Merge pull request #2559 from quickwit-oss/trinity/sstable-partial-automaton

allow warming partially an sstable for an automaton

Cargo.toml

@@ -67,6 +67,7 @@ tokenizer-api = { version = "0.3", path = "./tokenizer-api", package = "tantivy-
 sketches-ddsketch = { version = "0.3.0", features = ["use_serde"] }
 hyperloglogplus = { version = "0.4.1", features = ["const-loop"] }
 futures-util = { version = "0.3.28", optional = true }
+futures-channel = { version = "0.3.28", optional = true }
 fnv = "1.0.7"
 
 [target.'cfg(windows)'.dependencies]
@@ -121,7 +122,7 @@ zstd-compression = ["zstd"]
 failpoints = ["fail", "fail/failpoints"]
 unstable = []                            # useful for benches.
 
-quickwit = ["sstable", "futures-util"]
+quickwit = ["sstable", "futures-util", "futures-channel"]
 
 # Compares only the hash of a string when indexing data.
 # Increases indexing speed, but may lead to extremely rare missing terms, when there's a hash collision.

src/index/inverted_index_reader.rs

@@ -3,6 +3,12 @@ use std::io;
 use common::json_path_writer::JSON_END_OF_PATH;
 use common::BinarySerializable;
 use fnv::FnvHashSet;
+#[cfg(feature = "quickwit")]
+use futures_util::{FutureExt, StreamExt, TryStreamExt};
+#[cfg(feature = "quickwit")]
+use itertools::Itertools;
+#[cfg(feature = "quickwit")]
+use tantivy_fst::automaton::{AlwaysMatch, Automaton};
 
 use crate::directory::FileSlice;
 use crate::positions::PositionReader;
@@ -219,13 +225,18 @@ impl InvertedIndexReader {
         self.termdict.get_async(term.serialized_value_bytes()).await
     }
 
-    async fn get_term_range_async(
+    async fn get_term_range_async<'a, A: Automaton + 'a>(
-        &self,
+        &'a self,
         terms: impl std::ops::RangeBounds<Term>,
+        automaton: A,
         limit: Option<u64>,
-    ) -> io::Result<impl Iterator<Item = TermInfo> + '_> {
+        merge_holes_under_bytes: usize,
+    ) -> io::Result<impl Iterator<Item = TermInfo> + 'a>
+    where
+        A::State: Clone,
+    {
         use std::ops::Bound;
-        let range_builder = self.termdict.range();
+        let range_builder = self.termdict.search(automaton);
         let range_builder = match terms.start_bound() {
             Bound::Included(bound) => range_builder.ge(bound.serialized_value_bytes()),
             Bound::Excluded(bound) => range_builder.gt(bound.serialized_value_bytes()),
@@ -242,7 +253,9 @@ impl InvertedIndexReader {
             range_builder
         };
 
-        let mut stream = range_builder.into_stream_async().await?;
+        let mut stream = range_builder
+            .into_stream_async_merging_holes(merge_holes_under_bytes)
+            .await?;
 
         let iter = std::iter::from_fn(move || stream.next().map(|(_k, v)| v.clone()));
 
@@ -288,7 +301,9 @@ impl InvertedIndexReader {
         limit: Option<u64>,
         with_positions: bool,
     ) -> io::Result<bool> {
-        let mut term_info = self.get_term_range_async(terms, limit).await?;
+        let mut term_info = self
+            .get_term_range_async(terms, AlwaysMatch, limit, 0)
+            .await?;
 
         let Some(first_terminfo) = term_info.next() else {
             // no key matches, nothing more to load
@@ -315,6 +330,84 @@ impl InvertedIndexReader {
         Ok(true)
     }
 
