iqdb_cache/cached.rs
1//! The [`CachedIndex`] wrapper.
2
3use core::time::Duration;
4use std::sync::Arc;
5use std::sync::Mutex;
6use std::sync::atomic::{AtomicU64, Ordering};
7
8use clock_lib::{Clock, Monotonic, SystemClock};
9use iqdb_index::{IndexCore, IndexStats};
10use iqdb_types::{DistanceMetric, Hit, Metadata, Result, SearchParams, VectorId};
11
12use crate::config::{CacheConfig, EvictionPolicy};
13use crate::key::ResultKey;
14use crate::policy::PolicyCache;
15use crate::stats::CacheStats;
16
17/// A cached search result plus the moment it was stored.
18///
19/// `stamp` is recorded only when a TTL is configured; with no TTL it is `None`
20/// and the entry never expires on time.
21struct CacheEntry {
22 /// The memoized hits, ready to clone out on a hit.
23 hits: Box<[Hit]>,
24 /// When the entry was written, for TTL expiry; `None` when no TTL applies.
25 stamp: Option<Monotonic>,
26}
27
28/// A drop-in [`IndexCore`] wrapper that memoizes search results.
29///
30/// `CachedIndex` holds any `I: IndexCore` and forwards every call to it, with
31/// one addition: identical [`search`](IndexCore::search) calls — same query
32/// and same [`SearchParams`] — are served from an in-memory LRU cache instead
33/// of re-running the search. Because it *is* an [`IndexCore`], it slots in
34/// anywhere the wrapped index does, including behind `Box<dyn IndexCore>`.
35///
36/// ## Correctness
37///
38/// The cache never returns a stale result. Every mutation that can change the
39/// search space — [`insert`](IndexCore::insert),
40/// [`insert_batch`](IndexCore::insert_batch), and
41/// [`delete`](IndexCore::delete) — invalidates the cache, so a search after a
42/// write always recomputes against the current index. Operations that do not
43/// change the result set ([`flush`](IndexCore::flush) and the read-only
44/// accessors) leave the cache intact.
45///
46/// ## Opt-in
47///
48/// Caching is an optimization a caller chooses by wrapping an index; the
49/// database leaves indexes unwrapped by default. Construct a cache that holds
50/// a fixed number of recent searches with [`new`](CachedIndex::new) or
51/// [`with_capacity`](CachedIndex::with_capacity), or tune it through a
52/// [`CacheConfig`] with [`with_config`](CachedIndex::with_config). A capacity of
53/// `0` disables caching entirely: every search passes straight through, which is
54/// useful for A/B measuring the cache's effect without changing call sites.
55///
56/// ## Time-to-live
57///
58/// A [`CacheConfig::ttl`] gives entries an expiry: a cached result older than
59/// the TTL is treated as a miss and recomputed. Mutations through this wrapper
60/// already invalidate exactly, so the TTL exists to bound staleness from changes
61/// the wrapper *cannot* see — for example, the wrapped index mutated through
62/// another handle. With no TTL (the default) the clock is never consulted.
63///
64/// ## Concurrency
65///
66/// `CachedIndex<I>` is `Send + Sync` whenever `I` is (which every `IndexCore`
67/// is). Reads share the cache behind a [`Mutex`] held only for the lookup and
68/// the insert — never across the wrapped search — so concurrent misses run the
69/// underlying search in parallel rather than serializing on the lock.
70///
71/// # Examples
72///
73/// ```
74/// use std::sync::Arc;
75///
76/// use iqdb_cache::CachedIndex;
77/// use iqdb_index::{Index, IndexCore, IndexStats};
78/// use iqdb_types::{DistanceMetric, Hit, IqdbError, Metadata, Result, SearchParams, VectorId};
79///
80/// // A minimal index that returns one hit per search; enough to show the wrap.
