armdb 0.15.0

sharded bitcask key-value storage optimized for NVMe
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
//! Bounded ordered merge. Layer ownership exists only inside a refill.

use super::{
    block::EntryIter,
    cache::SegBlockCache,
    core::SegCore,
    generation::{GenLoc, Generation},
    hash::key_hash,
    page_index::IndexGroup,
    query::{SegRange, bound_key},
    segment::Segment,
};
use crate::{
    error::{DbError, DbResult},
    key::Key,
    map_index::Dir,
};
use std::{
    cmp::{Ordering, Reverse},
    collections::{BinaryHeap, HashMap, VecDeque},
    hash::Hash,
    sync::Arc,
};

const ROW_LIMIT: usize = 256;
const BYTE_LIMIT: usize = 1 << 20;
const CONSUMED_LIMIT: usize = 4096;

/// A segment's current block, parsed, kept by an internal scan between
/// refills. It owns only bytes: no segment, file or layer is retained.
struct CarriedBlock {
    group: Option<Arc<IndexGroup>>,
    block: usize,
    payload: Arc<[u8]>,
    offsets: Arc<[usize]>,
}

enum Payload {
    Tombstone,
    Journal(GenLoc),
    Block { offset: usize },
}
struct Item<K> {
    key: K,
    payload: Payload,
}

struct GenerationStream<K: Key + Hash + Eq + Send + Sync> {
    generation: Arc<Generation<K>>,
    shard: usize,
    query: SegRange<K>,
    cursor: Option<K>,
    entries: std::vec::IntoIter<(K, GenLoc)>,
    batch: usize,
    done: bool,
}
impl<K: Key + Hash + Eq + Send + Sync> GenerationStream<K> {
    fn next(&mut self) -> Option<Item<K>> {
        if self.entries.len() == 0 && !self.done {
            let query = self.query.clone().after(self.cursor.as_ref());
            let batch = if query.empty {
                Vec::new()
            } else {
                self.generation.ordered_batch(
                    self.shard,
                    query.start.as_ref(),
                    query.end.as_ref(),
                    query.dir,
                    self.batch,
                )
            };
            self.done = batch.len() < self.batch;
            self.entries = batch.into_iter();
        }
        self.entries.next().map(|(key, loc)| {
            self.cursor = Some(key);
            Item {
                key,
                payload: if loc.tombstone {
                    Payload::Tombstone
                } else {
                    Payload::Journal(loc)
                },
            }
        })
    }
}

struct SegmentStream<K: Key> {
    segment: Arc<Segment>,
    query: SegRange<K>,
    block: usize,
    payload: Option<Arc<[u8]>>,
    parsed: Option<Arc<[usize]>>,
    offsets: VecDeque<usize>,
    carried: Option<CarriedBlock>,
    group: Option<Arc<IndexGroup>>,
    done: bool,
}
impl<K: Key> SegmentStream<K> {
    fn new(
        segment: Arc<Segment>,
        query: SegRange<K>,
        carried: Option<CarriedBlock>,
        cache: Option<&SegBlockCache>,
    ) -> DbResult<Self> {
        let bound = match query.dir {
            Dir::Asc => &query.start,
            Dir::Desc => &query.end,
        };
        let mut group = carried.as_ref().and_then(|c| c.group.clone());
        // The previous refill consumed every earlier directional block of
        // this stream. Its retained head can be ahead of the merged cursor,
        // even in the next group; seeking backwards would reread that group.
        // Query bounds are still applied to every entry in the carried block.
        let block = if let Some(carried) = &carried {
            carried.block
        } else {
            match bound_key(bound) {
                Some(k) => segment.locate_key_in(k.as_bytes(), &mut group, cache)?,
                None if query.dir == Dir::Asc => 0,
                None => segment.block_count()? - 1,
            }
        };
        Ok(Self {
            segment,
            query,
            block,
            payload: None,
            parsed: None,
            offsets: VecDeque::new(),
            carried,
            group,
            done: false,
        })
    }

