Your cache is silently draining your budget. Could 20 lines of code fix that?

Meet SIEVE – a drop-in eviction algorithm that’s simpler than LRU and more efficient than ARC, TwoQ, and even ML-based policies.

Why care?

Up to 63% lower miss ratio than ARC
2× higher throughput in multi-threaded workloads
Zero locks on cache hits = massive scalability boost
Implementation in <20 lines of code
No tuning. No parameters. No drama.

💰 What does that mean in dollars?

Let’s say your backend cache serves 50 million requests per hour and you’re spending $0.10 per GB-hour of memory + $0.02 per 1M backend requests due to cache misses. Switching from ARC to SIEVE can cut miss ratio by up to 63% – meaning fewer trips to the backend, less memory pressure, and better CPU efficiency. Even a 10% miss reduction can save tens of thousands of dollars per month at scale. And it costs almost nothing to try. SIEVE has been tested on 1,559 real-world traces (together contain 247,017 million requests to 14,852 million objects) from Twitter, Meta, and large CDNs — and consistently beats state-of-the-art algorithms in both efficiency and simplicity. It doesn’t just outperform LRU. It redefines what a cache eviction algorithm can be: elegant, fast, and production-ready.

Where can it help?

CDN edge caches
Internal KV stores
Redis-like in-memory systems
Any place you’re still relying on LRU as the default

What’s the magic?

SIEVE avoids LRU’s eager promotion and complex structures. Instead, it uses a clever combination of lazy promotion and quick demotion. Its internal logic is simple:

A FIFO queue
A moving “hand” pointer
A 1-bit visited flag per object

No multiple queues. No ML. No guesswork. And yet, it outperforms 9 state-of-the-art algorithms on 45%+ of traces. And because it’s lock-free on hits, it scales beautifully with modern CPUs.

Bottom line

If you’re still using LRU, ask yourself: Is it because it’s the best choice — or just the default one? Swap in SIEVE. Save compute. Save money.