Eviction Policies

When maxmemory is set, FerrumKV evicts keys according to the selected policy. Policies come in two families:

• allkeys-* — consider every key.
• volatile-* — consider only keys with a TTL set.

Policy	Type	Recency	Frequency	TTL-Aware	Self-Tuning
noeviction	—
allkeys-lru / volatile-lru	LRU	x		x
allkeys-lfu / volatile-lfu	LFU		x	x
allkeys-random / volatile-random	Random			x
volatile-ttl	TTL			x
allkeys-sieve / volatile-sieve	SIEVE (NSDI'24)	x
allkeys-sieves / volatile-sieves	SIEVE-S (FerrumKV)	x		x
allkeys-adaptiveclimb / volatile-adaptiveclimb	AdaptiveClimb (arXiv:2511.21235)	x	x		x
allkeys-ahe / volatile-ahe	Adaptive	x	x	x	x

SIEVE — Simple, Efficient Eviction (NSDI'24)

SIEVE (Zhang et al., NSDI'24) is a modern eviction algorithm that is simpler than LRU yet beats 9 state-of-the-art algorithms on 45%+ of 1,559 production traces, with ~2× LRU throughput. It keeps one FIFO queue plus one hand:

• On insert, a key joins the tail of the queue with its visited bit cleared.
• On access, its visited bit is set.
• On eviction, the hand sweeps forward clearing visited bits until it finds a key whose bit is still false — that key is the victim.

The key insight is quick demotion: a single missed access is enough to evict an object, whereas LRU gives every touch a "second chance". Eviction is O(1) amortised and needs no random sampling.

SIEVE-S — FerrumKV's TTL-aware variant

SIEVE-S adds a FerrumKV-original twist: keys within ~10% of expiry are force-demoted — treated as if visited were false regardless of recent access — so they become immediate eviction candidates and spare hotter, long-lived keys. (The NSDI'24 paper does not discuss TTL integration.)

Switch at runtime, exactly like any other policy:

CONFIG SET maxmemory-policy allkeys-sieve
CONFIG SET maxmemory-policy allkeys-sieves

AdaptiveClimb — Self-Tuning CLIMB (arXiv:2511.21235)

AdaptiveClimb (Berend et al., 2025) is a self-tuning variant of CLIMB. It keeps the keyspace in an ordered MRU→LRU list and a single scalar jump that controls how far a hit promotes a key toward the MRU end — and needs no per-item counters. On a hit it decrements jump and promotes the key by jump positions; on a miss it increments jump and inserts the new key jump positions from the LRU end. Because jump adapts from the live hit/miss ratio, the policy is self-tuning: more misses push jump up (new keys enter nearer the LRU end, resisting cache pollution), more hits push it down (frequent keys climb and stay protected).

It is the academic cousin of FerrumKV's AHE — both hill-climb a single parameter from hit-ratio feedback — but AdaptiveClimb trades TTL-awareness for a leaner, counter-free design.

CONFIG SET maxmemory-policy allkeys-adaptiveclimb
CONFIG SET maxmemory-policy volatile-adaptiveclimb

AHE — Adaptive Hybrid Eviction

AHE (Adaptive Hybrid Eviction) blends recency, frequency, and TTL urgency into a single self-tuning Eviction Priority Score. Unlike static policies, AHE continuously reweights each signal based on observed access patterns, so it adapts to workloads that shift between cache-friendly and scan-heavy behavior — without manual tuning.

Switch policies at runtime with CONFIG SET maxmemory-policy allkeys-ahe; no restart required.