lsm_tree/blob_tree/
mod.rs

1// Copyright (c) 2024-present, fjall-rs
2// This source code is licensed under both the Apache 2.0 and MIT License
3// (found in the LICENSE-* files in the repository)
4
5mod gc;
6pub mod handle;
7pub mod ingest;
8
9#[doc(hidden)]
10pub use gc::{FragmentationEntry, FragmentationMap};
11
12use crate::{
13    coding::Decode,
14    file::{fsync_directory, BLOBS_FOLDER},
15    iter_guard::{IterGuard, IterGuardImpl},
16    r#abstract::{AbstractTree, RangeItem},
17    table::Table,
18    tree::inner::MemtableId,
19    value::InternalValue,
20    version::Version,
21    vlog::{Accessor, BlobFile, BlobFileWriter, ValueHandle},
22    Cache, Config, DescriptorTable, Memtable, SeqNo, TableId, TreeId, UserKey, UserValue,
23};
24use handle::BlobIndirection;
25use std::{
26    io::Cursor,
27    ops::RangeBounds,
28    path::PathBuf,
29    sync::{Arc, MutexGuard},
30};
31
32/// Iterator value guard
33pub struct Guard {
34    tree: crate::BlobTree,
35    version: Version,
36    kv: crate::Result<InternalValue>,
37}
38
39impl IterGuard for Guard {
40    fn into_inner_if(
41        self,
42        pred: impl Fn(&UserKey) -> bool,
43    ) -> crate::Result<(UserKey, Option<UserValue>)> {
44        let kv = self.kv?;
45
46        if pred(&kv.key.user_key) {
47            resolve_value_handle(
48                self.tree.id(),
49                self.tree.blobs_folder.as_path(),
50                &self.tree.index.config.cache,
51                &self.tree.index.config.descriptor_table,
52                &self.version,
53                kv,
54            )
55            .map(|(k, v)| (k, Some(v)))
56        } else {
57            Ok((kv.key.user_key, None))
58        }
59    }
60
61    fn key(self) -> crate::Result<UserKey> {
62        self.kv.map(|kv| kv.key.user_key)
63    }
64
65    fn size(self) -> crate::Result<u32> {
66        let kv = self.kv?;
67
68        if kv.key.value_type.is_indirection() {
69            let mut cursor = Cursor::new(kv.value);
70            Ok(BlobIndirection::decode_from(&mut cursor)?.size)
71        } else {
72            #[expect(clippy::cast_possible_truncation, reason = "values are u32 max length")]
73            Ok(kv.value.len() as u32)
74        }
75    }
76
77    fn into_inner(self) -> crate::Result<(UserKey, UserValue)> {
78        resolve_value_handle(
79            self.tree.id(),
80            self.tree.blobs_folder.as_path(),
81            &self.tree.index.config.cache,
82            &self.tree.index.config.descriptor_table,
83            &self.version,
84            self.kv?,
85        )
86    }
87}
88
89fn resolve_value_handle(
90    tree_id: TreeId,
91    blobs_folder: &std::path::Path,
92    cache: &Arc<Cache>,
93    descriptor_table: &Arc<DescriptorTable>,
94    version: &Version,
95    item: InternalValue,
96) -> RangeItem {
97    if item.key.value_type.is_indirection() {
98        let mut cursor = Cursor::new(item.value);
99        let vptr = BlobIndirection::decode_from(&mut cursor)?;
100
101        // Resolve indirection using value log
102        match Accessor::new(&version.blob_files).get(
103            tree_id,
104            blobs_folder,
105            &item.key.user_key,
106            &vptr.vhandle,
107            cache,
108            descriptor_table,
109        ) {
110            Ok(Some(v)) => {
111                let k = item.key.user_key;
112                Ok((k, v))
113            }
114            Ok(None) => {
115                panic!(
116                    "value handle ({:?} => {:?}) did not match any blob - this is a bug; version={}",
117                    item.key.user_key, vptr.vhandle,
118                    version.id(),
119                );
120            }
121            Err(e) => Err(e),
122        }
123    } else {
124        let k = item.key.user_key;
125        let v = item.value;
126        Ok((k, v))
127    }
128}
129
130/// A key-value-separated log-structured merge tree
131///
132/// This tree is a composite structure, consisting of an
133/// index tree (LSM-tree) and a log-structured value log
134/// to reduce write amplification.
