mito2/cache/index/
result_cache.rs

1// Copyright 2023 Greptime Team
2//
3// Licensed under the Apache License, Version 2.0 (the "License");
4// you may not use this file except in compliance with the License.
5// You may obtain a copy of the License at
6//
7//     http://www.apache.org/licenses/LICENSE-2.0
8//
9// Unless required by applicable law or agreed to in writing, software
10// distributed under the License is distributed on an "AS IS" BASIS,
11// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12// See the License for the specific language governing permissions and
13// limitations under the License.
14
15use std::collections::BTreeMap;
16use std::sync::Arc;
17
18use index::bloom_filter::applier::InListPredicate;
19use index::inverted_index::search::predicate::{Predicate, RangePredicate};
20use index::target::IndexTarget;
21use moka::notification::RemovalCause;
22use moka::sync::Cache;
23use store_api::storage::FileId;
24
25use crate::metrics::{CACHE_BYTES, CACHE_EVICTION, CACHE_HIT, CACHE_MISS};
26use crate::sst::index::fulltext_index::applier::builder::{
27    FulltextLike, FulltextQuery, FulltextRequest, FulltextTerm,
28};
29use crate::sst::parquet::row_selection::RowGroupSelection;
30
31const INDEX_RESULT_TYPE: &str = "index_result";
32
33/// Cache for storing index query results.
34///
35/// The `RowGroupSelection` is a collection of row groups that match the predicate.
36///
37/// Row groups can be partially searched. Row groups that not contained in `RowGroupSelection` are not searched.
38/// User can retrieve the partial results and handle uncontained row groups required by the predicate subsequently.
39pub struct IndexResultCache {
40    cache: Cache<(PredicateKey, FileId), Arc<RowGroupSelection>>,
41}
42
43impl IndexResultCache {
44    /// Creates a new cache with the given capacity.
45    pub fn new(capacity: u64) -> Self {
46        fn to_str(cause: RemovalCause) -> &'static str {
47            match cause {
48                RemovalCause::Expired => "expired",
49                RemovalCause::Explicit => "explicit",
50                RemovalCause::Replaced => "replaced",
51                RemovalCause::Size => "size",
52            }
53        }
54
55        let cache = Cache::builder()
56            .max_capacity(capacity)
57            .weigher(Self::index_result_cache_weight)
58            .eviction_listener(|k, v, cause| {
59                let size = Self::index_result_cache_weight(&k, &v);
60                CACHE_BYTES
61                    .with_label_values(&[INDEX_RESULT_TYPE])
62                    .sub(size.into());
63                CACHE_EVICTION
64                    .with_label_values(&[INDEX_RESULT_TYPE, to_str(cause)])
65                    .inc();
66            })
67            .support_invalidation_closures()
68            .build();
69        Self { cache }
70    }
71
72    /// Puts a query result into the cache.
73    ///
74    /// Allow user to put a partial result (not containing all row groups) into the cache.
75    pub fn put(&self, key: PredicateKey, file_id: FileId, result: Arc<RowGroupSelection>) {
76        let key = (key, file_id);
77        let size = Self::index_result_cache_weight(&key, &result);
78        CACHE_BYTES
79            .with_label_values(&[INDEX_RESULT_TYPE])
80            .add(size.into());
81        self.cache.insert(key, result);
82    }
83
84    /// Gets a query result from the cache.
85    ///
86    /// Note: the returned `RowGroupSelection` only contains the row groups that are searched.
87    ///       Caller should handle the uncontained row groups required by the predicate subsequently.
88    pub fn get(&self, key: &PredicateKey, file_id: FileId) -> Option<Arc<RowGroupSelection>> {
89        let res = self.cache.get(&(key.clone(), file_id));
90        if res.is_some() {
91            CACHE_HIT.with_label_values(&[INDEX_RESULT_TYPE]).inc();
92        } else {
93            CACHE_MISS.with_label_values(&[INDEX_RESULT_TYPE]).inc()
94        }
95        res
96    }
97
98    /// Calculates the memory usage of a cache entry.