+    /// Warmup a block postings given a range of `Term`s.
+    /// This method is for an advanced usage only.
+    ///
+    /// returns a boolean, whether a term matching the range was found in the dictionary
+    pub async fn warm_postings_automaton<
+        A: Automaton + Clone + Send + 'static,
+        E: FnOnce(Box<dyn FnOnce() -> io::Result<()> + Send>) -> F,
+        F: std::future::Future<Output = io::Result<()>>,
+    >(
+        &self,
+        automaton: A,
+        // with_positions: bool, at the moment we have no use for it, and supporting it would add
+        // complexity to the coalesce
+        executor: E,
+    ) -> io::Result<bool>
+    where
+        A::State: Clone,
+    {
+        // merge holes under 4MiB, that's how many bytes we can hope to receive during a TTFB from
+        // S3 (~80MiB/s, and 50ms latency)
+        const MERGE_HOLES_UNDER_BYTES: usize = (80 * 1024 * 1024 * 50) / 1000;
+        // we build a first iterator to download everything. Simply calling the function already
+        // download everything we need from the sstable, but doesn't start iterating over it.
+        let _term_info_iter = self
+            .get_term_range_async(.., automaton.clone(), None, MERGE_HOLES_UNDER_BYTES)
+            .await?;
+
+        let (sender, posting_ranges_to_load_stream) = futures_channel::mpsc::unbounded();
+        let termdict = self.termdict.clone();
+        let cpu_bound_task = move || {
+            // then we build a 2nd iterator, this one with no holes, so we don't go through blocks
+            // we can't match.
+            // This makes the assumption there is a caching layer below us, which gives sync read
+            // for free after the initial async access. This might not always be true, but is in
+            // Quickwit.
+            // We build things from this closure otherwise we get into lifetime issues that can only
+            // be solved with self referential strucs. Returning an io::Result from here is a bit
+            // more leaky abstraction-wise, but a lot better than the alternative
+            let mut stream = termdict.search(automaton).into_stream()?;
+
+            // we could do without an iterator, but this allows us access to coalesce which simplify
+            // things
+            let posting_ranges_iter =
+                std::iter::from_fn(move || stream.next().map(|(_k, v)| v.postings_range.clone()));
+
+            let merged_posting_ranges_iter = posting_ranges_iter.coalesce(|range1, range2| {
+                if range1.end + MERGE_HOLES_UNDER_BYTES >= range2.start {
+                    Ok(range1.start..range2.end)
+                } else {
+                    Err((range1, range2))
+                }
+            });
+
+            for posting_range in merged_posting_ranges_iter {
+                if let Err(_) = sender.unbounded_send(posting_range) {
+                    // this should happen only when search is cancelled
+                    return Err(io::Error::other("failed to send posting range back"));
+                }
+            }
+            Ok(())
+        };
+        let task_handle = executor(Box::new(cpu_bound_task));
+
+        let posting_downloader = posting_ranges_to_load_stream
+            .map(|posting_slice| {
+                self.postings_file_slice
+                    .read_bytes_slice_async(posting_slice)
+                    .map(|result| result.map(|_slice| ()))
+            })
+            .buffer_unordered(5)
+            .try_collect::<Vec<()>>();
+
+        let (_, slices_downloaded) =
+            futures_util::future::try_join(task_handle, posting_downloader).await?;
+
+        Ok(!slices_downloaded.is_empty())
+    }
+
     /// Warmup the block postings for all terms.
     /// This method is for an advanced usage only.
     ///

src/termdict/fst_termdict/term_info_store.rs

@@ -93,6 +93,7 @@ impl TermInfoBlockMeta {
     }
 }
 
+#[derive(Clone)]
 pub struct TermInfoStore {
     num_terms: usize,
     block_meta_bytes: OwnedBytes,

src/termdict/fst_termdict/termdict.rs

@@ -1,4 +1,5 @@
 use std::io::{self, Write};
+use std::sync::Arc;
 
 use common::{BinarySerializable, CountingWriter};
 use once_cell::sync::Lazy;
@@ -113,8 +114,9 @@ static EMPTY_TERM_DICT_FILE: Lazy<FileSlice> = Lazy::new(|| {
 /// The `Fst` crate is used to associate terms to their
 /// respective `TermOrdinal`. The `TermInfoStore` then makes it
 /// possible to fetch the associated `TermInfo`.
+#[derive(Clone)]
 pub struct TermDictionary {
-    fst_index: tantivy_fst::Map<OwnedBytes>,
+    fst_index: Arc<tantivy_fst::Map<OwnedBytes>>,
     term_info_store: TermInfoStore,
 }
 
@@ -136,7 +138,7 @@ impl TermDictionary {
         let fst_index = open_fst_index(fst_file_slice)?;
         let term_info_store = TermInfoStore::open(values_file_slice)?;
         Ok(TermDictionary {
-            fst_index,
+            fst_index: Arc::new(fst_index),
             term_info_store,
         })
     }

src/termdict/mod.rs

@@ -74,6 +74,7 @@ const CURRENT_TYPE: DictionaryType = DictionaryType::SSTable;
 
 // TODO in the future this should become an enum of supported dictionaries
 /// A TermDictionary wrapping either an FST based dictionary or a SSTable based one.
+#[derive(Clone)]
 pub struct TermDictionary(InnerTermDict);
 
 impl TermDictionary {

src/termdict/sstable_termdict/mod.rs

@@ -28,6 +28,7 @@ pub type TermDictionaryBuilder<W> = sstable::Writer<W, TermInfoValueWriter>;
 pub type TermStreamer<'a, A = AlwaysMatch> = sstable::Streamer<'a, TermSSTable, A>;
 
 /// SSTable used to store TermInfo objects.
+#[derive(Clone)]
 pub struct TermSSTable;
 
 pub type TermStreamerBuilder<'a, A = AlwaysMatch> = sstable::StreamerBuilder<'a, TermSSTable, A>;

sstable/Cargo.toml

@@ -11,6 +11,8 @@ description = "sstables for tantivy"
 
