Skip to content

lfu.md

Latest commit

 

History

History
50 lines (38 loc) · 1.87 KB

lfu.md

File metadata and controls

50 lines (38 loc) · 1.87 KB

LFU (Least Frequently Used)

Feature: policy-lfu

Goal

Evict the entry with the lowest access frequency; break ties by recency (common) or arbitrarily.

Common Production Approach: Bucketed LFU (O(1))

The typical O(1) LFU design is:

  • HashMap<K, EntryId> to locate entries
  • Per-entry metadata: freq, plus pointers for an intrusive list inside its frequency bucket
  • HashMap<freq, BucketList> (or Vec<BucketList> with bounded freq) storing LRU order within each frequency
  • min_freq tracking the smallest frequency currently present

In cachekit, src/policy/lfu.rs implements LFU with O(1) eviction based on bucket + min_freq.

Operations

get(key)

  • Return value and increment its frequency:
    1. Remove entry from old frequency bucket list.
    2. Increment freq.
    3. Insert into new bucket list (typically at head as MRU within that freq).
    4. If old bucket becomes empty and was min_freq, increment min_freq.

insert(key, value)

  • If present: update value and treat as an access (commonly increment frequency).
  • Else:
    • If full: evict from min_freq bucket (usually the LRU within that bucket).
    • Insert new entry with freq = 1 and set min_freq = 1.

pop_lfu()

  • Pop from the min_freq bucket; if it becomes empty, advance min_freq.

Complexity & Overhead

  • Lookup: O(1)
  • Frequency bump: O(1)
  • Eviction: O(1)
  • Metadata: higher than LRU/FIFO due to frequency buckets and per-entry freq

Aging / Decay (Important)

Pure LFU can keep “once-hot” items forever. To avoid this, add one of:

  • Periodic global decay: divide all frequencies by 2 each interval (O(n) unless amortized)
  • Time-decayed score: maintain score = freq / f(age) (more complex)
  • Windowed LFU: only count accesses in a moving time window

See lfu-aging.md for common patterns.

References