Approximate LRU/LFU Sampling, LFU Decay, Hot-Set Behavior, and Cache Hit Rate

Learning outcomes

AtlasMart has a small hot set of products and a long cold tail. The team wants “perfect LRU,” but perfect global ordering would itself consume CPU/memory. Redis instead samples candidates and uses compact metadata, so the useful question is whether the approximation preserves the workload's valuable working set.

1. LRU is sampled, not a global exact list

Redis LRU samples a small number of keys, keeps good candidates, and evicts among those candidates rather than maintaining a perfect ordering of every key. maxmemory-samples controls the sample count. Larger samples can more closely approximate ideal LRU but consume additional CPU during eviction.

2. Start with LFU and inspect its knobs

docker volume create atlasmart-redis-ch18-policy_datadocker run --rm -v atlasmart-redis-ch18-policy_data:/data redis:8.10.1 sh -lc 'printf "%s\n" "user default off" "user academy-admin on >AtlasMart-Admin-Lab-Only-2026 ~* &* +@all" > /data/users.acl'docker run -d --name atlasmart-redis-ch18-policy --memory 192m -p 127.0.0.1:6393:6379 -v atlasmart-redis-ch18-policy_data:/data redis:8.10.1 redis-server --dir /data --aclfile /data/users.acl --save '' --appendonly no --maxmemory 24mb --maxmemory-policy allkeys-lfu --maxmemory-samples 5 --lfu-log-factor 10 --lfu-decay-time 1docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin PINGdocker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin ACL WHOAMI

docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin CONFIG GET maxmemory maxmemory-policy maxmemory-samples lfu-log-factor lfu-decay-timedocker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin INFO statsdocker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin INFO memory

3. LFU stores a compact probabilistic frequency estimate

LFU uses a logarithmic Morris-style counter rather than an exact request count, then decays old popularity over time. The default lfu-log-factor 10 controls counter growth and lfu-decay-time 1 means one-minute decay intervals when sampled. The counter is an eviction heuristic—not an analytics metric.

import os, redisHOST = "127.0.0.1"PORT = 6393r = redis.Redis(host=HOST, port=PORT, username="academy-admin", password="AtlasMart-Admin-Lab-Only-2026", decode_responses=False)assert r.ping()import randomrng = random.Random(1803)payload = b"v" * 16384for i in range(500): r.set(f"atlasmart:ch18:l3:item:{i}", payload)hot = list(range(20))warm = list(range(20, 100))for _ in range(6000):    x = rng.random()    if x < 0.75: i = rng.choice(hot)    elif x < 0.95: i = rng.choice(warm)    else: i = rng.randrange(100, 500)    r.get(f"atlasmart:ch18:l3:item:{i}")print("hot_freq", [r.object("freq", f"atlasmart:ch18:l3:item:{i}") for i in range(5)])print("cold_freq", [r.object("freq", f"atlasmart:ch18:l3:item:{i}") for i in range(450,455)])

docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin OBJECT FREQ atlasmart:ch18:l3:item:0docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin OBJECT FREQ atlasmart:ch18:l3:item:450docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin --hotkeys --pattern "atlasmart:ch18:l3:*" --count 100

4. OBJECT FREQ and IDLETIME are policy-specific diagnostics

OBJECT FREQ is available only under LFU policies. OBJECT IDLETIME is unavailable when the maxmemory policy is LFU because the compact object metadata is used for LFU frequency instead. Always verify policy before interpreting either command.

docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin OBJECT FREQ atlasmart:ch18:l3:item:0docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin OBJECT IDLETIME atlasmart:ch18:l3:item:0# Under an LFU policy, OBJECT IDLETIME should report that it is unavailable.

