# 🚀 Redis Memory Eviction: Deep Dive into the Algorithm Behind Redis Performance

Qian Liu — Mon, 29 Sep 2025 06:46:24 +0000

Redis memory eviction is a sophisticated mechanism that automatically manages memory when usage exceeds the maxmemory limit. Understanding this system is crucial for Redis optimization and production deployment.

🔑 Key Takeaways:

8 distinct eviction policies for different use cases
Innovative eviction pool algorithm with 16-slot optimization
Approximate LRU/LFU algorithms balancing accuracy and performance
Progressive eviction to maintain responsiveness

📊 Foundation: Two Core Data Structures

Redis maintains two essential dictionaries that form the foundation of its eviction system:

✅ db->keys: All key-value pairs (allkeys scope)

⏰ db->expires: Keys with expiration time (volatile scope)

🎯 The Eviction Pool: Redis's Secret Weapon

#define EVPOOL_SIZE 16
#define EVPOOL_CACHED_SDS_SIZE 255

struct evictionPoolEntry {
    unsigned long long idle;    // Idle time or score value
    sds key;                    // Key name
    sds cached;                 // Cached SDS object to avoid frequent allocation
    int dbid;                   // Database ID
    int slot;                   // Slot index
};

🎮 Eight Eviction Strategies: Choose Your Game Plan

Redis offers 8 distinct eviction policies, each optimized for different scenarios:

📋 Complete Policy Matrix

Policy	🎯 Use Case	Algorithm	Scope
`volatile-lru`	Cache with TTL	🕐 Least Recently Used	Keys with expiration
`volatile-lfu`	Popular data priority	📊 Least Frequently Used	Keys with expiration
`volatile-ttl`	Session management	⏱️ Shortest TTL first	Keys with expiration
`volatile-random`	Fair eviction	🎲 Random selection	Keys with expiration
`allkeys-lru`	General caching	🕐 Least Recently Used	All keys
`allkeys-lfu`	Hot data retention	📊 Least Frequently Used	All keys
`allkeys-random`	Simple eviction	🎲 Random selection	All keys
`noeviction`	Strict memory	🚫 No eviction, reject writes	-

🧠 The Eviction Pool Algorithm: Smart Selection Process

The eviction pool is Redis's optimization masterpiece - a 16-slot buffer that dramatically improves key selection quality:

⚡ How It Works (The Magic Behind the Scenes)

🔍 Step 1 - Sampling: Grab 5 random keys from target dictionary

📊 Step 2 - Scoring: Calculate "idle score" for each key

🎯 Step 3 - Pool Update: Insert best candidates into 16-slot pool

✨ Step 4 - Selection: Pick the highest-scoring key for eviction

💡 Why This Works:

Maintains 16 best candidates across multiple sampling rounds
Much better than naive "pick best of 5" approach
Balances selection quality with performance overhead

🎯 Scoring System: How Redis Decides

🕐 LRU: idle = estimateObjectIdleTime(kv);        // Older = Higher Score
📊 LFU: idle = 255 - LFUDecrAndReturn(kv);        // Less Used = Higher Score  
⏰ TTL: idle = ULLONG_MAX - kvobjGetExpire(kv);   // Expires Soon = Higher Score

🎲 Random Strategy: No Pool Needed

// Round-robin search across databases for random key
for (i = 0; i < server.dbnum; i++) {
    j = (++next_db) % server.dbnum;
    db = server.db + j;

    kvstore *kvs = (policy == MAXMEMORY_ALLKEYS_RANDOM) ? 
                   db->keys : db->expires;

    int slot = kvstoreGetFairRandomDictIndex(kvs);
    de = kvstoreDictGetRandomKey(kvs, slot);
    if (de) {
        bestkey = kvobjGetKey(dictGetKV(de));
        break;
    }
}

🔬 LRU vs LFU: The Algorithm Showdown

🕐 LRU Implementation: 24-bit Timestamp Magic

#define LRU_CLOCK_RESOLUTION 1000    // 1 second precision
#define LRU_CLOCK_MAX ((1<<LRU_BITS)-1)

// Get current LRU clock
unsigned int getLRUClock(void) {
    return (mstime()/LRU_CLOCK_RESOLUTION) & LRU_CLOCK_MAX;
}

// Estimate object idle time
unsigned long long estimateObjectIdleTime(robj *o) {
    unsigned long long lruclock = LRU_CLOCK();
    if (lruclock >= o->lru) {
        return (lruclock - o->lru) * LRU_CLOCK_RESOLUTION;
    } else {
        // Handle clock wraparound
        return (lruclock + (LRU_CLOCK_MAX - o->lru)) * LRU_CLOCK_RESOLUTION;
    }
}

🎯 Redis's Clever Trade-off:

Only 24 bits per object (vs full timestamp)
1-second precision (sufficient for most cases)
O(1) idle time calculation
Handles clock wraparound gracefully

📊 LFU Implementation: Logarithmic Frequency Counter

🧮 Brilliant 24-bit Split:

     16 bits       8 bits
+------------------+--------+
| Last access time | LOG_C  |
+------------------+--------+

16 bits: Last access time (minute precision)
8 bits: Logarithmic frequency counter (0-255)

🚀 Why Logarithmic?