    /// The loaded block, to carry into the next refill.
    fn into_carried(self) -> Option<CarriedBlock> {
        Some(CarriedBlock {
            group: self.group,
            block: self.block,
            payload: self.payload?,
            offsets: self.parsed?,
        })
    }

    fn next(&mut self, core: &SegCore<K>, use_cache: bool) -> DbResult<Option<Item<K>>>
    where
        K: Hash + Eq + Send + Sync,
    {
        loop {
            if self.done {
                return Ok(None);
            }
            if self.payload.is_none() {
                // Reuse owned bytes and parsed offsets from the previous
                // refill, then apply the fresh query bounds to its entries.
                let block = self.block;
                let (payload, parsed) = match self.carried.take_if(|c| c.block == block) {
                    Some(c) => {
                        self.group = c.group;
                        (c.payload, c.offsets)
                    }
                    None => {
                        let payload = self.segment.block_payload_in(
                            self.block,
                            &mut self.group,
                            use_cache.then_some(&core.cache),
                        )?;
                        let mut offsets = Vec::new();
                        let mut offset = 0;
                        for entry in EntryIter::new(&payload, self.segment.key_len()) {
                            let entry = entry.map_err(|e| {
                                e.into_db(
                                    self.segment.id(),
                                    Some(self.block as u32),
                                    self.segment.sdat_path(),
                                )
                            })?;
                            offsets.push(offset);
                            offset += entry.raw.len();
                        }
                        (payload, Arc::from(offsets))
                    }
                };
                self.offsets = parsed.iter().copied().collect();
                self.payload = Some(payload);
                self.parsed = Some(parsed);
            }
            let offset = match self.query.dir {
                Dir::Asc => self.offsets.pop_front(),
                Dir::Desc => self.offsets.pop_back(),
            };
            let Some(offset) = offset else {
                self.payload = None;
                self.parsed = None;
                match self.query.dir {
                    Dir::Asc if self.block + 1 < self.segment.block_count()? => self.block += 1,
                    Dir::Desc if self.block > 0 => self.block -= 1,
                    _ => self.done = true,
                }
                continue;
            };
            let payload = self.payload.as_ref().expect("loaded block");
            let entry = EntryIter::new(&payload[offset..], self.segment.key_len())
                .next()
                .expect("known entry offset")
                .map_err(|e| {
                    e.into_db(
                        self.segment.id(),
                        Some(self.block as u32),
                        self.segment.sdat_path(),
                    )
                })?;
            if !self.query.contains(entry.key) {
                // The initial block can straddle the near bound; beyond the
                // far bound no subsequent entry can belong to this query.
                let past = match self.query.dir {
                    Dir::Asc => {
                        bound_key(&self.query.end).is_some_and(|k| entry.key >= k.as_bytes())
                    }
                    Dir::Desc => {
                        bound_key(&self.query.start).is_some_and(|k| entry.key <= k.as_bytes())
                    }
                };
                if past {
                    self.done = true;
                    return Ok(None);
                }
                continue;
            }
            entry.verify_crc().map_err(|e| {
                e.into_db(
                    self.segment.id(),
                    Some(self.block as u32),
                    self.segment.sdat_path(),
                )
            })?;
            return Ok(Some(Item {
                key: K::from_bytes(entry.key),
                payload: if entry.is_tombstone() {
                    Payload::Tombstone
                } else {
                    Payload::Block { offset }
                },
            }));
        }
    }

    fn value(&self, offset: usize) -> DbResult<Vec<u8>> {
        let payload = self
            .payload
            .as_ref()
            .ok_or(DbError::Internal("ordered block head lost its payload"))?;
        let entry = EntryIter::new(&payload[offset..], self.segment.key_len())
            .next()
            .expect("known entry offset")
            .map_err(|e| {
                e.into_db(
                    self.segment.id(),
                    Some(self.block as u32),
                    self.segment.sdat_path(),
                )
            })?;
        Ok(entry.value.to_vec())
    }
}

enum Stream<K: Key + Hash + Eq + Send + Sync> {
    Generation(GenerationStream<K>),
    Segment(SegmentStream<K>),
}
impl<K: Key + Hash + Eq + Send + Sync> Stream<K> {
    fn next(&mut self, core: &SegCore<K>, use_cache: bool) -> DbResult<Option<Item<K>>> {
        match self {
            Self::Generation(s) => Ok(s.next()),
            Self::Segment(s) => s.next(core, use_cache),
        }
    }
    fn value(
        &self,
        key: &K,
        payload: Payload,
        seed: u64,
        want_values: bool,
    ) -> DbResult<Option<Vec<u8>>> {
        if matches!(payload, Payload::Tombstone) {
            return Ok(None);
        }
        if !want_values {
            return Ok(Some(Vec::new()));
        }
        match (self, payload) {
            (Self::Generation(s), Payload::Journal(loc)) => s
                .generation
                .read_value(key_hash(key.as_bytes(), seed), &loc.loc)
                .map(Some),
            (Self::Segment(s), Payload::Block { offset }) => s.value(offset).map(Some),
            _ => Err(DbError::Internal(
                "ordered payload disagrees with its stream",
            )),
        }
    }
}