135#[derive(Clone)]
136pub struct BlobTree {
137    /// Index tree that holds value handles or small inline values
138    #[doc(hidden)]
139    pub index: crate::Tree,
140
141    blobs_folder: Arc<PathBuf>,
142}
143
144impl BlobTree {
145    pub(crate) fn open(config: Config) -> crate::Result<Self> {
146        let index = crate::Tree::open(config)?;
147
148        let blobs_folder = index.config.path.join(BLOBS_FOLDER);
149        std::fs::create_dir_all(&blobs_folder)?;
150        fsync_directory(&blobs_folder)?;
151
152        let blob_file_id_to_continue_with = index
153            .current_version()
154            .blob_files
155            .list_ids()
156            .max()
157            .map(|x| x + 1)
158            .unwrap_or_default();
159
160        index
161            .0
162            .blob_file_id_counter
163            .set(blob_file_id_to_continue_with);
164
165        Ok(Self {
166            index,
167            blobs_folder: Arc::new(blobs_folder),
168        })
169    }
170}
171
172impl AbstractTree for BlobTree {
173    fn table_file_cache_size(&self) -> usize {
174        self.index.table_file_cache_size()
175    }
176
177    fn get_version_history_lock(
178        &self,
179    ) -> std::sync::RwLockWriteGuard<'_, crate::version::SuperVersions> {
180        self.index.get_version_history_lock()
181    }
182
183    fn next_table_id(&self) -> TableId {
184        self.index.next_table_id()
185    }
186
187    fn id(&self) -> crate::TreeId {
188        self.index.id()
189    }
190
191    fn get_internal_entry(&self, key: &[u8], seqno: SeqNo) -> crate::Result<Option<InternalValue>> {
192        self.index.get_internal_entry(key, seqno)
193    }
194
195    fn current_version(&self) -> Version {
196        self.index.current_version()
197    }
198
199    #[cfg(feature = "metrics")]
200    fn metrics(&self) -> &Arc<crate::Metrics> {
201        self.index.metrics()
202    }
203
204    fn version_free_list_len(&self) -> usize {
205        self.index.version_free_list_len()
206    }
207
208    fn prefix<K: AsRef<[u8]>>(
209        &self,
210        prefix: K,
211        seqno: SeqNo,
212        index: Option<(Arc<Memtable>, SeqNo)>,
213    ) -> Box<dyn DoubleEndedIterator<Item = IterGuardImpl> + Send + 'static> {
214        use crate::range::prefix_to_range;
215
216        let super_version = self.index.get_version_for_snapshot(seqno);
217        let tree = self.clone();
218
219        let range = prefix_to_range(prefix.as_ref());
220
221        Box::new(
222            crate::Tree::create_internal_range(super_version.clone(), &range, seqno, index).map(
223                move |kv| {
224                    IterGuardImpl::Blob(Guard {
225                        tree: tree.clone(),
226                        version: super_version.version.clone(),
227                        kv,
228                    })
229                },
230            ),
231        )
232    }
233
234    fn range<K: AsRef<[u8]>, R: RangeBounds<K>>(
235        &self,
236        range: R,
237        seqno: SeqNo,
238        index: Option<(Arc<Memtable>, SeqNo)>,
239    ) -> Box<dyn DoubleEndedIterator<Item = IterGuardImpl> + Send + 'static> {
240        let super_version = self.index.get_version_for_snapshot(seqno);
241        let tree = self.clone();
242
243        Box::new(
244            crate::Tree::create_internal_range(super_version.clone(), &range, seqno, index).map(
245                move |kv| {
246                    IterGuardImpl::Blob(Guard {
247                        tree: tree.clone(),
248                        version: super_version.version.clone(),
249                        kv,
250                    })
251                },
252            ),
253        )
254    }
255
256    fn tombstone_count(&self) -> u64 {
257        self.index.tombstone_count()
258    }
259
260    fn weak_tombstone_count(&self) -> u64 {
261        self.index.weak_tombstone_count()
262    }
263
264    fn weak_tombstone_reclaimable_count(&self) -> u64 {
265        self.index.weak_tombstone_reclaimable_count()
266    }
267
