From b1310a1bf8a7954e0ffb030dd3b438017b1a7ff8 Mon Sep 17 00:00:00 2001 From: Yujong Lee Date: Mon, 7 Sep 2026 22:38:45 -0700 Subject: [PATCH] refactor(rust): remove generic core cache --- .../core/src/caching/in_memory_cache.rs | 258 ------------------ litellm-rust/crates/core/src/caching/mod.rs | 1 - litellm-rust/crates/core/src/lib.rs | 1 - .../core/src/providers/bedrock/aws_base.rs | 59 +++- 4 files changed, 50 insertions(+), 269 deletions(-) delete mode 100644 litellm-rust/crates/core/src/caching/in_memory_cache.rs delete mode 100644 litellm-rust/crates/core/src/caching/mod.rs diff --git a/litellm-rust/crates/core/src/caching/in_memory_cache.rs b/litellm-rust/crates/core/src/caching/in_memory_cache.rs deleted file mode 100644 index 45d4bd69b79..00000000000 --- a/litellm-rust/crates/core/src/caching/in_memory_cache.rs +++ /dev/null @@ -1,258 +0,0 @@ -use std::cmp::Reverse; -use std::collections::{BinaryHeap, HashMap}; -use std::time::{Duration, SystemTime, UNIX_EPOCH}; - -const DEFAULT_MAX_SIZE_IN_MEMORY: usize = 200; -const DEFAULT_TTL: Duration = Duration::from_secs(600); - -pub struct InMemoryCache { - pub cache_dict: HashMap, - pub ttl_dict: HashMap, - pub expiration_heap: BinaryHeap>, - pub max_size_in_memory: usize, - pub default_ttl: Duration, - now: Box Duration + Send + Sync>, -} - -impl Default for InMemoryCache { - fn default() -> Self { - Self::new(None, None) - } -} - -impl InMemoryCache { - pub fn new(max_size_in_memory: Option, default_ttl: Option) -> Self { - Self::with_clock(max_size_in_memory, default_ttl, || { - SystemTime::now() - .duration_since(UNIX_EPOCH) - .unwrap_or_default() - }) - } - - pub fn with_clock( - max_size_in_memory: Option, - default_ttl: Option, - now: impl Fn() -> Duration + Send + Sync + 'static, - ) -> Self { - Self { - cache_dict: HashMap::new(), - ttl_dict: HashMap::new(), - expiration_heap: BinaryHeap::new(), - max_size_in_memory: max_size_in_memory.unwrap_or(DEFAULT_MAX_SIZE_IN_MEMORY), - default_ttl: default_ttl.unwrap_or(DEFAULT_TTL), - now: Box::new(now), - } - } - - pub fn evict_cache(&mut self) { - if self.max_size_in_memory == 0 { - return; - } - - let current_time = (self.now)(); - while let Some(Reverse((expiration_time, key))) = self.expiration_heap.peek().cloned() { - if self.ttl_dict.get(&key).copied() != Some(expiration_time) { - self.expiration_heap.pop(); - } else if expiration_time <= current_time { - self.expiration_heap.pop(); - self.remove_key(&key); - } else { - break; - } - } - - while self.cache_dict.len() >= self.max_size_in_memory { - let Some(Reverse((expiration_time, key))) = self.expiration_heap.pop() else { - break; - }; - if self.ttl_dict.get(&key).copied() == Some(expiration_time) { - self.remove_key(&key); - } - } - } - - pub fn allow_ttl_override(&self, key: &str) -> bool { - match self.ttl_dict.get(key).copied() { - None => true, - Some(expiration_time) => expiration_time < (self.now)(), - } - } - - pub fn set_cache(&mut self, key: impl Into, value: V, ttl: Option) { - if self.max_size_in_memory == 0 { - return; - } - - self.evict_cache(); - let key = key.into(); - self.cache_dict.insert(key.clone(), value); - if self.allow_ttl_override(&key) { - let expiration_time = (self.now)() + ttl.unwrap_or(self.default_ttl); - self.ttl_dict.insert(key.clone(), expiration_time); - self.expiration_heap.push(Reverse((expiration_time, key))); - } - } - - // Generic values intentionally omit Python's per-item size check. - pub fn get_cache(&mut self, key: &str) -> Option { - if self.cache_dict.contains_key(key) { - if self.is_key_expired(key) { - self.remove_key(key); - return None; - } - return self.cache_dict.get(key).cloned(); - } - None - } - - pub fn get_ttl(&self, key: &str) -> Option { - self.ttl_dict.get(key).copied() - } - - pub fn delete_cache(&mut self, key: &str) { - self.remove_key(key); - } - - pub fn flush_cache(&mut self) { - self.cache_dict.clear(); - self.ttl_dict.clear(); - self.expiration_heap.clear(); - } - - fn is_key_expired(&self, key: &str) -> bool { - self.ttl_dict - .get(key) - .is_some_and(|expiration_time| *expiration_time < (self.now)()) - } - - fn remove_key(&mut self, key: &str) { - self.cache_dict.remove(key); - self.ttl_dict.remove(key); - } -} - -#[cfg(test)] -mod tests { - use std::sync::{ - Arc, - atomic::{AtomicU64, Ordering}, - }; - - use super::InMemoryCache; - use std::time::Duration; - - fn cache(now: Arc, max_size: usize, default_ttl: Duration) -> InMemoryCache { - InMemoryCache::with_clock(Some(max_size), Some(default_ttl), move || { - Duration::from_secs(now.load(Ordering::Relaxed)) - }) - } - - #[test] - fn ttl_expiry_is_deterministic() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now.clone(), 10, Duration::from_secs(60)); - cache.set_cache("key", "value".to_string(), None); - assert_eq!(cache.get_cache("key"), Some("value".to_string())); - now.store(161, Ordering::Relaxed); - assert_eq!(cache.get_cache("key"), None); - assert_eq!(cache.get_ttl("key"), None); - } - - #[test] - fn default_and_per_set_ttl_are_applied() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now.clone(), 10, Duration::from_secs(60)); - cache.set_cache("default", "value".to_string(), None); - cache.set_cache("custom", "value".to_string(), Some(Duration::from_secs(20))); - assert_eq!(cache.get_ttl("default"), Some(Duration::from_secs(160))); - assert_eq!(cache.get_ttl("custom"), Some(Duration::from_secs(120))); - } - - #[test] - fn unexpired_entries_do_not_allow_ttl_override() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now.clone(), 10, Duration::from_secs(60)); - cache.set_cache("key", "first".to_string(), Some(Duration::from_secs(20))); - cache.set_cache("key", "second".to_string(), Some(Duration::from_secs(80))); - assert_eq!(cache.get_cache("key"), Some("second".to_string())); - assert_eq!(cache.get_ttl("key"), Some(Duration::from_secs(120))); - now.store(121, Ordering::Relaxed); - cache.set_cache("key", "third".to_string(), Some(Duration::from_secs(80))); - assert_eq!(cache.get_ttl("key"), Some(Duration::from_secs(201))); - } - - #[test] - fn max_size_evicts_earliest_expiration() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now, 2, Duration::from_secs(60)); - cache.set_cache("early", "value".to_string(), Some(Duration::from_secs(10))); - cache.set_cache("late", "value".to_string(), Some(Duration::from_secs(20))); - cache.set_cache("new", "value".to_string(), Some(Duration::from_secs(30))); - assert_eq!(cache.get_cache("early"), None); - assert!(cache.get_cache("late").is_some()); - assert!(cache.get_cache("new").is_some()); - } - - #[test] - fn expired_entries_are_evicted_before_live_entries() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now.clone(), 3, Duration::from_secs(60)); - cache.set_cache( - "expired-one", - "value".to_string(), - Some(Duration::from_secs(10)), - ); - cache.set_cache( - "expired-two", - "value".to_string(), - Some(Duration::from_secs(20)), - ); - cache.set_cache("live", "value".to_string(), Some(Duration::from_secs(100))); - now.store(121, Ordering::Relaxed); - cache.set_cache("new", "value".to_string(), Some(Duration::from_secs(100))); - assert_eq!(cache.get_cache("expired-one"), None); - assert_eq!(cache.get_cache("expired-two"), None); - assert!(cache.get_cache("live").is_some()); - assert!(cache.get_cache("new").is_some()); - } - - #[test] - fn stale_heap_entries_are_skipped() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now, 1, Duration::from_secs(60)); - cache.set_cache( - "removed", - "value".to_string(), - Some(Duration::from_secs(10)), - ); - cache.delete_cache("removed"); - cache.set_cache("kept", "value".to_string(), Some(Duration::from_secs(20))); - cache.set_cache("new", "value".to_string(), Some(Duration::from_secs(30))); - assert_eq!(cache.get_cache("removed"), None); - assert_eq!(cache.get_cache("kept"), None); - assert!(cache.get_cache("new").is_some()); - } - - #[test] - fn delete_and_flush_remove_values_and_ttls() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now, 10, Duration::from_secs(60)); - cache.set_cache("one", "value".to_string(), None); - cache.set_cache("two", "value".to_string(), None); - cache.delete_cache("one"); - assert_eq!