81/// struct Stub {
82/// dim: usize,
83/// metric: DistanceMetric,
84/// ids: Vec<VectorId>,
85/// }
86///
87/// impl IndexCore for Stub {
88/// fn insert(&mut self, id: VectorId, _v: Arc<[f32]>, _m: Option<Metadata>) -> Result<()> {
89/// self.ids.push(id);
90/// Ok(())
91/// }
92/// fn delete(&mut self, id: &VectorId) -> Result<()> {
93/// match self.ids.iter().position(|x| x == id) {
94/// Some(pos) => { let _ = self.ids.remove(pos); Ok(()) }
95/// None => Err(IqdbError::NotFound),
96/// }
97/// }
98/// fn search(&self, _q: &[f32], params: &SearchParams) -> Result<Vec<Hit>> {
99/// Ok(self.ids.iter().take(params.k).map(|id| Hit::new(id.clone(), 0.0)).collect())
100/// }
101/// fn len(&self) -> usize { self.ids.len() }
102/// fn dim(&self) -> usize { self.dim }
103/// fn metric(&self) -> DistanceMetric { self.metric }
104/// fn flush(&mut self) -> Result<()> { Ok(()) }
105/// fn stats(&self) -> IndexStats {
106/// IndexStats { n_vectors: self.ids.len(), index_type: "stub", ..IndexStats::default() }
107/// }
108/// }
109///
110/// # fn main() -> Result<()> {
111/// let stub = Stub { dim: 3, metric: DistanceMetric::Cosine, ids: vec![VectorId::from(1u64)] };
112/// let mut cached = CachedIndex::new(stub);
113///
114/// let params = SearchParams::new(1, DistanceMetric::Cosine);
115/// let first = cached.search(&[1.0, 0.0, 0.0], ¶ms)?; // miss: runs the search
116/// let again = cached.search(&[1.0, 0.0, 0.0], ¶ms)?; // hit: served from cache
117/// assert_eq!(first, again);
118///
119/// let stats = cached.cache_stats();
120/// assert_eq!(stats.hits, 1);
121/// assert_eq!(stats.misses, 1);
122/// # Ok(())
123/// # }
124/// ```
125pub struct CachedIndex<I> {
126 /// The wrapped index every call forwards to.
127 inner: I,
128 /// The result cache, guarded for `&self` search access.
129 cache: Mutex<PolicyCache<ResultKey, CacheEntry>>,
130 /// Configured capacity, mirrored here for `0`-means-disabled fast paths.
131 capacity: usize,
132 /// Configured eviction policy, mirrored here for introspection.
133 policy: EvictionPolicy,
134 /// Optional per-entry time-to-live; `None` means entries expire only on
135 /// mutation.
136 ttl: Option<Duration>,
137 /// Time source for TTL expiry. `SystemClock` in production; a mock clock is
138 /// injected in tests. Only read when `ttl` is `Some`.
139 clock: Arc<dyn Clock>,
140 /// Lifetime count of cache hits.
141 hits: AtomicU64,
142 /// Lifetime count of cache misses.
143 misses: AtomicU64,
144 /// Lifetime count of entries discarded by the eviction policy.
145 evictions: AtomicU64,
146}
147
148impl<I: IndexCore> CachedIndex<I> {
149 /// Wraps `inner` with a result cache of the default capacity (1024 recent
150 /// searches) and no TTL.
151 ///
152 /// # Examples
153 ///
154 /// ```
155 /// # use iqdb_cache::CachedIndex;
156 /// # use iqdb_cache::doc_stub::stub_index;
157 /// let cached = CachedIndex::new(stub_index());
158 /// assert!(cached.is_enabled());
159 /// ```
160 #[must_use]
161 pub fn new(inner: I) -> Self {
162 Self::with_config(inner, CacheConfig::new())
163 }
164
165 /// Wraps `inner` with a result cache that holds at most `capacity` recent
166 /// searches and no TTL.
167 ///
168 /// A `capacity` of `0` disables caching: searches pass straight through and
169 /// nothing is stored.