268    fn drop_range<K: AsRef<[u8]>, R: RangeBounds<K>>(&self, range: R) -> crate::Result<()> {
269        self.index.drop_range(range)
270    }
271
272    fn clear(&self) -> crate::Result<()> {
273        self.index.clear()
274    }
275
276    fn major_compact(&self, target_size: u64, seqno_threshold: SeqNo) -> crate::Result<()> {
277        self.index.major_compact(target_size, seqno_threshold)
278    }
279
280    fn clear_active_memtable(&self) {
281        self.index.clear_active_memtable();
282    }
283
284    fn l0_run_count(&self) -> usize {
285        self.index.l0_run_count()
286    }
287
288    fn blob_file_count(&self) -> usize {
289        self.current_version().blob_file_count()
290    }
291
292    // NOTE: We skip reading from the value log
293    // because the vHandles already store the value size
294    fn size_of<K: AsRef<[u8]>>(&self, key: K, seqno: SeqNo) -> crate::Result<Option<u32>> {
295        let Some(item) = self.index.get_internal_entry(key.as_ref(), seqno)? else {
296            return Ok(None);
297        };
298
299        Ok(Some(if item.key.value_type.is_indirection() {
300            let mut cursor = Cursor::new(item.value);
301            let vptr = BlobIndirection::decode_from(&mut cursor)?;
302            vptr.size
303        } else {
304            #[expect(clippy::cast_possible_truncation, reason = "values are u32 length max")]
305            {
306                item.value.len() as u32
307            }
308        }))
309    }
310
311    fn stale_blob_bytes(&self) -> u64 {
312        self.current_version().gc_stats().stale_bytes()
313    }
314
315    fn filter_size(&self) -> u64 {
316        self.index.filter_size()
317    }
318
319    fn pinned_filter_size(&self) -> usize {
320        self.index.pinned_filter_size()
321    }
322
323    fn pinned_block_index_size(&self) -> usize {
324        self.index.pinned_block_index_size()
325    }
326
327    fn sealed_memtable_count(&self) -> usize {
328        self.index.sealed_memtable_count()
329    }
330
331    fn get_flush_lock(&self) -> MutexGuard<'_, ()> {
332        self.index.get_flush_lock()
333    }
334
335    fn flush_to_tables(
336        &self,
337        stream: impl Iterator<Item = crate::Result<InternalValue>>,
338    ) -> crate::Result<Option<(Vec<Table>, Option<Vec<BlobFile>>)>> {
339        use crate::{coding::Encode, file::TABLES_FOLDER, table::multi_writer::MultiWriter};
340
341        let start = std::time::Instant::now();
342
343        let table_folder = self.index.config.path.join(TABLES_FOLDER);
344
345        let data_block_size = self.index.config.data_block_size_policy.get(0);
346
347        let data_block_restart_interval =
348            self.index.config.data_block_restart_interval_policy.get(0);
349        let index_block_restart_interval =
350            self.index.config.index_block_restart_interval_policy.get(0);
351
352        let data_block_compression = self.index.config.data_block_compression_policy.get(0);
353        let index_block_compression = self.index.config.index_block_compression_policy.get(0);
354
355        let data_block_hash_ratio = self.index.config.data_block_hash_ratio_policy.get(0);
356
357        let index_partitioning = self.index.config.index_block_partitioning_policy.get(0);
358        let filter_partitioning = self.index.config.filter_block_partitioning_policy.get(0);
359
360        log::debug!("Flushing memtable(s) and performing key-value separation, data_block_restart_interval={data_block_restart_interval}, index_block_restart_interval={index_block_restart_interval}, data_block_size={data_block_size}, data_block_compression={data_block_compression}, index_block_compression={index_block_compression}");
361        log::debug!("=> to table(s) in {}", table_folder.display());
362        log::debug!("=> to blob file(s) at {}", self.blobs_folder.display());
363
364        let mut table_writer = MultiWriter::new(
365            table_folder.clone(),
366            self.index.table_id_counter.clone(),
367            64 * 1_024 * 1_024,
368            0,
369        )?