Linear counters saturate too quickly
Log counters give diminishing returns for high-frequency items
Probabilistic increment: harder to increment as frequency grows

// Logarithmic increment: Higher frequency makes increment harder
uint8_t LFULogIncr(uint8_t counter) {
    if (counter == 255) return 255;
    double r = (double)rand()/RAND_MAX;
    double baseval = counter - LFU_INIT_VAL;
    if (baseval < 0) baseval = 0;
    double p = 1.0/(baseval * server.lfu_log_factor + 1);
    if (r < p) counter++;
    return counter;
}

// Time decay: Reduce frequency based on elapsed time
unsigned long LFUDecrAndReturn(robj *o) {
    unsigned long ldt = o->lru >> 8;
    unsigned long counter = o->lru & 255;
    unsigned long num_periods = server.lfu_decay_time ? 
        LFUTimeElapsed(ldt) / server.lfu_decay_time : 0;
    if (num_periods)
        counter = (num_periods > counter) ? 0 : counter - num_periods;
    return counter;
}

🔧 Eviction Execution Flow

⚙️ Main Function: `performEvictions()`

int performEvictions(void) {
    // 1. Safety check
    if (!isSafeToPerformEvictions()) return EVICT_OK;

    // 2. Memory state check
    if (getMaxmemoryState(&mem_reported, NULL, &mem_tofree, NULL) == C_OK) {
        return EVICT_OK;
    }

    // 3. Policy check
    if (server.maxmemory_policy == MAXMEMORY_NO_EVICTION) {
        return EVICT_FAIL;
    }

    // 4. Loop eviction until enough memory is freed
    while (mem_freed < mem_tofree) {
        bestkey = selectEvictionKey();
        deleteEvictedKeyAndPropagate(db, keyobj, &key_mem_freed);
        mem_freed += key_mem_freed;

        // Periodic checks every 16 keys
        if (keys_freed % 16 == 0) {
            flushSlavesOutputBuffers();
            checkTimeLimit();
        }
    }
    return EVICT_OK;
}

⏱️ Time Limit & Progressive Eviction

// Calculate time limit based on tenacity parameter
static unsigned long evictionTimeLimitUs(void) {
    if (server.maxmemory_eviction_tenacity <= 10) {
        return 50uL * server.maxmemory_eviction_tenacity;  // 0-500μs
    }
    if (server.maxmemory_eviction_tenacity < 100) {
        return (unsigned long)(500.0 * pow(1.15, server.maxmemory_eviction_tenacity - 10.0));
    }
    return ULONG_MAX;  // No limit
}

// Start background eviction when timeout occurs
void startEvictionTimeProc(void) {
    if (!isEvictionProcRunning) {
        isEvictionProcRunning = 1;
        aeCreateTimeEvent(server.el, 0, evictionTimeProc, NULL, NULL);
    }
}

⚡ Performance Optimizations That Matter

🚀 Sampling Strategy: 5 keys → 16-slot pool (smart caching)

⚡ Time Slicing: Progressive eviction prevents blocking

🧠 SDS Caching: Reuse string objects in eviction pool

🔄 Lazy Free: Background deletion for large objects

📊 Batch Processing: Check every 16 keys for efficiency

🛠️ Production Configuration Guide

⚙️ Critical Parameters

Parameter	Default	🎯 Impact	Tuning Tips
`maxmemory`	0	🚨 Memory limit	Set to 80% of available RAM
`maxmemory-policy`	`noeviction`	🎮 Strategy choice	Use `allkeys-lru` for cache
`maxmemory-samples`	5	🎯 Selection quality	Increase to 10 for better accuracy
`maxmemory-eviction-tenacity`	10	⚡ Responsiveness	Lower for latency-sensitive apps
`lfu-log-factor`	10	📊 LFU sensitivity	Higher = more gradual frequency growth
`lfu-decay-time`	1	⏰ LFU time decay	Adjust based on access patterns

📈 Monitoring & Best Practices

🔍 Essential Metrics to Watch

📊 evicted_keys: Track eviction frequency

⚠️ eviction_exceeded_time: Monitor blocking time

🧩 mem_fragmentation_ratio: Watch memory efficiency

💾 used_memory: Current memory usage

🎯 maxmemory: Your memory ceiling

🎯 Strategy Selection Guide

🏆 Production Recommendations:

🌐 Web Caching: allkeys-lru or allkeys-lfu

Most cached data follows access patterns
Evict across all keys for maximum efficiency

🔐 Session Storage: volatile-ttl

Sessions have natural expiration times
TTL-based eviction respects session lifecycle

📊 Analytics Cache: allkeys-lfu

Popular queries should stay longer
Frequency matters more than recency

🎲 Development/Testing: allkeys-random

Simple and predictable behavior
Good for non-production environments

⚡ Performance Tuning Tips

🚀 For High-Throughput Systems:

Increase maxmemory-samples to 10-20
Use maxmemory-eviction-tenacity 20-30
Monitor eviction_exceeded_time closely

🎯 For Low-Latency Systems:

Keep maxmemory-samples at 5
Set maxmemory-eviction-tenacity to 5-10
Consider lazyfree-lazy-eviction yes

💡 For Memory-Constrained Systems:

Use aggressive eviction policies (allkeys-*)
Set maxmemory to 70-80% of available RAM
Monitor fragmentation ratio regularly

🎉 Key Takeaways

✅ Smart Algorithm: 16-slot eviction pool beats naive sampling

✅ Balanced Trade-offs: Approximate algorithms for production performance

✅ Configurable: 8 policies + tunable parameters for any use case

✅ Production-Ready: Time limits and progressive eviction prevent blocking

This deep dive is based on Redis 8.2 source code analysis. Understanding these internals helps you optimize Redis for your specific use case!

🔗 Want to dive deeper? Check out the Redis source code in src/evict.c - it's surprisingly readable and well-commented!

DEV Community: Qian Liu