170 ///
171 /// # Examples
172 ///
173 /// ```
174 /// # use iqdb_cache::CachedIndex;
175 /// # use iqdb_cache::doc_stub::stub_index;
176 /// let cached = CachedIndex::with_capacity(stub_index(), 256);
177 /// assert_eq!(cached.capacity(), 256);
178 ///
179 /// let bypass = CachedIndex::with_capacity(stub_index(), 0);
180 /// assert!(!bypass.is_enabled());
181 /// ```
182 #[must_use]
183 pub fn with_capacity(inner: I, capacity: usize) -> Self {
184 Self::with_config(inner, CacheConfig::new().capacity(capacity))
185 }
186
187 /// Wraps `inner` with a result cache built from `config` (the Tier-2 path).
188 ///
189 /// Use [`CacheConfig`] to set the capacity and an optional TTL together.
190 ///
191 /// # Examples
192 ///
193 /// ```
194 /// use std::time::Duration;
195 ///
196 /// use iqdb_cache::{CacheConfig, CachedIndex};
197 /// # use iqdb_cache::doc_stub::stub_index;
198 /// let config = CacheConfig::new().capacity(512).ttl(Duration::from_secs(30));
199 /// let cached = CachedIndex::with_config(stub_index(), config);
200 /// assert_eq!(cached.capacity(), 512);
201 /// assert_eq!(cached.ttl(), Some(Duration::from_secs(30)));
202 /// ```
203 #[must_use]
204 pub fn with_config(inner: I, config: CacheConfig) -> Self {
205 Self::with_config_in(inner, config, Arc::new(SystemClock::new()))
206 }
207
208 /// Construction core shared by every public constructor, with an injectable
209 /// clock for deterministic TTL tests.
210 pub(crate) fn with_config_in(inner: I, config: CacheConfig, clock: Arc<dyn Clock>) -> Self {
211 Self {
212 inner,
213 cache: Mutex::new(PolicyCache::new(config.policy, config.capacity)),
214 capacity: config.capacity,
215 policy: config.policy,
216 ttl: config.ttl,
217 clock,
218 hits: AtomicU64::new(0),
219 misses: AtomicU64::new(0),
220 evictions: AtomicU64::new(0),
221 }
222 }
223
224 /// The configured cache capacity. `0` means caching is disabled.
225 #[inline]
226 #[must_use]
227 pub fn capacity(&self) -> usize {
228 self.capacity
229 }
230
231 /// The configured per-entry time-to-live, or `None` if results expire only
232 /// on mutation.
233 #[inline]
234 #[must_use]
235 pub fn ttl(&self) -> Option<Duration> {
236 self.ttl
237 }
238
239 /// The configured eviction policy.
240 #[inline]
241 #[must_use]
242 pub fn policy(&self) -> EvictionPolicy {
243 self.policy
244 }
245
246 /// Whether caching is active (`capacity > 0`).
247 #[inline]
248 #[must_use]
249 pub fn is_enabled(&self) -> bool {
250 self.capacity > 0
251 }
252
253 /// Borrows the wrapped index.
254 #[inline]
255 #[must_use]
256 pub fn get_ref(&self) -> &I {
257 &self.inner
258 }
259
260 /// Unwraps the cache, returning the index it held.
261 ///
262 /// # Examples
263 ///
264 /// ```
265 /// # use iqdb_cache::CachedIndex;
266 /// # use iqdb_cache::doc_stub::stub_index;
267 /// # use iqdb_index::IndexCore;
268 /// let cached = CachedIndex::new(stub_index());
269 /// let inner = cached.into_inner();
270 /// assert_eq!(inner.dim(), 3);
271 /// ```
272 #[must_use]
273 pub fn into_inner(self) -> I {
274 self.inner
275 }
276
277 /// Drops every cached result, keeping the wrapped index untouched.
278 ///
279 /// Mutations invalidate automatically; call this only to force a cold cache
280 /// (for example, after the wrapped index was changed through a handle other
281 /// than this wrapper).