5. Pressure the cache and measure population-level outcomes

The right test is not “did the mathematically oldest key disappear?” It is “under this reproducible access distribution and memory pressure, did the policy preserve the hot working set and deliver an acceptable miss/latency profile?”

import os, redisHOST = "127.0.0.1"PORT = 6393r = redis.Redis(host=HOST, port=PORT, username="academy-admin", password="AtlasMart-Admin-Lab-Only-2026", decode_responses=False)assert r.ping()stats0 = r.info("stats")hits0, misses0 = stats0["keyspace_hits"], stats0["keyspace_misses"]payload = b"n" * 16384for i in range(500, 1800): r.set(f"atlasmart:ch18:l3:item:{i}", payload)alive_hot = sum(r.exists(f"atlasmart:ch18:l3:item:{i}") for i in range(20))alive_cold = sum(r.exists(f"atlasmart:ch18:l3:item:{i}") for i in range(450,500))for i in range(20): r.get(f"atlasmart:ch18:l3:item:{i}")for i in range(450,500): r.get(f"atlasmart:ch18:l3:item:{i}")stats1 = r.info("stats")dh = stats1["keyspace_hits"] - hits0dm = stats1["keyspace_misses"] - misses0print("hot_alive", alive_hot, "of 20", "cold_alive", alive_cold, "of 50")print("probe_hit_rate", dh / max(1, dh + dm), "evicted_keys", stats1["evicted_keys"])

6. Change maxmemory-samples only as an experiment

The default sample count is a deliberate efficiency tradeoff. Increasing it can make LRU/LRM candidate selection closer to an ideal ordering, but more sampling adds CPU work. For LFU, sampling is only one part of behavior; the logarithmic counter and decay parameters also matter.

docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin CONFIG GET maxmemory-samplesdocker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin CONFIG SET maxmemory-samples 10docker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin CONFIG GET maxmemory-samplesdocker exec -e REDISCLI_AUTH=AtlasMart-Admin-Lab-Only-2026 atlasmart-redis-ch18-policy redis-cli --user academy-admin CONFIG SET maxmemory-samples 5

7. LFU decay is adaptation, not a timer that resets counters exactly

Decay prevents yesterday's hot key from staying protected forever. Redis applies decay as keys participate in the LFU mechanism; lfu-decay-time should not be interpreted as a background scheduler that resets every counter at an exact wall-clock boundary. A value of zero disables decay.

8. Redis 8.6+ LRM is a separate question

Least Recently Modified (LRM) was added in Redis 8.6. It tracks write recency rather than read+write access recency. That can be useful for read-heavy datasets where frequently read but stale values should still be considered old. It is not a replacement for LRU or LFU; it expresses a different notion of value.

9. Verification checklist

• Confirm policy before using OBJECT FREQ or OBJECT IDLETIME.
• Capture frequency counters as heuristic metadata, not exact request counts.
• Measure keyspace_hits/keyspace_misses deltas and evicted_keys.
• Compare hot-set and cold-tail survival under the same deterministic trace.
• Do not change LFU decay/log-factor or sample size without latency and workload evidence.

Production judgment and next bridge

Cache quality is a distribution-level property. Measure hit rate, tail latency, evictions, skew, and memory together. Lesson 4 moves from policy heuristics to pathological shapes: big keys, hot keys, asynchronous freeing, fragmentation, and active defrag diagnostics.

Summary and next step

Approximate LRU/LFU Sampling, LFU Decay, Hot-Set Behavior, and Cache Hit Rate is now connected to observable Redis behavior, bounded failure cases, and production tradeoffs. Keep the evidence and cleanup state from this lesson; next, continue with Big Keys, Hot Keys, Large Collections, Lazy Freeing, Active Defrag, and Memory Diagnostics.

Authoritative references

• Redis key eviction
• Redis memory optimization
• INFO
• MEMORY STATS
• MEMORY USAGE
• MEMORY DOCTOR
• MEMORY PURGE
• MEMORY MALLOC-STATS
• OBJECT FREQ
• OBJECT IDLETIME
• UNLINK
• Redis CLI keyspace analysis
• Redis latency monitoring
• Redis persistence
• Redis replication
• Redis Sentinel
• Redis Cluster specification
• Redis Search and query
• Redis 8.10 release notes
• Redis 8.10 whats new
• Redis ACLs
• Redis security
• Redis licenses
• Docker Official Redis image