 [dependencies]
 common = {version= "0.7", path="../common", package="tantivy-common"}
+futures-util = "0.3.30"
+itertools = "0.13.0"
 tantivy-bitpacker = { version= "0.6", path="../bitpacker" }
 tantivy-fst = "0.5"
 # experimental gives us access to Decompressor::upper_bound

sstable/src/block_match_automaton.rs

@@ -0,0 +1,271 @@
+use tantivy_fst::Automaton;
+
+/// Returns whether a block can match an automaton based on its bounds.
+///
+/// start key is exclusive, and optional to account for the first block. end key is inclusive and
+/// mandatory.
+pub(crate) fn can_block_match_automaton(
+    start_key_opt: Option<&[u8]>,
+    end_key: &[u8],
+    automaton: &impl Automaton,
+) -> bool {
+    let start_key = if let Some(start_key) = start_key_opt {
+        start_key
+    } else {
+        // if start_key_opt is None, we would allow an automaton matching the empty string to match
+        if automaton.is_match(&automaton.start()) {
+            return true;
+        }
+        &[]
+    };
+    can_block_match_automaton_with_start(start_key, end_key, automaton)
+}
+
+// similar to can_block_match_automaton, ignoring the edge case of the initial block
+fn can_block_match_automaton_with_start(
+    start_key: &[u8],
+    end_key: &[u8],
+    automaton: &impl Automaton,
+) -> bool {
+    // notation: in loops, we use `kb` to denotate a key byte (a byte taken from the start/end key),
+    // and `rb`, a range byte (usually all values higher than a `kb` when comparing with
+    // start_key, or all values lower than a `kb` when comparing with end_key)
+
+    if start_key >= end_key {
+        return false;
+    }
+
+    let common_prefix_len = crate::common_prefix_len(start_key, end_key);
+
+    let mut base_state = automaton.start();
+    for kb in &start_key[0..common_prefix_len] {
+        base_state = automaton.accept(&base_state, *kb);
+    }
+
+    // this is not required for correctness, but allows dodging more expensive checks
+    if !automaton.can_match(&base_state) {
+        return false;
+    }
+
+    // we have 3 distinct case:
+    // - keys are `abc` and `abcd` => we test for abc[\0-d].*
+    // - keys are `abcd` and `abce` => we test for abc[d-e].*
+    // - keys are `abcd` and `abc` => contradiction with start_key < end_key.
+    //
+    // ideally for (abc, abcde] we could test for abc([\0-c].*|d([\0-d].*|e)?)
+    // but let's start simple (and correct), and tighten our bounds latter
+    //
+    // and for (abcde, abcfg] we could test for abc(d(e.+|[f-\xff].*)|e.*|f([\0-f].*|g)?)
+    // abc (
+    //  d(e.+|[f-\xff].*) |
+    //  e.* |
+    //  f([\0-f].*|g)?
+    // )
+    //
+    // these are all written as regex, but can be converted to operations we can do:
+    // - [x-y] is a for c in x..=y
+    // - .* is a can_match()
+    // - .+ is a for c in 0..=255 { accept(c).can_match() }
+    // - ? is a the thing before can_match(), or current state.is_match()
+    // - | means test both side
+
+    // we have two cases, either start_key is a prefix of end_key (e.g. (abc, abcjp]),
+    // or it is not (e.g. (abcdg, abcjp]). It is not possible however that end_key be a prefix of
+    // start_key (or that both are equal) because we already handled start_key >= end_key.
+    //
+    // if we are in the first case, we want to visit the following states:
+    // abc (
+    //   [\0-i].* |
+    //   j (
+    //     [\0-o].* |
+    //     p
+    //   )?
+    // )
+    // Everything after `abc` is handled by `match_range_end`
+    //
+    // if we are in the 2nd case, we want to visit the following states:
+    // abc (
+    //   d(g.+|[h-\xff].*) | // this is handled by match_range_start
+    //
+    //   [e-i].* |           // this is handled here
+    //
+    //   j (                 // this is handled by match_range_end (but countrary to the other
+    //    [\0-o].* |         // case, j is already consumed so to not check [\0-i].* )
+    //    p
+    //   )?
+    // )
+
+    let Some(start_range) = start_key.get(common_prefix_len) else {
+        return match_range_end(&end_key[common_prefix_len..], &automaton, base_state);
+    };
+
+    let end_range = end_key[common_prefix_len];
+
+    // things starting with start_range were handled in match_range_start
+    // this starting with end_range are handled bellow.
+    // this can run for 0 iteration in cases such as (abc, abd]
+    for rb in (start_range + 1)..end_range {
+        let new_state = automaton.accept(&base_state, rb);
+        if automaton.can_match(&new_state) {
+            return true;
+        }
+    }
+
+    let state_for_start = automaton.accept(&base_state, *start_range);
