feldera-buffer-cache

feldera-buffer-cache provides SieveCache, a sharded, weighted, thread-safe in-memory buffer cache used by Feldera's storage layer.

The implementation is inspired by the NSDI'24 SIEVE paper: https://www.usenix.org/conference/nsdi24/presentation/zhang-yazhuo

Implementation

• Weighted capacity accounting. Every insert supplies a charge, and eviction is based on the sum of charges instead of entry count.
• Sharded concurrency. SieveCache::new(total_capacity) uses SieveCache::DEFAULT_SHARDS, while SieveCache::with_shards(...) lets the caller pick an explicit power-of-two shard count.
• Cheap hit handling. A cache hit only flips a visited bit; it does not move the entry through a recency list.
• Shared values. Public operations return Arc<V> so callers can keep values alive after removal or eviction.

API sketch

use feldera_buffer_cache::SieveCache;

let cache = SieveCache::<u64, Vec<u8>>::with_shards(64 << 20, 64);
cache.insert(7, vec![0; 4096], 4096);

let page = cache.get(&7).expect("entry should be present");
assert_eq!(page.len(), 4096);

Benchmarks

Comparisons against dbsp's old LRU buffer cache live under dbsp bench targets so they exercise the real in-tree implementation:

cargo bench -p dbsp --bench buffer_cache -- --threads 1,2,4,8 --max-duration 10