370        .use_data_block_restart_interval(data_block_restart_interval)
371        .use_index_block_restart_interval(index_block_restart_interval)
372        .use_data_block_compression(data_block_compression)
373        .use_index_block_compression(index_block_compression)
374        .use_data_block_size(data_block_size)
375        .use_data_block_hash_ratio(data_block_hash_ratio)
376        .use_bloom_policy({
377            use crate::config::FilterPolicyEntry::{Bloom, None};
378            use crate::table::filter::BloomConstructionPolicy;
379
380            match self.index.config.filter_policy.get(0) {
381                Bloom(policy) => policy,
382                None => BloomConstructionPolicy::BitsPerKey(0.0),
383            }
384        });
385
386        if index_partitioning {
387            table_writer = table_writer.use_partitioned_index();
388        }
389        if filter_partitioning {
390            table_writer = table_writer.use_partitioned_filter();
391        }
392
393        #[expect(
394            clippy::expect_used,
395            reason = "cannot create blob tree without defining kv separation options"
396        )]
397        let kv_opts = self
398            .index
399            .config
400            .kv_separation_opts
401            .as_ref()
402            .expect("kv separation options should exist");
403
404        let mut blob_writer = BlobFileWriter::new(
405            self.index.0.blob_file_id_counter.clone(),
406            self.index.config.path.join(BLOBS_FOLDER),
407        )?
408        .use_target_size(kv_opts.file_target_size)
409        .use_compression(kv_opts.compression);
410
411        let separation_threshold = kv_opts.separation_threshold;
412
413        for item in stream {
414            let item = item?;
415
416            if item.is_tombstone() {
417                // NOTE: Still need to add tombstone to index tree
418                // But no blob to blob writer
419                table_writer.write(InternalValue::new(item.key, UserValue::empty()))?;
420                continue;
421            }
422
423            let value = item.value;
424
425            #[expect(clippy::cast_possible_truncation, reason = "values are u32 length max")]
426            let value_size = value.len() as u32;
427
428            if value_size >= separation_threshold {
429                let offset = blob_writer.offset();
430                let blob_file_id = blob_writer.blob_file_id();
431                let on_disk_size = blob_writer.write(&item.key.user_key, item.key.seqno, &value)?;
432
433                let indirection = BlobIndirection {
434                    vhandle: ValueHandle {
435                        blob_file_id,
436                        offset,
437                        on_disk_size,
438                    },
439                    size: value_size,
440                };
441
442                table_writer.write({
443                    let mut vptr =
444                        InternalValue::new(item.key.clone(), indirection.encode_into_vec());
445                    vptr.key.value_type = crate::ValueType::Indirection;
446                    vptr
447                })?;
448
449                table_writer.register_blob(indirection);
450            } else {
451                table_writer.write(InternalValue::new(item.key, value))?;
452            }
453        }
454
455        let blob_files = blob_writer.finish()?;
456
457        let result = table_writer.finish()?;
458
459        log::debug!("Flushed memtable(s) in {:?}", start.elapsed());
460
461        let pin_filter = self.index.config.filter_block_pinning_policy.get(0);
462        let pin_index = self.index.config.index_block_pinning_policy.get(0);
463
464        // Load tables
465        let tables = result
466            .into_iter()
467            .map(|(table_id, checksum)| -> crate::Result<Table> {
468                Table::recover(
469                    table_folder.join(table_id.to_string()),
470                    checksum,
471                    0,
472                    self.index.id,
473                    self.index.config.cache.clone(),
474                    self.index.config.descriptor_table.clone(),
475                    pin_filter,
476                    pin_index,
477                    #[cfg(feature = "metrics")]
478                    self.index.metrics.clone(),
479                )
480            })
481            .collect::<crate::Result<Vec<_>>>()?;