struct Ranked<K> {
    key: K,
    rank: usize,
    dir: Dir,
}
impl<K: Key> Ord for Ranked<K> {
    fn cmp(&self, other: &Self) -> Ordering {
        let order = self.key.as_bytes().cmp(other.key.as_bytes());
        (if self.dir == Dir::Asc {
            order
        } else {
            order.reverse()
        })
        .then(self.rank.cmp(&other.rank))
    }
}
impl<K: Key> PartialOrd for Ranked<K> {
    fn partial_cmp(&self, o: &Self) -> Option<Ordering> {
        Some(self.cmp(o))
    }
}
impl<K: Key> PartialEq for Ranked<K> {
    fn eq(&self, o: &Self) -> bool {
        self.cmp(o) == Ordering::Equal
    }
}
impl<K: Key> Eq for Ranked<K> {}

/// A weakly consistent ordered iterator retaining only owned buffered rows.
/// Every refill resolves and releases its fresh view before returning control.
pub(crate) struct RawOrderedIter<'a, K: Key + Hash + Eq + Send + Sync> {
    core: &'a SegCore<K>,
    query: SegRange<K>,
    cursor: Option<K>,
    rows: VecDeque<(K, Vec<u8>)>,
    error: Option<DbError>,
    done: bool,
    want_values: bool,
    use_cache: bool,
    /// Visible rows per refill, at most [`ROW_LIMIT`].
    row_limit: usize,
    /// Internal scans bypass the block cache. Without this, every refill
    /// would read the current block of every overlapping segment again: a
    /// full scan would cost `S` block reads per refill instead of one read
    /// per block. Keyed by `(segment_id, version)`; at most one block per
    /// segment of the last view. Retained memory includes the block payload,
    /// parsed row offsets, and its `IndexGroup`, summed across those segments.
    carried: HashMap<(u64, u32), CarriedBlock>,
}

impl<K: Key + Hash + Eq + Send + Sync + 'static> SegCore<K> {
    pub(crate) fn ordered_iter(
        &self,
        query: SegRange<K>,
        want_values: bool,
        use_cache: bool,
    ) -> RawOrderedIter<'_, K> {
        RawOrderedIter {
            core: self,
            query,
            cursor: None,
            rows: VecDeque::new(),
            error: None,
            done: false,
            want_values,
            use_cache,
            row_limit: ROW_LIMIT,
            carried: HashMap::new(),
        }
    }
}

impl<K: Key + Hash + Eq + Send + Sync + 'static> RawOrderedIter<'_, K> {
    /// Read at most `rows` visible rows per refill, for a caller that takes
    /// only that many: `first`, `last` and small pages.
    pub(crate) fn limit_rows(&mut self, rows: usize) {
        self.row_limit = rows.clamp(1, ROW_LIMIT);
    }

    #[cfg(test)]
    pub(crate) fn buffered_stats(&self) -> (usize, usize) {
        (
            self.rows.len(),
            self.rows.iter().map(|(_, v)| v.len()).sum(),
        )
    }