+    if match_range_start(
+        &start_key[common_prefix_len + 1..],
+        &automaton,
+        state_for_start,
+    ) {
+        return true;
+    }
+
+    let state_for_end = automaton.accept(&base_state, end_range);
+    if automaton.is_match(&state_for_end) {
+        return true;
+    }
+    match_range_end(&end_key[common_prefix_len + 1..], &automaton, state_for_end)
+}
+
+fn match_range_start<S, A: Automaton<State = S>>(
+    start_key: &[u8],
+    automaton: &A,
+    mut state: S,
+) -> bool {
+    // case (abcdgj, abcpqr], `abcd` is already consumed, we need to handle:
+    // - [h-\xff].*
+    // - g[k-\xff].*
+    // - gj.+ == gf[\0-\xff].*
+
+    for kb in start_key {
+        // this is an optimisation, and is not needed for correctness
+        if !automaton.can_match(&state) {
+            return false;
+        }
+
+        // does the [h-\xff].* part. we skip if kb==255 as [\{0100}-\xff] is an empty range, and
+        // this would overflow in our u8 world
+        if *kb < u8::MAX {
+            for rb in (kb + 1)..=u8::MAX {
+                let temp_state = automaton.accept(&state, rb);
+                if automaton.can_match(&temp_state) {
+                    return true;
+                }
+            }
+        }
+        // push g
+        state = automaton.accept(&state, *kb);
+    }
+
+    // this isn't required for correctness, but can save us from looping 256 below
+    if !automaton.can_match(&state) {
+        return false;
+    }
+
+    // does the final `.+`, which is the same as `[\0-\xff].*`
+    for rb in 0..=u8::MAX {
+        let temp_state = automaton.accept(&state, rb);
+        if automaton.can_match(&temp_state) {
+            return true;
+        }
+    }
+    false
+}
+
+fn match_range_end<S, A: Automaton<State = S>>(
+    end_key: &[u8],
+    automaton: &A,
+    mut state: S,
+) -> bool {
+    // for (abcdef, abcmps]. the prefix `abcm` has been consumed, `[d-l].*` was handled elsewhere,
+    // we just need to handle
+    // - [\0-o].*
+    // - p
+    // - p[\0-r].*
+    // - ps
+    for kb in end_key {
+        // this is an optimisation, and is not needed for correctness
+        if !automaton.can_match(&state) {
+            return false;
+        }
+
+        // does the `[\0-o].*`
+        for rb in 0..*kb {
+            let temp_state = automaton.accept(&state, rb);
+            if automaton.can_match(&temp_state) {
+                return true;
+            }
+        }
+
+        // push p
+        state = automaton.accept(&state, *kb);
+        // verify the `p` case
+        if automaton.is_match(&state) {
+            return true;
+        }
+    }
+    false
+}
+
+#[cfg(test)]
+pub(crate) mod tests {
+    use proptest::prelude::*;
+    use tantivy_fst::Automaton;
+
+    use super::*;
+
+    pub(crate) struct EqBuffer(pub Vec<u8>);
+
+    impl Automaton for EqBuffer {
+        type State = Option<usize>;
+
+        fn start(&self) -> Self::State {
+            Some(0)
+        }
+
+        fn is_match(&self, state: &Self::State) -> bool {
+            *state == Some(self.0.len())
+        }
+
+        fn accept(&self, state: &Self::State, byte: u8) -> Self::State {
+            state
+                .filter(|pos| self.0.get(*pos) == Some(&byte))
+                .map(|pos| pos + 1)
+        }
+
+        fn can_match(&self, state: &Self::State) -> bool {
+            state.is_some()
+        }
+
+        fn will_always_match(&self, _state: &Self::State) -> bool {
+            false
+        }
+    }
+
+    fn gen_key_strategy() -> impl Strategy<Value = Vec<u8>> {
+        // we only generate bytes in [0, 1, 2, 254, 255] to reduce the search space without
+        // ignoring edge cases that might ocure with integer over/underflow
+        proptest::collection::vec(prop_oneof![0u8..=2, 254u8..=255], 0..5)
+    }
+
+    proptest! {
+        #![proptest_config(ProptestConfig {
+            cases: 10000, .. ProptestConfig::default()
+        })]
+
+        #[test]
+        fn test_proptest_automaton_match_block(start in gen_key_strategy(), end in gen_key_strategy(), key in gen_key_strategy()) {
+            let expected = start < key && end >= key;
+            let automaton = EqBuffer(key);
+
+            assert_eq!(can_block_match_automaton(Some(&start), &end, &automaton), expected);
+        }
+
+        #[test]
+        fn test_proptest_automaton_match_first_block(end in gen_key_strategy(), key in gen_key_strategy()) {
+            let expected = end >= key;
+            let automaton = EqBuffer(key);
+            assert_eq!(can_block_match_automaton(None, &end, &automaton), expected);
+        }
+    }
+}

