.rs

@@ -10,7 +10,9 @@
 /// The estimated number of documents in the intersection.
 pub fn estimate_intersection<I>(mut docset_sizes: I, max_docs: u32) -> u32
 where I: Iterator<Item = u32> {
-     if max_doc == 0u32 { return 0u32; }
+    if max_docs == 0u32 {
+        return 0u32;
+    }
     // Terms tend to be not really randomly distributed.
     // This factor is used to adjust the estimate.
     let mut co_loc_factor: f64 = 1.3;