    fn refill(&mut self) -> DbResult<()> {
        let query = self.query.clone().after(self.cursor.as_ref());
        if query.empty {
            self.done = true;
            return Ok(());
        }
        // Count actual fresh-view/stream rebuild attempts, including invisible
        // slices and attempts that later fail; buffered rows emit no metric.
        metrics::counter!("armdb.seg.ordered_refills", "collection" => self.core.meta.name)
            .increment(1);
        let view = self.core.view();
        let mut streams = Vec::new();
        for (i, generation) in std::iter::once(&view.active)
            .chain(view.frozen.iter())
            .enumerate()
        {
            // The active layer is never pruned by time, as in a point read,
            // even when an indeterminate freeze left it marked frozen.
            if i > 0
                && query
                    .since_ms
                    .is_some_and(|ms| generation.frozen_at_ms() < ms)
            {
                continue;
            }
            for shard in 0..generation.shard_count() {
                streams.push(Stream::Generation(GenerationStream {
                    generation: generation.clone(),
                    shard,
                    query: query.clone(),
                    cursor: None,
                    entries: Vec::new().into_iter(),
                    batch: self.row_limit,
                    done: false,
                }));
            }
        }
        for segment in &view.sealed {
            if segment
                .key_range()
                .is_some_and(|r| !query.overlaps(&r.min, &r.max))
            {
                continue;
            }
            if query
                .since_ms
                .is_some_and(|ms| segment.record().frozen_at_ms < ms)
            {
                continue;
            }
            let id = (segment.id(), segment.record().version);
            let filter_len = segment
                .prefix_filter_lens()
                .iter()
                .map(|n| usize::from(*n))
                .filter(|n| *n <= query.pruning_prefix.len())
                .max();
            // A carried segment passed this probe in the previous refill; its
            // file and `pruning_prefix` (kept by `after`) are unchanged.
            if !self.carried.contains_key(&id)
                && let Some(n) = filter_len
                && !segment.prefix_may_contain(
                    &query.pruning_prefix[..n],
                    self.use_cache.then_some(&self.core.cache),
                )?
            {
                continue;
            }
            // Count actual surviving segment stream initializations after range,
            // prefix and time pruning. Generation shard seeks are excluded.
            metrics::counter!("armdb.seg.ordered_reseeks", "collection" => self.core.meta.name)
                .increment(1);
            let carried = self.carried.remove(&id);
            streams.push(Stream::Segment(SegmentStream::new(
                segment.clone(),
                query.clone(),
                carried,
                self.use_cache.then_some(&self.core.cache),
            )?));
        }
        let result = self.merge(&query, &mut streams);
        // Blocks of segments this view no longer reaches are dropped here.
        self.carried.clear();
        for stream in streams {
            if let Stream::Segment(s) = stream {
                let id = (s.segment.id(), s.segment.record().version);
                if let Some(c) = s.into_carried() {
                    self.carried.insert(id, c);
                }
            }
        }
        result
    }

    fn merge(&mut self, query: &SegRange<K>, streams: &mut [Stream<K>]) -> DbResult<()> {
        let mut heads = Vec::with_capacity(streams.len());
        let mut heap = BinaryHeap::new();
        for (rank, stream) in streams.iter_mut().enumerate() {
            let item = stream.next(self.core, self.use_cache)?;
            if let Some(item) = &item {
                heap.push(Reverse(Ranked {
                    key: item.key,
                    rank,
                    dir: query.dir,
                }));
            }
            heads.push(item);
        }
        let mut bytes: usize = 0;
        for _ in 0..CONSUMED_LIMIT {
            let Some(Reverse(top)) = heap.pop() else {
                self.done = true;
                return Ok(());
            };
            let item = heads[top.rank].take().expect("heap head");
            let value = streams[top.rank].value(
                &item.key,
                item.payload,
                self.core.seed,
                self.want_values,
            )?;
            if let Some(value) = value {
                if !self.rows.is_empty() && bytes.saturating_add(value.len()) > BYTE_LIMIT {
                    return Ok(());
                }
                bytes += value.len();
                self.rows.push_back((item.key, value));
            }
            self.cursor = Some(item.key);
            let mut ranks = vec![top.rank];
            while heap
                .peek()
                .is_some_and(|Reverse(next)| next.key.as_bytes() == item.key.as_bytes())
            {
                let Reverse(next) = heap.pop().expect("peeked head");
                heads[next.rank] = None;
                ranks.push(next.rank);
            }
            // A completed batch need not prefetch a later block or value.
            if self.rows.len() == self.row_limit || bytes >= BYTE_LIMIT {
                return Ok(());
            }
            for rank in ranks {
                let next = streams[rank].next(self.core, self.use_cache)?;
                if let Some(item) = &next {
                    heap.push(Reverse(Ranked {
                        key: item.key,
                        rank,
                        dir: query.dir,
                    }));
                }
                heads[rank] = next;
            }
        }
        Ok(())
    }
}
impl<K: Key + Hash + Eq + Send + Sync + 'static> Iterator for RawOrderedIter<'_, K> {
    type Item = DbResult<(K, Vec<u8>)>;
    fn next(&mut self) -> Option<Self::Item> {
        loop {
            if let Some(row) = self.rows.pop_front() {
                return Some(Ok(row));
            }
            if let Some(error) = self.error.take() {
                return Some(Err(error));
            }
            if self.done {
                return None;
            }
            if let Err(error) = self.refill() {
                self.done = true;
                self.error = Some(error);
            }
        }
    }
}