sstable/src/block_reader.rs

@@ -7,6 +7,7 @@ use zstd::bulk::Decompressor;
 pub struct BlockReader {
     buffer: Vec<u8>,
     reader: OwnedBytes,
+    next_readers: std::vec::IntoIter<OwnedBytes>,
     offset: usize,
 }
 
@@ -15,6 +16,18 @@ impl BlockReader {
         BlockReader {
             buffer: Vec::new(),
             reader,
+            next_readers: Vec::new().into_iter(),
+            offset: 0,
+        }
+    }
+
+    pub fn from_multiple_blocks(readers: Vec<OwnedBytes>) -> BlockReader {
+        let mut next_readers = readers.into_iter();
+        let reader = next_readers.next().unwrap_or_else(OwnedBytes::empty);
+        BlockReader {
+            buffer: Vec::new(),
+            reader,
+            next_readers,
             offset: 0,
         }
     }
@@ -34,42 +47,52 @@ impl BlockReader {
         self.offset = 0;
         self.buffer.clear();
 
-        let block_len = match self.reader.len() {
+        loop {
-            0 => return Ok(false),
+            let block_len = match self.reader.len() {
-            1..=3 => {
+                0 => {
+                    // we are out of data for this block. Check if we have another block after
+                    if let Some(new_reader) = self.next_readers.next() {
+                        self.reader = new_reader;
+                        continue;
+                    } else {
+                        return Ok(false);
+                    }
+                }
+                1..=3 => {
+                    return Err(io::Error::new(
+                        io::ErrorKind::UnexpectedEof,
+                        "failed to read block_len",
+                    ))
+                }
+                _ => self.reader.read_u32() as usize,
+            };
+            if block_len <= 1 {
+                return Ok(false);
+            }
+            let compress = self.reader.read_u8();
+            let block_len = block_len - 1;
+
+            if self.reader.len() < block_len {
                 return Err(io::Error::new(
                     io::ErrorKind::UnexpectedEof,
-                    "failed to read block_len",
+                    "failed to read block content",
-                ))
+                ));
             }
-            _ => self.reader.read_u32() as usize,
+            if compress == 1 {
-        };
+                let required_capacity =
-        if block_len <= 1 {
+                    Decompressor::upper_bound(&self.reader[..block_len]).unwrap_or(1024 * 1024);
-            return Ok(false);
+                self.buffer.reserve(required_capacity);
-        }
+                Decompressor::new()?
-        let compress = self.reader.read_u8();
+                    .decompress_to_buffer(&self.reader[..block_len], &mut self.buffer)?;
-        let block_len = block_len - 1;
 
-        if self.reader.len() < block_len {
+                self.reader.advance(block_len);
-            return Err(io::Error::new(
+            } else {
-                io::ErrorKind::UnexpectedEof,
+                self.buffer.resize(block_len, 0u8);
-                "failed to read block content",
+                self.reader.read_exact(&mut self.buffer[..])?;
-            ));
+            }
-        }
-        if compress == 1 {
-            let required_capacity =
-                Decompressor::upper_bound(&self.reader[..block_len]).unwrap_or(1024 * 1024);
-            self.buffer.reserve(required_capacity);
-            Decompressor::new()?
-                .decompress_to_buffer(&self.reader[..block_len], &mut self.buffer)?;
 
-            self.reader.advance(block_len);
+            return Ok(true);
-        } else {
-            self.buffer.resize(block_len, 0u8);
-            self.reader.read_exact(&mut self.buffer[..])?;
         }
-
-        Ok(true)
     }
 
     #[inline(always)]

sstable/src/delta.rs

@@ -143,6 +143,16 @@ where TValueReader: value::ValueReader
         }
     }
 
+    pub fn from_multiple_blocks(reader: Vec<OwnedBytes>) -> Self {
+        DeltaReader {
+            idx: 0,
+            common_prefix_len: 0,
+            suffix_range: 0..0,
+            value_reader: TValueReader::default(),
+            block_reader: BlockReader::from_multiple_blocks(reader),
+        }
+    }
+
     pub fn empty() -> Self {
         DeltaReader::new(OwnedBytes::empty())
     }

sstable/src/dictionary.rs

@@ -7,6 +7,8 @@ use std::sync::Arc;
 use common::bounds::{transform_bound_inner_res, TransformBound};
 use common::file_slice::FileSlice;
 use common::{BinarySerializable, OwnedBytes};
+use futures_util::{stream, StreamExt, TryStreamExt};
+use itertools::Itertools;
 use tantivy_fst::automaton::AlwaysMatch;
 use tantivy_fst::Automaton;
 
@@ -98,20 +100,52 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
         &self,
         key_range: impl RangeBounds<[u8]>,
         limit: Option<u64>,
+        automaton: &impl Automaton,
+        merge_holes_under_bytes: usize,
     ) -> io::Result<DeltaReader<TSSTable::ValueReader>> {
-        let slice = self.file_slice_for_range(key_range, limit);
+        let match_all = automaton.will_always_match(&automaton.start());
-        let data = slice.read_bytes_async().await?;
+        if match_all {
-        Ok(TSSTable::delta_reader(data))
+            let slice = self.file_slice_for_range(key_range, limit);
+            let data = slice.read_bytes_async().await?;
+            Ok(TSSTable::delta_reader(data))
+        } else {
+            let blocks = stream::iter(self.get_block_iterator_for_range_and_automaton(
+                key_range,
+                automaton,
+                merge_holes_under_bytes,
+            ));
+            let data = blocks
+                .map(|block_addr| {
+                    self.sstable_slice
+                        .read_bytes_slice_async(block_addr.byte_range)
+                })
+                .buffered(5)
+                .try_collect::<Vec<_>>()
+                .await?;
+            Ok(DeltaReader::from_multiple_blocks(data))
+        }
     }
 
     pub(crate) fn sstable_delta_reader_for_key_range(
         &self,
         key_range: impl RangeBounds<[u8]>,
         limit: Option<u64>,
+        automaton: &impl Automaton,
     ) -> io::Result<DeltaReader<TSSTable::ValueReader>> {
-        let slice = self.file_slice_for_range(key_range, limit);
+        let match_all = automaton.will_always_match(&automaton.start());
-        let data = slice.read_bytes()?;
+        if match_all {
-        Ok(TSSTable::delta_reader(data))
+            let slice = self.file_slice_for_range(key_range, limit);
+            let data = slice.read_bytes()?;
+            Ok(TSSTable::delta_reader(data))
+        } else {
+            // if operations are sync, we assume latency is almost null, and there is no point in
+            // merging accross holes
+            let blocks = self.get_block_iterator_for_range_and_automaton(key_range, automaton, 0);
+            let data = blocks
+                .map(|block_addr| self.sstable_slice.read_bytes_slice(block_addr.byte_range))
+                .collect::<Result<Vec<_>, _>>()?;
+            Ok(DeltaReader::from_multiple_blocks(data))
+        }
     }
 
     pub(crate) fn sstable_delta_reader_block(
@@ -204,6 +238,42 @@ impl<TSSTable: SSTable> Dictionary<TSSTable> {
         self.sstable_slice.slice((start_bound, end_bound))
     }
 