99    fn index_result_cache_weight(k: &(PredicateKey, FileId), v: &Arc<RowGroupSelection>) -> u32 {
100        k.0.mem_usage() as u32 + v.mem_usage() as u32
101    }
102
103    /// Removes cached results for the given file.
104    pub fn invalidate_file(&self, file_id: FileId) {
105        self.cache
106            .invalidate_entries_if(move |(_, cached_file_id), _| *cached_file_id == file_id)
107            .expect("cache should support invalidation closures");
108    }
109}
110
111/// Key for different types of index predicates.
112#[derive(Debug, Clone, PartialEq, Eq, Hash)]
113pub enum PredicateKey {
114    /// Fulltext index predicate.
115    Fulltext(FulltextIndexKey),
116    /// Bloom filter predicate.
117    Bloom(BloomFilterKey),
118    /// Inverted index predicate.
119    Inverted(InvertedIndexKey),
120    /// Min-max pruning predicate.
121    MinMax(MinMaxKey),
122}
123
124impl PredicateKey {
125    /// Creates a new fulltext index key.
126    pub fn new_fulltext(predicates: Arc<BTreeMap<IndexTarget, FulltextRequest>>) -> Self {
127        Self::Fulltext(FulltextIndexKey::new(predicates))
128    }
129
130    /// Creates a new bloom filter key.
131    pub fn new_bloom(predicates: Arc<BTreeMap<IndexTarget, Vec<InListPredicate>>>) -> Self {
132        Self::Bloom(BloomFilterKey::new(predicates))
133    }
134
135    /// Creates a new inverted index key.
136    pub fn new_inverted(predicates: Arc<BTreeMap<IndexTarget, Vec<Predicate>>>) -> Self {
137        Self::Inverted(InvertedIndexKey::new(predicates))
138    }
139
140    /// Creates a new min-max pruning key.
141    pub fn new_minmax(exprs: Arc<Vec<String>>, schema_version: u64, skip_fields: bool) -> Self {
142        Self::MinMax(MinMaxKey::new(exprs, schema_version, skip_fields))
143    }
144
145    /// Returns the memory usage of this key.
146    pub fn mem_usage(&self) -> usize {
147        match self {
148            Self::Fulltext(key) => key.mem_usage,
149            Self::Bloom(key) => key.mem_usage,
150            Self::Inverted(key) => key.mem_usage,
151            Self::MinMax(key) => key.mem_usage,
152        }
153    }
154}
155
156/// Key for fulltext index queries.
157#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
158pub struct FulltextIndexKey {
159    predicates: Arc<BTreeMap<IndexTarget, FulltextRequest>>,
160    mem_usage: usize,
161}
162
163impl FulltextIndexKey {
164    /// Creates a new fulltext index key with the given predicates.
165    /// Calculates memory usage based on the size of queries and terms.
166    pub fn new(predicates: Arc<BTreeMap<IndexTarget, FulltextRequest>>) -> Self {
167        let mem_usage = predicates
168            .iter()
169            .map(|(target, request)| {
170                let target_size = target.approx_mem_size();
171                let query_size = request
172                    .queries
173                    .iter()
174                    .map(|query| query.0.len() + size_of::<FulltextQuery>())
175                    .sum::<usize>();
176                let term_size = request
177                    .terms
178                    .iter()
179                    .map(|term| term.term.len() + size_of::<FulltextTerm>())
180                    .sum::<usize>();
181                let like_size = request
182                    .likes
183                    .iter()
184                    .map(|like| like.pattern.len() + size_of::<FulltextLike>())
185                    .sum::<usize>();
186                target_size + query_size + term_size + like_size
187            })
188            .sum();
189        Self {
190            predicates,
191            mem_usage,
192        }
193    }
194}
195
196/// Key for bloom filter queries.
