explore: foyer hybrid cache for search (memory + disk)

quickwit/Cargo.toml

@@ -116,6 +116,7 @@ dyn-clone = "1.0"
 enum-iterator = "2.3"
 env_logger = { version = "0.11", default-features = false, features = ["auto-color"] }
 fail = "0.5"
+foyer = { version = "0.22.3", features = ["serde"] }
 flate2 = "1.1"
 flume = "0.12"
 fnv = "1"

quickwit/quickwit-config/src/node_config/mod.rs

@@ -283,6 +283,15 @@ pub struct SearcherConfig {
                                 {ByteSize::mb(256).as_u64()}>")]
     pub predicate_cache: CacheConfig,
 
+    /// Cache for term dictionaries (.term files). Disabled (0) by default.
+    #[serde(alias = "term_dict_cache_capacity")]
+    #[serde(deserialize_with = "CacheConfig::deserialize_with_default::<_, 0>")]
+    pub term_dict_cache: CacheConfig,
+    /// Cache for posting lists (.idx files). Disabled (0) by default.
+    #[serde(alias = "posting_list_cache_capacity")]
+    #[serde(deserialize_with = "CacheConfig::deserialize_with_default::<_, 0>")]
+    pub posting_list_cache: CacheConfig,
+
     pub max_num_concurrent_split_searches: usize,
     pub max_splits_per_search: Option<usize>,
     // Deprecated: stream search requests are no longer supported.
@@ -302,6 +311,17 @@ pub struct SearcherConfig {
     pub storage_timeout_policy: Option<StorageTimeoutPolicy>,
     pub warmup_memory_budget: ByteSize,
     pub warmup_single_split_initial_allocation: ByteSize,
+    /// Enable the disk tier for the partial request cache.
+    /// When enabled, evicted entries from the in-memory partial request cache
+    /// spill to `{data_dir}/partial-request-cache/` (NVMe recommended) instead
+    /// of being lost. Disabled by default.
+    #[serde(default)]
+    pub partial_request_disk_cache_enabled: bool,
+    /// Disk cache capacity for the partial request cache (default: 1 GB).
+    /// Ignored if partial_request_disk_cache_enabled is false.
+    #[serde(default = "SearcherConfig::default_partial_request_disk_cache_capacity")]
+    pub partial_request_disk_cache_capacity: ByteSize,
+
     /// Lambda configuration for serverless leaf search execution.
     /// If set, enables Lambda execution for leaf search.
     ///
@@ -514,6 +534,8 @@ impl Default for SearcherConfig {
             split_footer_cache: CacheConfig::default_with_capacity(ByteSize::mb(500)),
             partial_request_cache: CacheConfig::default_with_capacity(ByteSize::mb(64)),
             predicate_cache: CacheConfig::default_with_capacity(ByteSize::mb(256)),
+            term_dict_cache: CacheConfig::default_with_capacity(ByteSize(0)),
+            posting_list_cache: CacheConfig::default_with_capacity(ByteSize(0)),
             max_num_concurrent_split_searches: 100,
             max_splits_per_search: None,
             _max_num_concurrent_split_streams: None,
@@ -525,12 +547,18 @@ impl Default for SearcherConfig {
             storage_timeout_policy: None,
             warmup_memory_budget: ByteSize::gb(100),
             warmup_single_split_initial_allocation: ByteSize::mb(300),
+            partial_request_disk_cache_enabled: false,
+            partial_request_disk_cache_capacity: Self::default_partial_request_disk_cache_capacity(),
             lambda: None,
         }
     }
 }
 
 impl SearcherConfig {
+    fn default_partial_request_disk_cache_capacity() -> ByteSize {
+        ByteSize::gb(1)
+    }
+
     /// The timeout applied at the gRPC layer for search requests
     pub fn request_timeout(&self) -> Duration {
         Duration::from_secs(self.request_timeout_secs.get())

quickwit/quickwit-search/Cargo.toml

@@ -17,6 +17,7 @@ base64 = { workspace = true }
 bytes = { workspace = true }
 bytesize = { workspace = true }
 fnv = { workspace = true }
+foyer = { workspace = true }
 futures = { workspace = true }
 http = { workspace = true }
 itertools = { workspace = true }