+    fn get_block_iterator_for_range_and_automaton<'a>(
+        &'a self,
+        key_range: impl RangeBounds<[u8]>,
+        automaton: &'a impl Automaton,
+        merge_holes_under_bytes: usize,
+    ) -> impl Iterator<Item = BlockAddr> + 'a {
+        let lower_bound = match key_range.start_bound() {
+            Bound::Included(key) | Bound::Excluded(key) => {
+                self.sstable_index.locate_with_key(key).unwrap_or(u64::MAX)
+            }
+            Bound::Unbounded => 0,
+        };
+
+        let upper_bound = match key_range.end_bound() {
+            Bound::Included(key) | Bound::Excluded(key) => {
+                self.sstable_index.locate_with_key(key).unwrap_or(u64::MAX)
+            }
+            Bound::Unbounded => u64::MAX,
+        };
+        let block_range = lower_bound..=upper_bound;
+        self.sstable_index
+            .get_block_for_automaton(automaton)
+            .filter(move |(block_id, _)| block_range.contains(block_id))
+            .map(|(_, block_addr)| block_addr)
+            .coalesce(move |first, second| {
+                if first.byte_range.end + merge_holes_under_bytes >= second.byte_range.start {
+                    Ok(BlockAddr {
+                        first_ordinal: first.first_ordinal,
+                        byte_range: first.byte_range.start..second.byte_range.end,
+                    })
+                } else {
+                    Err((first, second))
+                }
+            })
+    }
+
     /// Opens a `TermDictionary`.
     pub fn open(term_dictionary_file: FileSlice) -> io::Result<Self> {
         let (main_slice, footer_len_slice) = term_dictionary_file.split_from_end(20);

sstable/src/lib.rs

@@ -3,6 +3,7 @@ use std::ops::Range;
 
 use merge::ValueMerger;
 
+mod block_match_automaton;
 mod delta;
 mod dictionary;
 pub mod merge;

sstable/src/sstable_index_v2.rs

@@ -1,10 +1,12 @@
 use common::OwnedBytes;
+use tantivy_fst::Automaton;
 
+use crate::block_match_automaton::can_block_match_automaton;
 use crate::{BlockAddr, SSTable, SSTableDataCorruption, TermOrdinal};
 
 #[derive(Default, Debug, Clone)]
 pub struct SSTableIndex {
-    blocks: Vec<BlockMeta>,
+    pub(crate) blocks: Vec<BlockMeta>,
 }
 
 impl SSTableIndex {
@@ -74,6 +76,31 @@ impl SSTableIndex {
         // locate_with_ord always returns an index within range
         self.get_block(self.locate_with_ord(ord)).unwrap()
     }
+
+    pub(crate) fn get_block_for_automaton<'a>(
+        &'a self,
+        automaton: &'a impl Automaton,
+    ) -> impl Iterator<Item = (u64, BlockAddr)> + 'a {
+        std::iter::once((None, &self.blocks[0]))
+            .chain(self.blocks.windows(2).map(|window| {
+                let [prev, curr] = window else {
+                    unreachable!();
+                };
+                (Some(&*prev.last_key_or_greater), curr)
+            }))
+            .enumerate()
+            .filter_map(move |(pos, (prev_key, current_block))| {
+                if can_block_match_automaton(
+                    prev_key,
+                    &current_block.last_key_or_greater,
+                    automaton,
+                ) {
+                    Some((pos as u64, current_block.block_addr.clone()))
+                } else {
+                    None
+                }
+            })
+    }
 }
 
 #[derive(Debug, Clone)]
@@ -99,3 +126,106 @@ impl SSTable for IndexSSTable {
 
     type ValueWriter = crate::value::index::IndexValueWriter;
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::block_match_automaton::tests::EqBuffer;
+
+    #[test]
+    fn test_get_block_for_automaton() {
+        let sstable = SSTableIndex {
+            blocks: vec![
+                BlockMeta {
+                    last_key_or_greater: vec![0, 1, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 0,
+                        byte_range: 0..10,
+                    },
+                },
+                BlockMeta {
+                    last_key_or_greater: vec![0, 2, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 5,
+                        byte_range: 10..20,
+                    },
+                },
+                BlockMeta {
+                    last_key_or_greater: vec![0, 3, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 10,
+                        byte_range: 20..30,
+                    },
+                },
+            ],
+        };
+
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 1, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                0,
+                BlockAddr {
+                    first_ordinal: 0,
+                    byte_range: 0..10
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 2, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                1,
+                BlockAddr {
+                    first_ordinal: 5,
+                    byte_range: 10..20
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 3, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                2,
+                BlockAddr {
+                    first_ordinal: 10,
+                    byte_range: 20..30
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 4, 1]))
+            .collect::<Vec<_>>();
+        assert!(res.is_empty());
+
+        let complex_automaton = EqBuffer(vec![0, 1, 1]).union(EqBuffer(vec![0, 3, 1]));
+        let res = sstable
+            .get_block_for_automaton(&complex_automaton)
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![
+                (
+                    0,
+                    BlockAddr {
+                        first_ordinal: 0,
+                        byte_range: 0..10
+                    }
+                ),
+                (
+                    2,
+                    BlockAddr {
+                        first_ordinal: 10,
+                        byte_range: 20..30
+                    }
+                )
+            ]
+        );
+    }
+}