282 pub fn clear_cache(&mut self) {
283 match self.cache.get_mut() {
284 Ok(cache) => cache.clear(),
285 Err(poisoned) => poisoned.into_inner().clear(),
286 }
287 }
288
289 /// A snapshot of the cache's hit/miss counters and occupancy.
290 #[must_use]
291 pub fn cache_stats(&self) -> CacheStats {
292 let len = self.lock_cache().len();
293 CacheStats {
294 hits: self.hits.load(Ordering::Relaxed),
295 misses: self.misses.load(Ordering::Relaxed),
296 evictions: self.evictions.load(Ordering::Relaxed),
297 len,
298 capacity: self.capacity,
299 }
300 }
301
302 /// Locks the cache, recovering the guard if a previous holder panicked.
303 ///
304 /// A poisoned result cache is safe to keep using: a half-finished insert
305 /// can at worst drop or duplicate a memoized entry, never corrupt a result,
306 /// so recovery is preferable to propagating the panic.
307 fn lock_cache(&self) -> std::sync::MutexGuard<'_, PolicyCache<ResultKey, CacheEntry>> {
308 self.cache
309 .lock()
310 .unwrap_or_else(|poisoned| poisoned.into_inner())
311 }
312
313 /// Whether a cached entry is still live under the configured TTL.
314 ///
315 /// Always `true` when no TTL is set, so the clock is never read on the
316 /// non-TTL hot path.
317 #[inline]
318 fn is_live(&self, entry: &CacheEntry) -> bool {
319 match (self.ttl, entry.stamp) {
320 (Some(ttl), Some(stamp)) => self.clock.now().saturating_duration_since(stamp) < ttl,
321 _ => true,
322 }
323 }
324
325 /// Empties the cache through `&mut self` after a mutation.
326 fn invalidate(&mut self) {
327 // `&mut self` guarantees exclusive access, so no lock is contended.
328 match self.cache.get_mut() {
329 Ok(cache) => cache.clear(),
330 Err(poisoned) => poisoned.into_inner().clear(),
331 }
332 }
333}
334
335impl<I: IndexCore> IndexCore for CachedIndex<I> {
336 fn insert(
337 &mut self,
338 id: VectorId,
339 vector: std::sync::Arc<[f32]>,
340 metadata: Option<Metadata>,
341 ) -> Result<()> {
342 let result = self.inner.insert(id, vector, metadata);
343 if result.is_ok() {
344 self.invalidate();
345 }
346 result
347 }
348
349 fn insert_batch(
350 &mut self,
351 items: Vec<(VectorId, std::sync::Arc<[f32]>, Option<Metadata>)>,
352 ) -> Result<()> {
353 // `insert_batch` is fail-fast and may apply partially, so any outcome
354 // can have changed the search space: always invalidate.
355 let result = self.inner.insert_batch(items);
356 self.invalidate();
357 result
358 }
359
360 fn delete(&mut self, id: &VectorId) -> Result<()> {
361 let result = self.inner.delete(id);
362 if result.is_ok() {
363 self.invalidate();
364 }
365 result
366 }
367
368 fn search(&self, query: &[f32], params: &SearchParams) -> Result<Vec<Hit>> {
369 if self.capacity == 0 {
370 let _ = self.misses.fetch_add(1, Ordering::Relaxed);
371 return self.inner.search(query, params);
372 }
373
374 let key = ResultKey::new(query, params);
375 {
376 let mut cache = self.lock_cache();
377 if let Some(entry) = cache.get(&key) {
378 if self.is_live(entry) {
379 let _ = self.hits.fetch_add(1, Ordering::Relaxed);
380 return Ok(entry.hits.to_vec());
381 }
382 // Expired: fall through to recompute. The stale entry stays
383 // until the `put` below overwrites it with a fresh result.
384 }
385 }
386
387 // Miss (or expired): run the search without holding the lock so
388 // concurrent misses do not serialize on it.