197#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
198pub struct BloomFilterKey {
199    predicates: Arc<BTreeMap<IndexTarget, Vec<InListPredicate>>>,
200    mem_usage: usize,
201}
202
203impl BloomFilterKey {
204    /// Creates a new bloom filter key with the given predicates.
205    /// Calculates memory usage based on the size of predicate lists.
206    pub fn new(predicates: Arc<BTreeMap<IndexTarget, Vec<InListPredicate>>>) -> Self {
207        let mem_usage = predicates
208            .iter()
209            .map(|(target, predicates)| {
210                target.approx_mem_size()
211                    + predicates
212                        .iter()
213                        .map(|predicate| {
214                            predicate.list.iter().map(|list| list.len()).sum::<usize>()
215                        })
216                        .sum::<usize>()
217            })
218            .sum();
219        Self {
220            predicates,
221            mem_usage,
222        }
223    }
224}
225
226/// Key for inverted index queries.
227#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
228pub struct InvertedIndexKey {
229    predicates: Arc<BTreeMap<IndexTarget, Vec<Predicate>>>,
230    mem_usage: usize,
231}
232
233impl InvertedIndexKey {
234    /// Creates a new inverted index key with the given predicates.
235    /// Calculates memory usage based on the type and size of predicates.
236    pub fn new(predicates: Arc<BTreeMap<IndexTarget, Vec<Predicate>>>) -> Self {
237        let mem_usage = predicates
238            .iter()
239            .map(|(target, predicates)| {
240                target.approx_mem_size()
241                    + predicates
242                        .iter()
243                        .map(|predicate| match predicate {
244                            Predicate::InList(predicate) => {
245                                predicate.list.iter().map(|list| list.len()).sum::<usize>()
246                            }
247                            Predicate::Range(_) => size_of::<RangePredicate>(),
248                            Predicate::RegexMatch(predicate) => predicate.pattern.len(),
249                        })
250                        .sum::<usize>()
251            })
252            .sum();
253
254        Self {
255            predicates,
256            mem_usage,
257        }
258    }
259}
260
261/// Key for min-max pruning.
262#[derive(Debug, Clone, PartialEq, Eq, Hash, Default)]
263pub struct MinMaxKey {
264    exprs: Arc<Vec<String>>,
265    schema_version: u64,
266    skip_fields: bool,
267    mem_usage: usize,
268}
269
270impl MinMaxKey {
271    pub fn new(exprs: Arc<Vec<String>>, schema_version: u64, skip_fields: bool) -> Self {
272        let mem_usage = size_of::<Self>()
273            + size_of::<Vec<String>>()
274            + exprs.len() * size_of::<String>()
275            + exprs.iter().map(|s| s.len()).sum::<usize>();
276        Self {
277            exprs,
278            schema_version,
279            skip_fields,
280            mem_usage,
281        }
282    }
283}
284
285#[cfg(test)]
286#[allow(clippy::single_range_in_vec_init)]
287mod tests {
288    use std::collections::{BTreeMap, BTreeSet};
289    use std::sync::Arc;
290
291    use index::bloom_filter::applier::InListPredicate as BloomInListPredicate;
292    use index::inverted_index::search::predicate::{Predicate, Range, RangePredicate};
293
294    use super::*;
295    use crate::sst::index::fulltext_index::applier::builder::{
296        FulltextQuery, FulltextRequest, FulltextTerm,
297    };
298    use crate::sst::parquet::row_selection::RowGroupSelection;
299
300    #[test]
301    fn test_inverted_cache_separates_json_paths() {
302        let key = |path: &str| {
303            PredicateKey::new_inverted(Arc::new(BTreeMap::from([(
304                IndexTarget::new_json_path(1, vec![path.into()]).unwrap(),
305                vec![Predicate::Range(RangePredicate {
306                    range: Range {
307                        lower: None,
308                        upper: None,
309                    },
310                })],
311            )])))
312        };