sstable/src/sstable_index_v3.rs

@@ -5,8 +5,9 @@ use std::sync::Arc;
 use common::{BinarySerializable, FixedSize, OwnedBytes};
 use tantivy_bitpacker::{compute_num_bits, BitPacker};
 use tantivy_fst::raw::Fst;
-use tantivy_fst::{IntoStreamer, Map, MapBuilder, Streamer};
+use tantivy_fst::{Automaton, IntoStreamer, Map, MapBuilder, Streamer};
 
+use crate::block_match_automaton::can_block_match_automaton;
 use crate::{common_prefix_len, SSTableDataCorruption, TermOrdinal};
 
 #[derive(Debug, Clone)]
@@ -64,6 +65,41 @@ impl SSTableIndex {
             SSTableIndex::V3Empty(v3_empty) => v3_empty.get_block_with_ord(ord),
         }
     }
+
+    pub fn get_block_for_automaton<'a>(
+        &'a self,
+        automaton: &'a impl Automaton,
+    ) -> impl Iterator<Item = (u64, BlockAddr)> + 'a {
+        match self {
+            SSTableIndex::V2(v2_index) => {
+                BlockIter::V2(v2_index.get_block_for_automaton(automaton))
+            }
+            SSTableIndex::V3(v3_index) => {
+                BlockIter::V3(v3_index.get_block_for_automaton(automaton))
+            }
+            SSTableIndex::V3Empty(v3_empty) => {
+                BlockIter::V3Empty(std::iter::once((0, v3_empty.block_addr.clone())))
+            }
+        }
+    }
+}
+
+enum BlockIter<V2, V3, T> {
+    V2(V2),
+    V3(V3),
+    V3Empty(std::iter::Once<T>),
+}
+
+impl<V2: Iterator<Item = T>, V3: Iterator<Item = T>, T> Iterator for BlockIter<V2, V3, T> {
+    type Item = T;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        match self {
+            BlockIter::V2(v2) => v2.next(),
+            BlockIter::V3(v3) => v3.next(),
+            BlockIter::V3Empty(once) => once.next(),
+        }
+    }
 }
 
 #[derive(Debug, Clone)]
@@ -123,6 +159,59 @@ impl SSTableIndexV3 {
     pub(crate) fn get_block_with_ord(&self, ord: TermOrdinal) -> BlockAddr {
         self.block_addr_store.binary_search_ord(ord).1
     }
+
+    pub(crate) fn get_block_for_automaton<'a>(
+        &'a self,
+        automaton: &'a impl Automaton,
+    ) -> impl Iterator<Item = (u64, BlockAddr)> + 'a {
+        // this is more complicated than other index formats: we don't have a ready made list of
+        // blocks, and instead need to stream-decode the sstable.
+
+        GetBlockForAutomaton {
+            streamer: self.fst_index.stream(),
+            block_addr_store: &self.block_addr_store,
+            prev_key: None,
+            automaton,
+        }
+    }
+}
+
+// TODO we iterate over the entire Map to find matching blocks,
+// we could manually iterate on the underlying Fst and skip whole branches if our Automaton says
+// cannot match. this isn't as bad as it sounds given the fst is a lot smaller than the rest of the
+// sstable.
+// To do that, we can't use tantivy_fst's Stream with an automaton, as we need to know 2 consecutive
+// fst keys to form a proper opinion on whether this is a match, which we wan't translate into a
+// single automaton
+struct GetBlockForAutomaton<'a, A: Automaton> {
+    streamer: tantivy_fst::map::Stream<'a>,
+    block_addr_store: &'a BlockAddrStore,
+    prev_key: Option<Vec<u8>>,
+    automaton: &'a A,
+}
+
+impl<A: Automaton> Iterator for GetBlockForAutomaton<'_, A> {
+    type Item = (u64, BlockAddr);
+
+    fn next(&mut self) -> Option<Self::Item> {
+        while let Some((new_key, block_id)) = self.streamer.next() {
+            if let Some(prev_key) = self.prev_key.as_mut() {
+                if can_block_match_automaton(Some(prev_key), new_key, self.automaton) {
+                    prev_key.clear();
+                    prev_key.extend_from_slice(new_key);
+                    return Some((block_id, self.block_addr_store.get(block_id).unwrap()));
+                }
+                prev_key.clear();
+                prev_key.extend_from_slice(new_key);
+            } else {
+                self.prev_key = Some(new_key.to_owned());
+                if can_block_match_automaton(None, new_key, self.automaton) {
+                    return Some((block_id, self.block_addr_store.get(block_id).unwrap()));
+                }
+            }
+        }
+        None
+    }
 }
 