(cache.get_cache("one"), None); - cache.flush_cache(); - assert!(cache.cache_dict.is_empty()); - assert!(cache.ttl_dict.is_empty()); - assert!(cache.expiration_heap.is_empty()); - } - - #[test] - fn zero_max_size_does_not_cache() { - let now = Arc::new(AtomicU64::new(100)); - let mut cache = cache(now, 0, Duration::from_secs(60)); - cache.set_cache("key", "value".to_string(), None); - assert_eq!(cache.get_cache("key"), None); - assert!(cache.cache_dict.is_empty()); - } -} diff --git a/litellm-rust/crates/core/src/caching/mod.rs b/litellm-rust/crates/core/src/caching/mod.rs deleted file mode 100644 index 5fb8a0e5174..00000000000 --- a/litellm-rust/crates/core/src/caching/mod.rs +++ /dev/null @@ -1 +0,0 @@ -pub mod in_memory_cache; diff --git a/litellm-rust/crates/core/src/lib.rs b/litellm-rust/crates/core/src/lib.rs index 89c3c05dfbb..98fd54ba58d 100644 --- a/litellm-rust/crates/core/src/lib.rs +++ b/litellm-rust/crates/core/src/lib.rs @@ -1,5 +1,4 @@ pub mod audio_transcription; -pub mod caching; pub mod chat_completions; pub mod constants; pub mod error; diff --git a/litellm-rust/crates/core/src/providers/bedrock/aws_base.rs b/litellm-rust/crates/core/src/providers/bedrock/aws_base.rs index e5e52bfce95..072488fb8d9 100644 --- a/litellm-rust/crates/core/src/providers/bedrock/aws_base.rs +++ b/litellm-rust/crates/core/src/providers/bedrock/aws_base.rs @@ -1,9 +1,8 @@ -use std::collections::BTreeMap; +use std::cmp::Reverse; +use std::collections::{BTreeMap, BinaryHeap, HashMap}; use std::sync::{Mutex, OnceLock}; -use std::time::Duration; -use std::time::{SystemTime, UNIX_EPOCH}; +use std::time::{Duration, SystemTime, UNIX_EPOCH}; -use crate::caching::in_memory_cache::InMemoryCache; use crate::error::Error; use aws_credential_types::Credentials; use aws_credential_types::provider::ProvideCredentials; @@ -25,8 +24,50 @@ use super::constants::{ const STATIC_CREDENTIALS_TTL: Duration = Duration::from_secs(3600 - 60); const AMBIENT_CREDENTIALS_TTL: Duration = Duration::from_secs(600); +const MAX_CACHED_CREDENTIALS: usize = 200; -static IAM_CREDENTIALS_CACHE: OnceLock>> = OnceLock::new(); +#[derive(Default)] +struct CredentialCache { + entries: HashMap, + expirations: BinaryHeap>, +} + +impl CredentialCache { + fn get(&mut self, key: &str) -> Option { + let now = unix_time(); + let (credentials, expiration) = self.entries.get(key)?; + if *expiration > now { + return Some(credentials.clone()); + } + self.entries.remove(key); + None + } + + fn insert(&mut self, key: String, credentials: Credentials, ttl: Duration) { + let now = unix_time(); + while let Some(Reverse((expiration, key))) = self.expirations.peek().cloned() { + if self.entries.get(&key).map(|(_, current)| *current) != Some(expiration) { + self.expirations.pop(); + } else if expiration <= now || self.entries.len() >= MAX_CACHED_CREDENTIALS { + self.expirations.pop(); + self.entries.remove(&key); + } else { + break; + } + } + let expiration = now + ttl; + self.entries.insert(key.clone(), (credentials, expiration)); + self.expirations.push(Reverse((expiration, key))); + } +} + +static IAM_CREDENTIALS_CACHE: OnceLock> = OnceLock::new(); + +fn unix_time() -> Duration { + SystemTime::now() + .duration_since(UNIX_EPOCH) + .unwrap_or_default() +} fn credential_cache_ttl(flow: &AwsAuthFlow) -> Option { match flow { @@ -108,15 +149,15 @@ fn cache_key(config: &AwsAuthConfig, flow: &AwsAuthFlow) -> String { } fn get_cached_credentials(key: &str) -> Option { - let cache = IAM_CREDENTIALS_CACHE.get_or_init(|| Mutex::new(InMemoryCache::default())); + let cache = IAM_CREDENTIALS_CACHE.get_or_init(|| Mutex::new(CredentialCache::default())); let mut entries = cache.lock().ok()?; - entries.get_cache(key) + entries.get(key) } fn set_cached_credentials(key: String, credentials: Credentials, ttl: Duration) { - let cache = IAM_CREDENTIALS_CACHE.get_or_init(|| Mutex::new(InMemoryCache::default())); + let cache = IAM_CREDENTIALS_CACHE.get_or_init(|| Mutex::new(CredentialCache::default())); if let Ok(mut entries) = cache.lock() { - entries.set_cache(key, credentials, Some(ttl)); + entries.insert(key, credentials, ttl); } }