313        let a = key("a");
314        let b = key("b");
315        let cache = IndexResultCache::new(10000);
316        let file = FileId::random();
317        cache.put(a.clone(), file, Arc::new(RowGroupSelection::new(10, 10)));
318        assert!(cache.get(&a, file).is_some());
319        assert!(cache.get(&b, file).is_none());
320    }
321
322    #[test]
323    fn test_cache_basic_operations() {
324        let cache = IndexResultCache::new(1000);
325        let file_id = FileId::random();
326
327        // Create a test key and value
328        let predicates = BTreeMap::new();
329        let key = PredicateKey::new_fulltext(Arc::new(predicates));
330        let selection = Arc::new(RowGroupSelection::from_row_ids(
331            [1, 2, 3].into_iter().collect(),
332            1,
333            10,
334        ));
335
336        // Test put and get
337        cache.put(key.clone(), file_id, selection.clone());
338        let retrieved = cache.get(&key, file_id);
339        assert!(retrieved.is_some());
340        assert_eq!(
341            retrieved.unwrap().as_ref().row_count(),
342            selection.as_ref().row_count()
343        );
344
345        // Test get non-existent key
346        let non_existent_file_id = FileId::random();
347        assert!(cache.get(&key, non_existent_file_id).is_none());
348    }
349
350    #[test]
351    fn test_minmax_key_should_distinguish_schema_version_and_skip_fields() {
352        let exprs = Arc::new(vec!["col > 1".to_string()]);
353
354        let key1 = PredicateKey::new_minmax(exprs.clone(), 1, false);
355        let key2 = PredicateKey::new_minmax(exprs.clone(), 2, false);
356        assert_ne!(key1, key2);
357
358        let key3 = PredicateKey::new_minmax(exprs, 1, true);
359        assert_ne!(key1, key3);
360    }
361
362    #[test]
363    fn test_cache_capacity_limit() {
364        // Create a cache with small capacity (100 bytes)
365        let cache = IndexResultCache::new(100);
366        let file_id1 = FileId::random();
367        let file_id2 = FileId::random();
368
369        // Create two large keys that will exceed capacity
370        let mut predicates1 = BTreeMap::new();
371        let request1 = FulltextRequest {
372            queries: vec![
373                FulltextQuery(
374                    "test query 1 with a very long string to ensure large weight".to_string(),
375                ),
376                FulltextQuery("another long query string".to_string()),
377            ],
378            terms: vec![],
379            likes: vec![],
380        };
381        predicates1.insert(IndexTarget::ColumnId(1), request1);
382        let key1 = PredicateKey::new_fulltext(Arc::new(predicates1));
383        let selection1 = Arc::new(RowGroupSelection::default());
384
385        let mut predicates2 = BTreeMap::new();
386        let request2 = FulltextRequest {
387            queries: vec![
388                FulltextQuery(
389                    "test query 2 with a very long string to ensure large weight".to_string(),
390                ),
391                FulltextQuery("another long query string".to_string()),
392            ],
393            terms: vec![],
394            likes: vec![],
395        };
396        predicates2.insert(IndexTarget::ColumnId(1), request2);
397        let key2 = PredicateKey::new_fulltext(Arc::new(predicates2));
398        let selection2 = Arc::new(RowGroupSelection::default());
399
400        // Calculate weights
401        let weight1 =
402            IndexResultCache::index_result_cache_weight(&(key1.clone(), file_id1), &selection1);
403        let weight2 =
404            IndexResultCache::index_result_cache_weight(&(key2.clone(), file_id2), &selection2);
405        assert!(weight1 > 100);
406        assert!(weight2 > 100);
407
408        // Put first key-value pair
409        cache.put(key1.clone(), file_id1, selection1.clone());
410
411        // Verify first key is in cache
412        let retrieved1 = cache.get(&key1, file_id1);
413        assert!(retrieved1.is_some());
414        assert_eq!(
415            retrieved1.unwrap().as_ref().row_count(),
416            selection1.as_ref().row_count()