 #[derive(Debug, Clone)]
@@ -734,7 +823,8 @@ fn find_best_slope(elements: impl Iterator<Item = (usize, u64)> + Clone) -> (u32
 mod tests {
     use common::OwnedBytes;
 
-    use super::{BlockAddr, SSTableIndexBuilder, SSTableIndexV3};
+    use super::*;
+    use crate::block_match_automaton::tests::EqBuffer;
     use crate::SSTableDataCorruption;
 
     #[test]
@@ -823,4 +913,108 @@ mod tests {
             (12345, 1)
         );
     }
+
+    #[test]
+    fn test_get_block_for_automaton() {
+        let sstable_index_builder = SSTableIndexBuilder {
+            blocks: vec![
+                BlockMeta {
+                    last_key_or_greater: vec![0, 1, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 0,
+                        byte_range: 0..10,
+                    },
+                },
+                BlockMeta {
+                    last_key_or_greater: vec![0, 2, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 5,
+                        byte_range: 10..20,
+                    },
+                },
+                BlockMeta {
+                    last_key_or_greater: vec![0, 3, 2],
+                    block_addr: BlockAddr {
+                        first_ordinal: 10,
+                        byte_range: 20..30,
+                    },
+                },
+            ],
+        };
+
+        let mut sstable_index_bytes = Vec::new();
+        let fst_len = sstable_index_builder
+            .serialize(&mut sstable_index_bytes)
+            .unwrap();
+
+        let sstable = SSTableIndexV3::load(OwnedBytes::new(sstable_index_bytes), fst_len).unwrap();
+
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 1, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                0,
+                BlockAddr {
+                    first_ordinal: 0,
+                    byte_range: 0..10
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 2, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                1,
+                BlockAddr {
+                    first_ordinal: 5,
+                    byte_range: 10..20
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 3, 1]))
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![(
+                2,
+                BlockAddr {
+                    first_ordinal: 10,
+                    byte_range: 20..30
+                }
+            )]
+        );
+        let res = sstable
+            .get_block_for_automaton(&EqBuffer(vec![0, 4, 1]))
+            .collect::<Vec<_>>();
+        assert!(res.is_empty());
+
+        let complex_automaton = EqBuffer(vec![0, 1, 1]).union(EqBuffer(vec![0, 3, 1]));
+        let res = sstable
+            .get_block_for_automaton(&complex_automaton)
+            .collect::<Vec<_>>();
+        assert_eq!(
+            res,
+            vec![
+                (
+                    0,
+                    BlockAddr {
+                        first_ordinal: 0,
+                        byte_range: 0..10
+                    }
+                ),
+                (
+                    2,
+                    BlockAddr {
+                        first_ordinal: 10,
+                        byte_range: 20..30
+                    }
+                )
+            ]
+        );
+    }
 }

sstable/src/streamer.rs

@@ -86,16 +86,24 @@ where
             bound_as_byte_slice(&self.upper),
         );
         self.term_dict
-            .sstable_delta_reader_for_key_range(key_range, self.limit)
+            .sstable_delta_reader_for_key_range(key_range, self.limit, &self.automaton)
     }
 
-    async fn delta_reader_async(&self) -> io::Result<DeltaReader<TSSTable::ValueReader>> {
+    async fn delta_reader_async(
+        &self,
+        merge_holes_under_bytes: usize,
+    ) -> io::Result<DeltaReader<TSSTable::ValueReader>> {
         let key_range = (
             bound_as_byte_slice(&self.lower),
             bound_as_byte_slice(&self.upper),
         );
         self.term_dict
-            .sstable_delta_reader_for_key_range_async(key_range, self.limit)
+            .sstable_delta_reader_for_key_range_async(
+                key_range,
+                self.limit,
+                &self.automaton,
+                merge_holes_under_bytes,
+            )
             .await
     }
 
@@ -130,7 +138,16 @@ where
 
     /// See `into_stream(..)`
     pub async fn into_stream_async(self) -> io::Result<Streamer<'a, TSSTable, A>> {
-        let delta_reader = self.delta_reader_async().await?;
+        self.into_stream_async_merging_holes(0).await
+    }
+
+    /// Same as `into_stream_async`, but tries to issue a single io operation when requesting
+    /// blocks that are not consecutive, but also less than `merge_holes_under_bytes` bytes appart.
+    pub async fn into_stream_async_merging_holes(
+        self,
+        merge_holes_under_bytes: usize,
+    ) -> io::Result<Streamer<'a, TSSTable, A>> {
+        let delta_reader = self.delta_reader_async(merge_holes_under_bytes).await?;
         self.into_stream_given_delta_reader(delta_reader)
     }
 
@@ -327,4 +344,7 @@ mod tests {
         assert!(!term_streamer.advance());
         Ok(())
     }
+
+    // TODO add test for sparse search with a block of poison (starts with 0xffffffff) => such a
+    // block instantly causes an unexpected EOF error
 }