389 let hits = self.inner.search(query, params)?;
390 let _ = self.misses.fetch_add(1, Ordering::Relaxed);
391 let stamp = self.ttl.map(|_| self.clock.now());
392 let evicted = {
393 let mut cache = self.lock_cache();
394 cache.put(
395 key,
396 CacheEntry {
397 hits: hits.clone().into_boxed_slice(),
398 stamp,
399 },
400 )
401 };
402 if evicted {
403 let _ = self.evictions.fetch_add(1, Ordering::Relaxed);
404 }
405 Ok(hits)
406 }
407
408 fn len(&self) -> usize {
409 self.inner.len()
410 }
411
412 fn is_empty(&self) -> bool {
413 self.inner.is_empty()
414 }
415
416 fn dim(&self) -> usize {
417 self.inner.dim()
418 }
419
420 fn metric(&self) -> DistanceMetric {
421 self.inner.metric()
422 }
423
424 fn flush(&mut self) -> Result<()> {
425 // Flush commits durable state without changing the searchable set, so
426 // the cache stays valid.
427 self.inner.flush()
428 }
429
430 fn stats(&self) -> IndexStats {
431 self.inner.stats()
432 }
433}
434
435#[cfg(test)]
436mod tests {
437 #![allow(clippy::unwrap_used)]
438
439 use clock_lib::ManualClock;
440
441 use super::*;
442 use crate::doc_stub::stub_index;
443
444 fn params() -> SearchParams {
445 SearchParams::new(1, DistanceMetric::Cosine)
446 }
447
448 #[test]
449 fn ttl_entry_is_recomputed_after_expiry() {
450 let clock = Arc::new(ManualClock::new());
451 let config = CacheConfig::new().capacity(8).ttl(Duration::from_secs(60));
452 let cached = CachedIndex::with_config_in(stub_index(), config, clock.clone());
453
454 let _miss = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
455 let _hit = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
456 assert_eq!(cached.cache_stats().hits, 1);
457
458 // Just inside the TTL: still a hit.
459 clock.advance(Duration::from_secs(59));
460 let _hit2 = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
461 assert_eq!(cached.cache_stats().hits, 2);
462
463 // Past the TTL: the entry expires and the search recomputes (a miss).
464 clock.advance(Duration::from_secs(2));
465 let _expired = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
466 assert_eq!(cached.cache_stats().hits, 2);
467 assert_eq!(cached.cache_stats().misses, 2);
468
469 // The recompute refreshed the entry, so the next search hits again.
470 let _hit3 = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
471 assert_eq!(cached.cache_stats().hits, 3);
472 }
473
474 #[test]
475 fn ttl_boundary_is_exclusive() {
476 let clock = Arc::new(ManualClock::new());
477 let config = CacheConfig::new().capacity(8).ttl(Duration::from_secs(10));
478 let cached = CachedIndex::with_config_in(stub_index(), config, clock.clone());
479
480 let _miss = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
481 // Exactly at the TTL counts as expired (`elapsed >= ttl`).
482 clock.advance(Duration::from_secs(10));
483 let _again = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
484 assert_eq!(cached.cache_stats().hits, 0);
485 assert_eq!(cached.cache_stats().misses, 2);
486 }
487
488 #[test]
489 fn no_ttl_never_expires_even_as_time_passes() {
490 let clock = Arc::new(ManualClock::new());
491 let config = CacheConfig::new().capacity(8); // no TTL
492 let cached = CachedIndex::with_config_in(stub_index(), config, clock.clone());
493
494 let _miss = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
495 let _hit = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
496 // Advance far beyond any plausible TTL: still a hit, because none is set.
497 clock.advance(Duration::from_secs(60 * 60 * 24 * 365));
498 let _hit2 = cached.search(&[1.0, 0.0, 0.0], ¶ms()).unwrap();
499 assert_eq!(cached.cache_stats().hits, 2);
500 assert_eq!(cached.cache_stats().misses, 1);
501 }
502}