417        );
418
419        // Put second key-value pair, which should trigger eviction
420        cache.put(key2.clone(), file_id2, selection2.clone());
421
422        // Verify second key is in cache
423        let retrieved2 = cache.get(&key2, file_id2);
424        assert!(retrieved2.is_some());
425        assert_eq!(
426            retrieved2.unwrap().as_ref().row_count(),
427            selection2.as_ref().row_count()
428        );
429
430        // Verify first key was evicted
431        cache.cache.run_pending_tasks();
432        let retrieved1_after_eviction = cache.get(&key1, file_id1);
433        assert!(
434            retrieved1_after_eviction.is_none(),
435            "First key should have been evicted"
436        );
437    }
438
439    #[test]
440    fn test_index_result_cache_weight() {
441        let file_id = FileId::random();
442
443        // Test empty values
444        let empty_predicates = BTreeMap::new();
445        let empty_key = PredicateKey::new_fulltext(Arc::new(empty_predicates));
446        let empty_selection = Arc::new(RowGroupSelection::default());
447        let empty_weight = IndexResultCache::index_result_cache_weight(
448            &(empty_key.clone(), file_id),
449            &empty_selection,
450        );
451        assert_eq!(empty_weight, 0);
452        assert_eq!(
453            empty_weight,
454            empty_key.mem_usage() as u32 + empty_selection.mem_usage() as u32
455        );
456
457        // Test 1: FulltextIndexKey
458        let mut predicates1 = BTreeMap::new();
459        let request1 = FulltextRequest {
460            queries: vec![FulltextQuery("test query".to_string())],
461            terms: vec![FulltextTerm {
462                col_lowered: false,
463                term: "test term".to_string(),
464            }],
465            likes: vec![],
466        };
467        predicates1.insert(IndexTarget::ColumnId(1), request1);
468        let key1 = PredicateKey::new_fulltext(Arc::new(predicates1));
469        let selection1 = Arc::new(RowGroupSelection::new(100, 250));
470        let weight1 =
471            IndexResultCache::index_result_cache_weight(&(key1.clone(), file_id), &selection1);
472        assert!(weight1 > 0);
473        assert_eq!(
474            weight1,
475            key1.mem_usage() as u32 + selection1.mem_usage() as u32
476        );
477
478        // Test 2: BloomFilterKey
479        let mut predicates2 = BTreeMap::new();
480        let predicate2 = BloomInListPredicate {
481            list: BTreeSet::from([b"test1".to_vec(), b"test2".to_vec()]),
482        };
483        predicates2.insert(IndexTarget::ColumnId(1), vec![predicate2]);
484        let key2 = PredicateKey::new_bloom(Arc::new(predicates2));
485        let selection2 = Arc::new(RowGroupSelection::from_row_ids(
486            [1, 2, 3].into_iter().collect(),
487            100,
488            1,
489        ));
490        let weight2 =
491            IndexResultCache::index_result_cache_weight(&(key2.clone(), file_id), &selection2);
492        assert!(weight2 > 0);
493        assert_eq!(
494            weight2,
495            key2.mem_usage() as u32 + selection2.mem_usage() as u32
496        );
497
498        // Test 3: InvertedIndexKey
499        let mut predicates3 = BTreeMap::new();
500        let predicate3 = Predicate::Range(RangePredicate {
501            range: Range {
502                lower: None,
503                upper: None,
504            },
505        });
506        predicates3.insert(IndexTarget::ColumnId(1), vec![predicate3]);
507        let key3 = PredicateKey::new_inverted(Arc::new(predicates3));
508        let selection3 = Arc::new(RowGroupSelection::from_row_ranges(
509            vec![(0, vec![5..15])],
510            20,
511        ));
512        let weight3 =
513            IndexResultCache::index_result_cache_weight(&(key3.clone(), file_id), &selection3);
514        assert!(weight3 > 0);
515        assert_eq!(
516            weight3,
517            key3.mem_usage() as u32 + selection3.mem_usage() as u32
518        );
519    }
520}
mito2/cache/index/result_cache.rs

mito2/cache/index/
result_cache.rs