482
483        Ok(Some((tables, Some(blob_files))))
484    }
485
486    fn register_tables(
487        &self,
488        tables: &[Table],
489        blob_files: Option<&[BlobFile]>,
490        frag_map: Option<FragmentationMap>,
491        sealed_memtables_to_delete: &[MemtableId],
492        gc_watermark: SeqNo,
493    ) -> crate::Result<()> {
494        self.index.register_tables(
495            tables,
496            blob_files,
497            frag_map,
498            sealed_memtables_to_delete,
499            gc_watermark,
500        )
501    }
502
503    fn compact(
504        &self,
505        strategy: Arc<dyn crate::compaction::CompactionStrategy>,
506        seqno_threshold: SeqNo,
507    ) -> crate::Result<()> {
508        self.index.compact(strategy, seqno_threshold)
509    }
510
511    fn get_next_table_id(&self) -> TableId {
512        self.index.get_next_table_id()
513    }
514
515    fn tree_config(&self) -> &Config {
516        &self.index.config
517    }
518
519    fn get_highest_seqno(&self) -> Option<SeqNo> {
520        self.index.get_highest_seqno()
521    }
522
523    fn active_memtable(&self) -> Arc<Memtable> {
524        self.index.active_memtable()
525    }
526
527    fn tree_type(&self) -> crate::TreeType {
528        crate::TreeType::Blob
529    }
530
531    fn rotate_memtable(&self) -> Option<Arc<Memtable>> {
532        self.index.rotate_memtable()
533    }
534
535    fn table_count(&self) -> usize {
536        self.index.table_count()
537    }
538
539    fn level_table_count(&self, idx: usize) -> Option<usize> {
540        self.index.level_table_count(idx)
541    }
542
543    fn approximate_len(&self) -> usize {
544        self.index.approximate_len()
545    }
546
547    // NOTE: Override the default implementation to not fetch
548    // data from the value log, so we get much faster key reads
549    fn is_empty(&self, seqno: SeqNo, index: Option<(Arc<Memtable>, SeqNo)>) -> crate::Result<bool> {
550        self.index.is_empty(seqno, index)
551    }
552
553    // NOTE: Override the default implementation to not fetch
554    // data from the value log, so we get much faster key reads
555    fn contains_key<K: AsRef<[u8]>>(&self, key: K, seqno: SeqNo) -> crate::Result<bool> {
556        self.index.contains_key(key, seqno)
557    }
558
559    // NOTE: Override the default implementation to not fetch
560    // data from the value log, so we get much faster scans
561    fn len(&self, seqno: SeqNo, index: Option<(Arc<Memtable>, SeqNo)>) -> crate::Result<usize> {
562        self.index.len(seqno, index)
563    }
564
565    fn disk_space(&self) -> u64 {
566        let version = self.current_version();
567        self.index.disk_space() + version.blob_files.on_disk_size()
568    }
569
570    fn get_highest_memtable_seqno(&self) -> Option<SeqNo> {
571        self.index.get_highest_memtable_seqno()
572    }
573
574    fn get_highest_persisted_seqno(&self) -> Option<SeqNo> {
575        self.index.get_highest_persisted_seqno()
576    }
577
578    fn insert<K: Into<UserKey>, V: Into<UserValue>>(
579        &self,
580        key: K,
581        value: V,
582        seqno: SeqNo,
583    ) -> (u64, u64) {
584        self.index.insert(key, value.into(), seqno)
585    }
586
587    fn get<K: AsRef<[u8]>>(&self, key: K, seqno: SeqNo) -> crate::Result<Option<crate::UserValue>> {
588        let key = key.as_ref();
589
590        #[expect(clippy::expect_used, reason = "lock is expected to not be poisoned")]
591        let super_version = self
592            .index
593            .version_history
594            .read()
595            .expect("lock is poisoned")
596            .get_version_for_snapshot(seqno);
597
598        let Some(item) = crate::Tree::get_internal_entry_from_version(&super_version, key, seqno)?
599        else {
600            return Ok(None);
601        };
602
603        let (_, v) = resolve_value_handle(
604            self.id(),
605            self.blobs_folder.as_path(),
606            &self.index.config.cache,
607            &self.index.config.descriptor_table,
608            &super_version.version,
609            item,
610        )?;
611
612        Ok(Some(v))
613    }
614
615    fn remove<K: Into<UserKey>>(&self, key: K, seqno: SeqNo) -> (u64, u64) {
616        self.index.remove(key, seqno)
617    }
618
619    fn remove_weak<K: Into<UserKey>>(&self, key: K, seqno: SeqNo) -> (u64, u64) {
620        self.index.remove_weak(key, seqno)
621    }
622}